CN114036821A - Thickener control method and device based on non-deterministic hidden space model - Google Patents

Thickener control method and device based on non-deterministic hidden space model Download PDF

Info

Publication number
CN114036821A
CN114036821A CN202111227806.0A CN202111227806A CN114036821A CN 114036821 A CN114036821 A CN 114036821A CN 202111227806 A CN202111227806 A CN 202111227806A CN 114036821 A CN114036821 A CN 114036821A
Authority
CN
China
Prior art keywords
thickener
deterministic
space model
thickener system
time state
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202111227806.0A
Other languages
Chinese (zh)
Inventor
班晓娟
张子轩
袁兆麟
李潇睿
阮竹恩
王贻明
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
University of Science and Technology Beijing USTB
Original Assignee
University of Science and Technology Beijing USTB
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by University of Science and Technology Beijing USTB filed Critical University of Science and Technology Beijing USTB
Priority to CN202111227806.0A priority Critical patent/CN114036821A/en
Publication of CN114036821A publication Critical patent/CN114036821A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Computer Hardware Design (AREA)
  • Geometry (AREA)
  • Feedback Control In General (AREA)

Abstract

The invention discloses a thickener control method and device based on a non-deterministic hidden space model, and relates to the technical field of mining intelligent control. The method comprises the following steps: acquiring current operating parameters of a thickener system, wherein the current operating parameters comprise feeding and discharging flow and feeding and discharging concentration; inputting the current operating parameters into a trained non-deterministic discrete time state space model; obtaining mud layer pressure variation distribution of the thickener system based on the current operation parameters and the trained non-deterministic discrete time state space model; and optimizing the input control sequence of the thickener system according to a cross entropy optimization algorithm based on a result obtained by sampling the mud bed pressure variation distribution of the thickener system to obtain an optimal input control sequence of the thickener system, and controlling the thickener system. The method can better represent the complex noise disturbance and the nondeterminiseness of the thickener system, so that the whole set of prediction and control method has better prediction precision and control precision.

Description

Thickener control method and device based on non-deterministic hidden space model
Technical Field
The invention relates to the technical field of mining intelligent control, in particular to a thickener control method and device based on a non-deterministic hidden space model.
Background
The optimization control problem of the complex process industrial system is widely concerned in the fields of industrial informatization and intelligent control. In modern mining technology, a thickener is an important large-scale sedimentation tool, so that tailings particles form high-concentration underflow under the actions of gravity, certain height of mud layer pressure intensity and stirring of a rake frame, and the functions of reducing water and concentrating are achieved. When the thickener is controlled, the mud layer pressure is a core control index of the system, and other key variables of the thickener, such as underflow concentration and mud layer pressure, can be indirectly controlled by controlling the mud layer pressure of the thickener. Because the mud layer pressure and other process monitoring variables such as feeding flow, feeding concentration, discharging flow and mud layer height have complex nonlinear and time-delay relations, and because the thickener system has high operation cost and low operation fault tolerance, the system is similar to a screw thread; yuan M Lin; liuting; lijia; the posture of the patient is moistened; a thickener online control method based on reinforcement learning comprises the following steps: the model-free online learning control method adopted in CN103454176 (P/OL) in China has the problems of cold start and uncertain convergence time, and has certain limitation in the control application of a real thickener system.
With the development of the technology in the industrial field, the model predictive control technology based on the optimal control theory idea is widely applied. A thickener prediction or simulation model is constructed by utilizing system off-line data, and the optimization control of the running parameters of the thickener system is realized by adopting a model-based control method, so that the control idea of the thickener is safer and more effective. Article [ 2 ]
Figure BDA0003314913090000011
F.,Langarica,S.,Díaz,P.,Torres,M.,&Salas,J.C.(2020).Neural Network-Based Model Predictive Control of a Paste Thickener over an Industrial Internet Platform.IEEE Transactions on Industrial Informatics,16(4),2859-2867.https://doi.org/10.1109/TII.2019.2953275]A multi-step prediction model based on an Encoder-Decoder framework is used for predicting future underflow concentration and mud layer pressure variation sequences, and a particle swarm optimization algorithm is used for solving control sequences, but the deterministic time sequence prediction model adopted by the method does not take the nondeterministic characteristic of a thickener system into consideration, so that the prediction accuracy of the model is poor, and the control accuracy of the system is poor.
Disclosure of Invention
The invention provides a method for controlling a thickener system, which aims at solving the problem that the control precision of the system is poor due to the poor prediction precision of a model caused by the non-deterministic characteristic of the thickener system.
In order to solve the technical problems, the invention provides the following technical scheme:
in one aspect, a thickener control method based on a non-deterministic hidden space model is provided, and the method is applied to electronic equipment and comprises the following steps:
and S1, acquiring current operation parameters of the thickener system, wherein the current operation parameters comprise feeding and discharging flow and feeding and discharging concentration.
And S2, inputting the inlet and outlet flow and the inlet and outlet concentration into the trained non-deterministic discrete time state space model.
S3, obtaining the mud layer pressure variation distribution of the thickener system based on the feeding and discharging flow, the feeding and discharging concentration and the trained non-deterministic discrete time state space model.
And S4, optimizing the input control sequence of the thickener system according to a cross entropy optimization algorithm based on the result obtained by sampling from the mud bed pressure variation distribution of the thickener system to obtain the optimal input control sequence of the thickener system, and controlling the thickener system based on the optimal input control sequence.
Optionally, the trained non-deterministic discrete-time state-space model in S2 includes:
s21, acquiring historical operating parameters of the thickener system; the historical operation parameters comprise sample feeding and discharging flow, sample feeding and discharging concentration and sample mud layer pressure.
S22, constructing a non-deterministic discrete time state space model of the thickener system based on the deep neural network containing the hidden variables, inputting the sample feeding and discharging flow, the sample feeding and discharging concentration and the sample mud layer pressure into the non-deterministic discrete time state space model to obtain a reconstructed predicted mud layer pressure, and training the non-deterministic discrete time state space model according to the predicted mud layer pressure and the reconstruction error of the sample mud layer pressure.
Optionally, the obtaining of the historical operating parameters of the thickener system in S21 includes:
and acquiring original operation parameters monitored by each sensor of the thickener system.
And counting the mean value and the variance of each parameter in the original operation parameters, and carrying out normalized scaling on the original operation parameters based on the counted mean value and variance of each parameter to obtain historical operation parameters.
Optionally, the training of the non-deterministic discrete-time state space model in S22 includes: and estimating the gradient of the loss function to the parameters of the non-deterministic discrete time state space model, and after obtaining each gradient, performing optimization training on the non-deterministic discrete time state space model by adopting a random gradient descent method.
Optionally, the non-deterministic discrete-time state space model comprises an a posteriori coding module and an a priori prediction module.
The posterior coding module is used for hidden variable reasoning and coding the historical operation data of the thickener system.
And the prior prediction module is used for implicit variable prior distribution representation to realize prediction of the mud layer pressure of the thickener system.
Optionally, the non-deterministic discrete-time state space model comprising an a posteriori coding module and an a priori prediction module comprises:
based on the variational self-encoder method, an approximate posterior inference model from the observed quantity of the thickener system to the hidden variable of the thickener system is constructed, the lower bound of the variational evidence is used as an optimization target of the approximate posterior inference model, and the approximate posterior inference model is trained and used for an observation posterior coding module and a prior prediction module.
Optionally, the optimizing the input control sequence of the thickener system according to the cross entropy optimization algorithm based on the result obtained by sampling from the mud bed pressure variation distribution of the thickener system in S4 to obtain the optimal input control sequence of the thickener system, and controlling the thickener system based on the optimal input control sequence includes:
and S41, constructing the optimal input control sequence distribution of the initial state, wherein the optimal input control sequence distribution of the initial state obeys Gaussian distribution, and sampling to obtain the optimal input control sequence.
S42, establishing an evaluation function, obtaining the error between the mud layer pressure of the thickener system and a set value and the instability degree of the thickener optimal input control sequence based on the optimal input control sequence and the trained non-deterministic discrete time state space model, and re-estimating the optimal input control sequence distribution according to the evaluation function and the optimal input control sequence obtained by sampling.
And S43, repeating the step S42, and after preset iteration turns, taking the mean value of the optimal input control sequence obtained by final solution as the system action of the thickener system at the next moment.
In another aspect, a thickener control apparatus based on a non-deterministic hidden space model is provided, the apparatus being applied to an electronic device, and the apparatus comprising:
and the data acquisition module is used for acquiring the current operation parameters of the thickener system, wherein the current operation parameters comprise the flow rate of feeding and discharging and the concentration of feeding and discharging.
And the non-deterministic discrete time state space model prediction module is used for inputting the current operation parameters into the trained non-deterministic discrete time state space model.
And the nondeterministic discrete time state space model output module is used for obtaining the mud layer pressure variation distribution of the thickener system based on the current operating parameters and the trained nondeterministic discrete time state space model.
And the optimal input control module is used for optimizing the input control sequence of the thickener system based on the result obtained by sampling from the mud bed pressure variation distribution of the thickener system to obtain the optimal input control sequence of the thickener system.
Optionally, the non-deterministic discrete-time state space model prediction module is further configured to:
a trained non-deterministic discrete-time state-space model comprising:
s21, acquiring historical operating parameters of the thickener system; the historical operation parameters comprise sample feeding and discharging flow, sample feeding and discharging concentration and sample mud layer pressure.
S22, constructing a non-deterministic discrete time state space model of the thickener system based on the deep neural network containing the hidden variables, inputting the sample feeding and discharging flow, the sample feeding and discharging concentration and the sample mud layer pressure into the non-deterministic discrete time state space model to obtain a reconstructed predicted mud layer pressure, and training the non-deterministic discrete time state space model according to the predicted mud layer pressure and the reconstruction error of the sample mud layer pressure.
Optionally, the non-deterministic discrete-time state space model prediction module is further configured to:
the obtaining of the historical operating parameters of the thickener system in the step S21 includes:
and acquiring original operation parameters monitored by each sensor of the thickener system.
And counting the mean value and the variance of each parameter in the original operation parameters, and carrying out normalized scaling on the original operation parameters based on the counted mean value and variance of each parameter to obtain historical operation parameters.
Optionally, the non-deterministic discrete-time state space model prediction module is further configured to:
training the non-deterministic discrete-time state-space model in S22 includes: and estimating the gradient of the loss function to the parameters of the non-deterministic discrete time state space model, and after obtaining each gradient, performing optimization training on the non-deterministic discrete time state space model by adopting a random gradient descent method.
Optionally, the non-deterministic discrete-time state space model prediction module is further configured to:
the non-deterministic discrete time state space model includes an a posteriori coding module and an a priori prediction module.
The posterior coding module is used for hidden variable reasoning and coding the historical operation data of the thickener system.
And the prior prediction module is used for implicit variable prior distribution representation to realize prediction of the mud layer pressure of the thickener system.
Optionally, the non-deterministic discrete-time state space model prediction module is further configured to:
the non-deterministic discrete time state space model comprises a posterior coding module and a prior prediction module, and comprises:
based on the variational self-encoder method, an approximate posterior inference model from the observed quantity of the thickener system to the hidden variable of the thickener system is constructed, the lower bound of the variational evidence is used as an optimization target of the approximate posterior inference model, and the approximate posterior inference model is trained and used for an observation posterior coding module and a prior prediction module.
Optionally, the optimal input control module, further for,
based on the result obtained by sampling from the mud layer pressure variation distribution of the thickener system, the input control sequence of the thickener system is optimized according to the cross entropy optimization algorithm to obtain the optimal input control sequence of the thickener system, and the thickener system is controlled based on the optimal input control sequence, which comprises the following steps:
and S41, constructing the optimal input control sequence distribution of the initial state, wherein the optimal input control sequence distribution of the initial state obeys Gaussian distribution, and sampling to obtain the optimal input control sequence.
S42, establishing an evaluation function, obtaining the error between the mud layer pressure of the thickener system and a set value and the instability degree of the thickener optimal input control sequence based on the optimal input control sequence and the trained non-deterministic discrete time state space model, and re-estimating the optimal input control sequence distribution according to the evaluation function and the optimal input control sequence obtained by sampling.
And S43, repeating the step S42, and after preset iteration turns, taking the mean value of the optimal input control sequence obtained by final solution as the system action of the thickener system at the next moment.
In one aspect, an electronic device is provided, and the electronic device includes a processor and a memory, where the memory stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the above-mentioned non-deterministic hidden space model-based thickener control method.
In one aspect, a computer-readable storage medium is provided, and at least one instruction is stored in the storage medium and loaded and executed by a processor to implement the non-deterministic hidden space model based thickener control method.
The technical scheme of the embodiment of the invention at least has the following beneficial effects:
the invention creatively provides that a non-deterministic hidden space model with randomness is adopted as a prediction model of a thickener system, and the method constructs and learns a non-deterministic hidden variable dynamic model of a thickener system mud layer pressure dynamic change process based on the running data of the thickener system. In the control stage, the trained non-deterministic hidden variable dynamic model is used for predicting the mud layer pressure change of the system in a period of time in the future under the given control input, and the cross entropy algorithm is used for calculating the optimal control input sequence of the future system. Compared with the traditional prediction control method of the thickener based on the deterministic model, the method uses the non-deterministic hidden variable dynamic model with randomness as the prediction model of the thickener system, and can better represent the complex noise disturbance and the non-determinacy of the thickener system, so that the whole set of prediction and control method has better prediction precision and control precision.
Drawings
In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings needed to be used in the description of the embodiments will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.
FIG. 1 is a schematic flow chart of a thickener control method based on a non-deterministic hidden space model according to an embodiment of the present invention;
FIG. 2 is a diagram of a thickener apparatus overview provided by an embodiment of the present invention;
FIG. 3 is a flow chart of a construction based on a non-deterministic hidden space model according to an embodiment of the present invention;
FIG. 4 is a diagram of a non-deterministic hidden space model architecture provided by an embodiment of the present invention;
FIG. 5 is a diagram of a model predictive control framework provided by an embodiment of the present invention;
FIG. 6 is a flow chart of an optimization of an optimal input control sequence for a thickener system according to an embodiment of the present invention;
FIG. 7 is a diagram of a thickener control service based on the http protocol of the python flash service framework according to an embodiment of the present invention;
FIG. 8 is a schematic flow chart of a thickener control device based on a non-deterministic hidden space model according to an embodiment of the present invention;
fig. 9 is a schematic structural diagram of an electronic device according to an embodiment of the present invention.
Detailed Description
In order to make the technical problems, technical solutions and advantages of the present invention more apparent, the following detailed description is given with reference to the accompanying drawings and specific embodiments.
As shown in fig. 1, an embodiment of the present invention provides a thickener control method based on a non-deterministic hidden space model, where the method is applied to an electronic device, and the method includes:
and S1, acquiring current operation parameters of the thickener system, wherein the current operation parameters comprise feeding and discharging flow and feeding and discharging concentration.
And S2, inputting the inlet and outlet flow and the inlet and outlet concentration into the trained non-deterministic discrete time state space model.
Optionally, the trained non-deterministic discrete-time state-space model in S2 comprises:
s21, acquiring historical operating parameters of the thickener system; the historical operation parameters comprise sample feeding and discharging flow, sample feeding and discharging concentration and sample mud layer pressure.
S22, constructing a non-deterministic discrete time state space model of the thickener system based on the deep neural network containing the hidden variables, inputting the sample feeding and discharging flow, the sample feeding and discharging concentration and the sample mud layer pressure into the non-deterministic discrete time state space model to obtain a reconstructed predicted mud layer pressure, and training the non-deterministic discrete time state space model according to the predicted mud layer pressure and the reconstruction error of the sample mud layer pressure.
Optionally, the obtaining of the historical operating parameters of the thickener system in S21 includes:
and acquiring original operation parameters monitored by each sensor of the thickener system.
And counting the mean value and the variance of each parameter in the original operation parameters, and carrying out normalized scaling on the original operation parameters based on the counted mean value and variance of each parameter to obtain historical operation parameters.
Optionally, the training of the non-deterministic discrete-time state-space model in S22 includes: and estimating the gradient of the loss function to the parameters of the non-deterministic discrete time state space model, and after obtaining each gradient, performing optimization training on the non-deterministic discrete time state space model by adopting a random gradient descent method.
Optionally, the non-deterministic discrete-time state space model comprises an a posteriori coding module and an a priori prediction module.
The posterior coding module is used for hidden variable reasoning and coding the historical operation data of the thickener system.
And the prior prediction module is used for implicit variable prior distribution representation to realize prediction of the mud layer pressure of the thickener system.
Optionally, the non-deterministic discrete-time state space model including the a posteriori coding module and the a priori prediction module includes:
based on the variational self-encoder method, an approximate posterior inference model from the observed quantity of the thickener system to the hidden variable of the thickener system is constructed, the lower bound of the variational evidence is used as an optimization target of the approximate posterior inference model, and the approximate posterior inference model is trained and used for an observation posterior coding module and a prior prediction module.
S3, obtaining the mud layer pressure variation distribution of the thickener system based on the feeding and discharging flow, the feeding and discharging concentration and the trained non-deterministic discrete time state space model.
And S4, optimizing the input control sequence of the thickener system according to a cross entropy optimization algorithm based on the result obtained by sampling from the mud bed pressure variation distribution of the thickener system to obtain the optimal input control sequence of the thickener system, and controlling the thickener system based on the optimal input control sequence.
Optionally, the optimizing the input control sequence of the thickener system according to the cross entropy optimization algorithm based on the result obtained by sampling from the mud bed pressure variation distribution of the thickener system in S4 to obtain the optimal input control sequence of the thickener system, and controlling the thickener system based on the optimal input control sequence includes:
and S41, constructing the optimal input control sequence distribution of the initial state, wherein the optimal input control sequence distribution of the initial state obeys Gaussian distribution, and sampling to obtain the optimal input control sequence.
S42, establishing an evaluation function, obtaining the error between the mud layer pressure of the thickener system and a set value and the instability degree of the thickener optimal input control sequence based on the optimal input control sequence and the trained non-deterministic discrete time state space model, and re-estimating the optimal input control sequence distribution according to the evaluation function and the optimal input control sequence obtained by sampling.
And S43, repeating the step S42, and after preset iteration turns, taking the mean value of the optimal input control sequence obtained by final solution as the system action of the thickener system at the next moment.
In a feasible implementation manner, the embodiment of the invention provides a modeling prediction control method of mud layer pressure of a thickener based on a nondeterministic hidden space model aiming at the nondeterministic characteristic of the thickener system, the method comprises the steps of firstly establishing a depth time sequence network consisting of a prior module and a posterior module, and obtaining the nondeterministic hidden space state space model for identifying the thickener system by utilizing offline operation data of the thickener system; the network is used to identify a state space model of the thickener system. On the basis of the state space model, an online control algorithm of the underflow concentration of the thickener based on model predictive control is provided, an evaluation function is established by initializing action distribution, and feedback correction and rolling optimization are performed by using an optimization algorithm, so that the optimal control sequence of the thickener system is calculated. Compared with the traditional model prediction control method using a deterministic model, the thickener mud layer pressure modeling and control method based on the non-deterministic hidden space has higher prediction and control precision, and is more suitable for modeling and controlling a thickener system in an actual production environment.
As shown in fig. 2, the thickener is a typical mud sedimentation separation tool, and is widely used in the process industry fields of metallurgy, mining, petrochemical industry and the like. The upstream section can generate low-concentration slurry with constantly fluctuating concentration and flow. By utilizing the characteristic that the density of silt particles is higher than that of water and the flocculation effect of a flocculating agent, sand particles can be continuously settled and form high-concentration underflow at the bottom of the thickener, and the underflow is sucked into a conveying pipeline under the pressure action of an underflow pump.
When the thickener is controlled, the underflow concentration is a core control index, and the parameter has a complex coupling relation with other process monitoring variables such as feeding flow, feeding concentration, discharging flow and mud layer pressure. And relevant researches show that the change of the mud layer pressure of the thickener in a plurality of parameters influencing the underflow concentration can reflect the change of the underflow concentration to a great extent, namely, the underflow concentration of the thickener can be indirectly stabilized by stabilizing the mud layer pressure of the thickener. Therefore, the mud layer pressure is a core control target of the sedimentation control process of the thickener for the system.
The performance of the dense sedimentation process is evaluated, and the core control index is the mud layer pressure y1This factor is influenced by control inputs, system state parameters and other external noise. The control input is underflow flow u (k), and the system state parameters are mud layer height h (k) and rake rack rotating speed c1(k) And flocculant pump speed c2(k) Is the external noise input. The corresponding relation between the feeding flow rate of the thickener and the concentration of slurry during feeding and discharging exists on the pressure of a mud layer, and the variable can not be manually controlled in the actual production process of the thickener, so the variable is also used as a predictive variable y2,y3,y4∈R3。
According to the above definition, y ═ y1(k),y2(k),y3(k),y4(k)]∈R4For the system control quantity, u (k) epsilon R is a controllable input quantity which is an important parameter for representing the state of the current thickener, and c (k) epsilon [ c [ [ c ]1(k),c2(k)]∈R2For controllable system noise quantity, h (k) is a system state quantity, which is an important parameter characterizing the current state of the thickener and can be indirectly controlled but not be a control target. In the industrial field, the feeding particle size and the feeding components of the thickener can influence the underflow concentration of the thickener, and further influence the mud bed pressure, but because the variables cannot be directly observed and have small fluctuation, the thickener is modeled into the following system for simplifying the problem:
Figure BDA0003314913090000091
wherein the content of the first and second substances,
Figure BDA0003314913090000092
is the sequence of external input variables of the system, including the underflow flow u (k). N is a radical ofxIs the sequence length;
Figure BDA0003314913090000093
is a system state variable comprising mud layer pressure, feeding flow, feeding concentration and underflow concentration. The output y (t +1) is the system state value at the next time.
In the embodiment of the invention, a non-deterministic hidden space model with randomness is creatively provided as a prediction model of a thickener system, and the method is used for constructing and learning the non-deterministic hidden variable dynamic model of the thickener system mud layer pressure dynamic change process based on the running data of the thickener system. In the control stage, the trained non-deterministic hidden variable dynamic model is used for predicting the mud layer pressure change of the system in a period of time in the future under the given control input, and the cross entropy algorithm is used for calculating the optimal control input sequence of the future system. Compared with the traditional prediction control method of the thickener based on the deterministic model, the method uses the non-deterministic hidden variable dynamic model with randomness as the prediction model of the thickener system, and can better represent the complex noise disturbance and the non-determinacy of the thickener system, so that the whole set of prediction and control method has better prediction precision and control precision.
As shown in fig. 3, an embodiment of the present invention provides a building flowchart based on a non-deterministic hidden space model, where a processing flow of the method may include:
s310, obtaining historical operating parameters of the thickener system; the historical operation parameters comprise sample feeding and discharging flow, sample feeding and discharging concentration and sample mud layer pressure.
Optionally, the obtaining of the historical operating parameters of the thickener system in S310 includes steps 3101 and 3102:
s3101, collecting original operation parameters monitored by each sensor of the thickener system.
In one possible embodiment, historical operating parameters of the thickener system monitored by various sensors of the thickener system are obtained: the embodiment of the application realizes data reading of an industrial DCS (Distributed Control System) System through an OPC (Object Linking and Embedding for Process Control) technology, develops a data reading service by using an OpenOPC tool package, deploys the service to a computing terminal, connects the computing terminal with an industrial field DCS Control room OPC server through a network cable, realizes real-time reading of DCS System sensor data, and stores the data in a local MySQL database and transfers the data to a MongoDB database. Wherein the data detected by the sensor is archived once per minute.
S3102, the mean value and the variance of each parameter in the original operation parameters are counted, and the original operation parameters are normalized and scaled based on the counted mean value and variance of each parameter to obtain historical operation parameters.
In a possible implementation manner, in the embodiment of the present application, data of the operation of the thickener is derived from the MongoDB, and data including five monitoring points, i.e., a feed flow rate, a feed concentration, and a mud layer pressure, is recorded in a CSV (Comma-Separated Values) file. Because the value difference of different physical quantities is large, the network can not be effectively learned and the setting of the super-parameters is difficult, so that the initial operation parameters are preprocessed through the numpy toolkit, specifically, the mean value and the variance of the initial operation parameters are normalized, and the specific formula is as follows:
Figure BDA0003314913090000111
wherein x is the original operating parameter, xscaleFor historical operating parameters, xmeanIs the mean value of the original operating parameter, xstdIs the original operating parameter variance.
Dividing the history operation parameter set obtained after preprocessing into three parts: training set (60%), test set (20%), and validation set (20%).
In addition, the extreme values of the parameters in the historical operating parameters of the thickener system need to be counted, and the extreme values are used for constraining the optimal control input sequence obtained by the subsequent calculation so as to meet the fault-tolerant range of the thickener system.
S320, constructing a non-deterministic discrete time state space model of the thickener system based on the deep neural network containing the hidden variables.
Optionally, the non-deterministic discrete time state space model shown in fig. 4 is a structure diagram, and the model predictive control framework diagram shown in fig. 5 is a diagram, and the non-deterministic discrete time state space model includes an a posteriori coding module and an a priori prediction module. Specifically, S3201-S3202-S3203 are included.
Based on the variational self-encoder method, an approximate posterior inference model from the observed quantity of the thickener system to the hidden variable of the thickener system is constructed, the lower bound of the variational evidence is used as an optimization target of the approximate posterior inference model, and the approximate posterior inference model is trained and used for an observation posterior coding module and a prior prediction module.
S3201, the posterior coding module is used for hidden variable reasoning to realize the coding of the historical operating data of the thickener system.
In one possible embodiment, the thickener system is modeled using a non-deterministic discrete-time state-space model, embodied as describing the thickener system using a state-space model containing hidden variables. The transfer process of the hidden variables is a discrete non-deterministic model of the running parameters of the thickener system, the initial probability distribution and the transfer condition probability distribution of the hidden variables are subjected to Gaussian distribution, and a deep neural network is adopted to model the parameters of the probability transfer process.
Specifically, RSSM (recursive State Space Model) is used to perform forward prediction in hidden Space, and this Model can be regarded as a non-linear kalman filter or a sequence VAE (variant automatic encoder). The model includes deterministic branches and stochastic branches. The state of the RSSM model is divided into a random part ztAnd a deterministic portion htThis depends on the random and deterministic part of the RNN (Recurrent Neural Network) that the previous time step passed.
Further, the non-deterministic discrete time state space is used for modeling the thickener system to construct an approximate posterior model from the historical operating parameters x and y of the system to the hidden variable z, and the system input sequence x in the time T of the thickener system is input1:TAnd system output sequence y1:TCoded as a sequence of hidden variables z1:TIn the historical operating parameter sequence acquired from the system, x is the system input, y is the system output, and t is the system operating time.
Based on given system input sequence x1:TAnd system output sequence y1:TObtaining a given system input xtSystem output oftProbability distribution of, i.e. p (y)t|xt). Implicit variable sequence z introduced into system in embodiment of application1：TTo represent the randomness of the system and the long-term impact of the system inputs on the system.
The embodiment of the application dynamically expresses a thickener system as p (y)t|xt)＝p(zt|zt-1,xt-1)p(yt|zt) Wherein p (z)t|zt-1,xt-1) Input x at a given system is describedtLower, hidden variable ztTransition conditional probability distribution of (c), p (y)t|zt) Describing the hidden variable z at a given time in a given systemtNext, the system outputs ytDistribution of (2).
To learn the parameterized model p (z)t|zt-1,xt-1) And p (y)t|zt) To achieve a given system input sequence x1:TAnd predict the system output sequence y1:TAnd implementing a given system output sequence y1:TAnd input sequence x1:TIn this case, the sequence of hidden variables z is estimated1:TThe embodiment of the application adopts a variational Bayes method to introduce the approximate posterior distribution q (z) of the hidden variable1:T|y1:T,x1:T) To approximate the true posterior distribution p (z) of the hidden variable1:T|y1:T,x1:T) And minimizing the KL divergence between the two distributions to make an approximate posterior distribution q (z) of the hidden variable1:T|y1:T,x1:T) Successive approximation to the true posterior distribution p (z) of hidden variables1:T|y1:T,x1:T). In particular, the hidden variable random part ztThe approximate posterior distribution of (a) is as shown in formula (1):
Figure BDA0003314913090000121
an observation sequence representing the past T moments from the system andthe motion sequence approximates and deduces the posterior distribution of hidden variables; wherein q (z)t|zt-1,xt-1,yt) Is a parameterized diagonal gaussian distribution.
Since there is a great uncertainty in the thickener system and the prediction model used is non-linear and cannot directly calculate the required state posteriori using parameter learning, an encoder is used to infer an approximate hidden state posteriori from past observations and actions, taking into account the deterministic portion h of the RSSM modeltImplicit variable deterministic part htThe approximate posterior distribution of (a) is as shown in the following formula (2):
Figure BDA0003314913090000122
wherein q (z)t|ht-1,yt) For a mean value μ parameterized by a neural network and a feedforward neural network0Variance is σ0,tIs of a diagonal gaussian distribution zt～N(μ0,t,diag(σ0,t))。
S3202, inputting the sample feeding and discharging flow, the sample feeding and discharging concentration and the sample mud layer pressure into the non-deterministic discrete time state space model to obtain reconstructed predicted mud layer pressure, and training the non-deterministic discrete time state space model according to the predicted mud layer pressure and the reconstruction error of the sample mud layer pressure.
Optionally, the training of the non-deterministic discrete-time state space model in S3202 includes: and estimating the gradient of the loss function to the parameters of the non-deterministic discrete time state space model, and after obtaining each gradient, performing optimization training on the non-deterministic discrete time state space model by adopting a random gradient descent method.
In a feasible implementation mode, an ELBO (Evidence Lower Bound) of observation data likelihood is solved by using a variational Bayesian method, and the ELBO is used as an optimization target of a model to train a model posterior coding module and a system dynamic prior prediction module.
Thus, an ELBO of log-likelihood of the observed data can be established, as shown in equation (3):
Figure BDA0003314913090000131
the standard variation lower bound comprises a reconstruction term aiming at an observed variable and a KL divergence regular term approximating posterior distribution of an implicit variable and prior distribution of the implicit variable, and a network p (y) is generated through the KL divergence regular termt|zt) Calculating the conditional generation distribution of x, which generates the log-likelihood lnp of the distribution (y)t|zt) The larger the size, the better the reconstruction of the model, and the need to ensure an approximate posterior distribution q (z) in order to enable the model to generate new samplest|y≤t,x≤t) With a known prior distribution p (z)t|zt-1,xt-1) As close as possible, i.e., the KL divergence regularization term of the second term of the formula. The two formulas together serve as the optimization target of the prediction model. Because the information of a plurality of time steps is difficult to be reliably memorized by a conversion model due to pure random conversion, the deterministic hidden variable sequence is additionally introduced into the model by the embodiment of the application
Figure BDA0003314913090000132
So that the deterministic hidden variable and the stochastic hidden variable are simultaneously transmitted in the model, and the equations (4) and (5) represent the deterministic hidden variable h in the systemtRandom hidden variable ztAnd the association between them:
ht＝f(ht-1,zt-1,xt-1) (4)
zt～p(zt|ht) (5)
where f is a basic recurrent neural network. Because of the large non-determinism of the thickener system and the non-linear prediction model used, the required state posteriori cannot be directly calculated. The encoder is used to parameterize the approximate state a posteriori. In order to enable the gradients used to optimize the model parameters during training to be propagated back into the computational graph for multi-prediction, i.e. the gradients flow through p (z)t|zt-1,xt-1) After can pass through zt-1Proceed to p (z)t-1|zt-2,xt-2) And a number of previously predicted samples zt-i，
In the training process of the model, the input of the module is the input x of the historical system of the thickenertOutput y from the history systemtAnd historical system input xtCoded latent variable ztAnd a hidden state h (t-1) output by the a posteriori coding module at the last moment, and further explaining the formula (4) by using a formula (6), wherein the formula (6) is used for updating the hidden state of the system and updating the network parameters:
Figure BDA0003314913090000141
wherein f isθAn RNN network based on
Figure BDA0003314913090000142
The feature extractors, which can be regarded as x and z, are all fully connected networks of a two-layer structure, in which
Figure BDA0003314913090000143
The number of input nodes is 1, the underflow flow of the thickener is shown, the number of hidden state nodes in the middle layer is 32, the number of hidden layers is 1, and the number of output nodes is 32;
Figure BDA0003314913090000144
the number of the input nodes is 4, the input nodes comprise the mud layer pressure intensity, the underflow concentration, the feeding concentration and the feeding flow rate of the thickener, the number of the hidden state nodes in the middle layer is 32, the number of the hidden layer layers is 1, and the number of the output nodes is 32. Specifically, the network output layer activation function is a tanh function.
In the experimental process of the embodiment of the application, the length N of the historical data trained by the posterior prediction module is 160, and the RNN network fθThe number of input nodes of (1) is 96, the number of output nodes is 32, the number of layers is 1, and theta is a network parameter of the network.
Therefore, the embodiment of the present application extends the lower bound of the standard variation to the lower bound of the evidence of the multi-part prediction in step D, as shown in the following formula (7):
Figure BDA0003314913090000145
wherein
Figure BDA0003314913090000146
Is a weighting factor used to adjust whether the decision is more focused on short-term prediction or long-term prediction. D is the step length of multiple predictions, D is the step length of each iteration in the training process (D is gradually increased from 1 to D), and the prediction performance of the model with D being 1,3 and 5 is tested in the experimental process. And determining the model and a standard variation lower bound for optimization, and further, representing each probability function and a deterministic function in the non-deterministic discrete time state space model by using a plurality of deep neural network parameterizations. Training the model by utilizing the acquired and preprocessed training data; and testing the prediction effect of the verification model by using the verification set data.
In a feasible implementation mode, in the model verification stage, after the system hidden state is updated, the system hidden state is decoded by the observation variable state decoder to obtain the reconstruction of the observation variable of the thickener, and the error between the observation variable and the historical operating parameter of the thickener is calculated to observe the prediction effect of the model. Specifically, the Loss adopted in the model verification process is a Relative square Root Error (Root Relative Squared Error RRSE) of a real value and a predicted value of the control target of the thickener, and a calculation formula of the Loss is shown as follows (8):
Figure BDA0003314913090000151
wherein P is(ij)Is the predicted value of model i for variable j (among n variables); t isjIs the target value of variable j;
Figure BDA0003314913090000152
then given by the following equation (9):
Figure BDA0003314913090000153
and S3203, the prior prediction module is used for implicit variable prior distribution representation to realize prediction of the mud layer pressure of the thickener system.
In one possible embodiment, the a priori prediction module construction includes: the prior coding module inputs the system input x of the thickener system for a period of time in the future according to the model parameters trained by the model posterior coding module in the stept+LLatent variable z encoded from system inputt+LAnd the hidden layer h of the thickener system at the current momenttOutputting the predicted hidden space state h of the future L-length thickener system by the modelt+LAs shown in the following formula (10):
Figure BDA0003314913090000154
the network part is basically the same as the posterior coding module, and finally the obtained thickener is predicted to be hidden in the state h by using an observation variable decodert+LAnd decoding the operation parameters of the thickener to obtain the prediction target of the model.
Further, the trained model is used for the prediction of the thickener system, and can be specifically expressed as the following two processes of formula (11) (12):
Figure BDA0003314913090000155
Figure BDA0003314913090000156
where τ is the backward prediction time, and (T: T + τ) denotes the backward prediction τ from time T.
Figure BDA0003314913090000157
As a result of the predicted future time τ. Formula (11) represents that a posteriori coding module is adopted to code input and output data of a system of the thickener system in the past period of time, and an initial state of a hidden variable of the thickener system at the T moment is obtained through prediction and sampling, wherein the transfer process of the hidden variable in a hidden space is discrete and non-deterministic, and the hidden variable z at an initial position is0The probability distribution and the transition condition probability distribution are Gaussian distributions coded by using a thickener system historical operation parameter sequence.
Equation (12) represents the prediction process using the system a priori prediction module, with the input including the hidden state of the system obtained using the encoder
Figure BDA0003314913090000161
Thickener system input xT+τAnd the output is the system output under the input of a given system, the external input data of the thickener system at the current moment is coded to obtain the hidden variable of the system at the current moment, and the decoder in the model is combined to obtain the system output under the current input of the system. Wherein the system output includes the mud layer pressure and the underflow concentration.
Specifically, in the present embodiment, training epochs is 800, batch size is 1024, learning rate is 0.0005, attenuation rate is 0.98, attenuation step number is 10, and training is performed using SGD (stochastic gradient descent) back propagation. Model implementation and training are completed by using a pyrrch framework, and after training, the model is saved as a _.
In the embodiment of the invention, a non-deterministic hidden space model with randomness is creatively provided as a prediction model of a thickener system, and the method is used for constructing and learning the non-deterministic hidden variable dynamic model of the thickener system mud layer pressure dynamic change process based on the running data of the thickener system. In the control stage, the trained non-deterministic hidden variable dynamic model is used for predicting the mud layer pressure change of the system in a period of time in the future under the given control input, and the cross entropy algorithm is used for calculating the optimal control input sequence of the future system. Compared with the traditional prediction control method of the thickener based on the deterministic model, the method uses the non-deterministic hidden variable dynamic model with randomness as the prediction model of the thickener system, and can better represent the complex noise disturbance and the non-determinacy of the thickener system, so that the whole set of prediction and control method has better prediction precision and control precision.
As shown in fig. 6, an embodiment of the present invention provides a flowchart for optimizing an optimal input control sequence of a thickener system, where a processing flow of the method may include:
s610, determining an optimization function: according to the control requirements of the thickener, the control sequence of the thickener is as stable as possible, and the underflow concentration is stable at the process set value of the thickener. The optimization function designed here is shown in the following equation (13):
Figure BDA0003314913090000171
wherein the content of the first and second substances,
Figure BDA0003314913090000172
is an estimate of the output of the predictive model,
Figure BDA0003314913090000173
n represents the dimension of the state value as the manual setting value of the control target;
Figure BDA0003314913090000174
indicating that the system outputs an artificial set point as close to the control target as possible. Δ ukRepresenting the action variation for the optimization goal;
Figure BDA0003314913090000175
means that the motion variation is made as small as possible by optimization;
Figure BDA0003314913090000176
the method is a penalty item, and in order to ensure that the calculated amount of the action is within the normal orientation, once the calculated amount of the action exceeds the range, the value of the penalty item is increased, so that the aim of controlling the action variation range of the system is fulfilled.
In a feasible implementation mode, in the aspect of control, on the basis of a trained non-deterministic hidden space state space model, the mud layer pressure change of the system is predicted according to the current operating parameters, and according to the change prediction result, the input control sequence of the thickener system is optimized by using a CEM (Cross Entropy optimization algorithm) to obtain the optimal input control sequence of the thickener system.
Because the experiment cost for controlling the thickener in a real industrial scene is higher, the experiments carried out in the embodiment of the application are all simulation experiments, and the effectiveness of the control algorithm is verified through the simulation experiments.
S620, selecting an optimization algorithm: for such a non-deterministic prediction model currently used, the Optimization algorithm based on gradient and the conventional PSO algorithm (Particle Swarm Optimization) can make the Optimization time too long, so a cross entropy Optimization algorithm is selected, which predicts the distribution of the action sequence in a short time in the future, is more suitable for matching a stochastic prediction model and taking the stochastic prediction model as a transfer model to perform Optimization control on the system, and is therefore more suitable for being used as an Optimization algorithm of a thickener system.
S630, the optimization process comprises: optimizing an input control sequence of the thickener system according to a cross entropy optimization algorithm based on a result obtained by sampling from mud layer pressure variation distribution of the thickener system to obtain an optimal input control sequence of the thickener system, and controlling the thickener system based on the optimal input control sequence, wherein the method specifically comprises the following steps of S6301-S6303:
and S6301, constructing optimal input control sequence distribution in the initial state, wherein the optimal input control sequence distribution in the initial state obeys Gaussian distribution, and sampling to obtain the optimal input control sequence.
In one possible embodiment, initializing the optimal input control sequence distribution includes: and (3) constructing the optimal input control sequence distribution in an initial state by the optimal control sequence distribution obtained by current solution (each position of the sequence distribution is in standard normal distribution under the initial condition), and sampling a large number of control sequences. And under the control input sequence obtained by predicting each sample according to the non-deterministic discrete time state space model, the future change distribution of partial key variables in the thickener system is obtained, namely the optimal input control sequence of the system is obtained by sampling.
And S6302, constructing an evaluation function, obtaining an error between the mud layer pressure of the thickener system and a set value and the instability degree of the thickener optimal input control sequence based on the optimal input control sequence and the trained non-deterministic discrete time state space model, and re-estimating the distribution of the optimal input control sequence according to the evaluation function and the optimal input control sequence obtained by sampling.
And S6303, repeating the step S6302, and taking the mean value of the optimal input control sequence distribution obtained by final solution as the optimal input control sequence of the thickener system after preset iteration turns.
In one possible embodiment, the input control sequence includes underflow flow, underflow concentration, feed flow, and feed concentration; the optimization algorithm is a cross entropy optimization algorithm. The algorithm method comprises the following execution flows:
construction of optimal input control sequence distribution A of current thickener system obeying Gaussian distribution using mean and variance of original operation parameters of thickener (a)t:t+H) Initialize the action confidence matrix A (a)t:t+H) And (c) either oid (0,1), which generates J motion sequences from the system motion confidence level distribution samples
Figure BDA0003314913090000181
Predicting system states over a period of time in the future by a predictive model
Figure BDA0003314913090000182
Sampling to obtain j control input sequences with length of H
Figure BDA0003314913090000183
Predicting the running track of each control action sequence in the system H time by using a non-deterministic discrete time state space model
Figure BDA0003314913090000184
As shown in equation (14). Where o is the system observation.
Figure BDA0003314913090000185
Wherein q(s)t|o1:t,a1:t-1) For the system state s at time ttRepresents the ratio of time 1: systematic observation of t and 1: system action a within t-11:t-1The obtained system state s at the moment tt。
Figure BDA0003314913090000186
Representing the state s of the system at a future time ττRepresents the state s of the system at the previous moment of useτ-1And system actions in the jth action sequence at the previous time
Figure BDA0003314913090000187
Predicted to obtain sτ。
Figure BDA0003314913090000188
Representing multiplication by sτA priori predicts the system state of the H length
Figure BDA0003314913090000189
The error between the system output predicted from the sampled action sequence and its artificial set value is calculated by an evaluation function expression (15), evaluation values of a plurality of input control sequences are given, and then sorted from small to large according to the evaluation values, and the optimal control input distributions (16), (17) are updated using the first K control input sequences with the smallest evaluation values.
Figure BDA00033149130900001810
Figure BDA00033149130900001811
Figure BDA00033149130900001812
Wherein R is(j)The rating value (or rewarded reward) for the jth action sequence, rτAn evaluation value at time τ is represented,
Figure BDA0003314913090000191
indicating the system state at time instant t of the j-th motion sequence.
μt:t+HIs the mean, σ, of the first K control input sequencest:t+HIs its standard deviation.
Figure BDA0003314913090000192
Represents the mean value of μt:t+HVariance is σt:t+HIs sampled from the control input sequence a (a)t:t+H)
And after I iterations, the mean value mu of the optimal control input sequence obtained by sampling the motion sequence distribution obtained by optimization is used as the system input quantity to act on the thickener system.
Specifically, the mean value μ of the first action sequence in the first K action sequences is output as the system action after I iterations. And in the simulation experiment process of the thickener, the iteration number I is 50, and K is 10. In particular, for said thickener system modeled using a non-deterministic discrete-time state-space model, the observed variable of the system, i.e. s in the algorithmtThe action variables a of the system are mud layer pressure, feeding flow, feeding concentration and discharging concentrationtIs the discharge flow.
In the embodiment of the invention, a non-deterministic hidden space model with randomness is creatively provided as a prediction model of a thickener system, and the method is used for constructing and learning the non-deterministic hidden variable dynamic model of the thickener system mud layer pressure dynamic change process based on the running data of the thickener system. In the control stage, the trained non-deterministic hidden variable dynamic model is used for predicting the mud layer pressure change of the system in a period of time in the future under the given control input, and the cross entropy algorithm is used for calculating the optimal control input sequence of the future system. Compared with the traditional prediction control method of the thickener based on the deterministic model, the method uses the non-deterministic hidden variable dynamic model with randomness as the prediction model of the thickener system, and can better represent the complex noise disturbance and the non-determinacy of the thickener system, so that the whole set of prediction and control method has better prediction precision and control precision.
As shown in fig. 7, an embodiment of the present invention provides a thick machine control service based on the http protocol of the python flash service framework. The frame comprises a thickener mud layer pressure controller and a flash _ service server based on python.
Specifically, the mud layer pressure controller of the thickener comprises a data transmitter, a data receiver and a state controller, wherein the data transmitter transmits the current system state of the thickener system to the flash server through an http protocol, and transmits a post request to call each function module of the flash _ service section. And the data receiver receives system data updated by the thickener system received by the flash server and the optimal control input of the thickener obtained by the thickener mud layer pressure modeling and control method based on the nondeterministic hidden space model, and the optimal control input of the thickener controls the mud layer pressure of the thickener by the state controller. Wherein the state controller is an underflow flow controller of the thickener.
Specifically, the flash _ service server based on python comprises a state updating module and an optimization algorithm module, receives a module calling instruction calling module sent by a thickener mud layer pressure controller through a receiver, and sends the return output of each module to the thickener mud layer pressure controller.
Specifically, the state updating module updates the hidden state of the thickener system through an RSSM model posterior coding module trained by using the historical operating parameters of the thickener, and returns the updated hidden state of the thickener system and the current data of the system.
Specifically, the optimization algorithm module performs control optimization according to the hidden state of the thickener system and the current state of the system by using an optimization algorithm and returns the optimal control input for the thickener system at the next moment.
In the embodiment of the invention, a non-deterministic hidden space model with randomness is creatively provided as a prediction model of a thickener system, and the method is used for constructing and learning the non-deterministic hidden variable dynamic model of the thickener system mud layer pressure dynamic change process based on the running data of the thickener system. In the control stage, the trained non-deterministic hidden variable dynamic model is used for predicting the mud layer pressure change of the system in a period of time in the future under the given control input, and the cross entropy algorithm is used for calculating the optimal control input sequence of the future system. Compared with the traditional prediction control method of the thickener based on the deterministic model, the method uses the non-deterministic hidden variable dynamic model with randomness as the prediction model of the thickener system, and can better represent the complex noise disturbance and the non-determinacy of the thickener system, so that the whole set of prediction and control method has better prediction precision and control precision.
As shown in fig. 8, an embodiment of the present invention provides a thickener control apparatus 800 based on a non-deterministic hidden space model, where the apparatus 800 is used to implement the above thickener control method based on the non-deterministic hidden space model, and the apparatus 800 includes:
the data acquisition module 810 is configured to acquire current operating parameters of the thickener system, where the current operating parameters include a feed/discharge flow rate and a feed/discharge concentration;
a prediction module 820 for inputting current operating parameters into the trained non-deterministic discrete time state space model;
the output module 830 is configured to obtain mud layer pressure variation distribution of the thickener system based on the current operating parameters and the trained non-deterministic discrete time state space model;
the optimal input control module 840 is used for optimizing the input control sequence of the thickener system based on the result obtained by sampling from the mud layer pressure variation distribution of the thickener system to obtain the optimal input control sequence of the thickener system.
Optionally, the apparatus 800 further comprises a training module 850;
the training module 850 is configured to:
acquiring historical operating parameters of a thickener system; the historical operation parameters comprise sample feeding and discharging flow, sample feeding and discharging concentration and sample mud layer pressure;
the method comprises the steps of constructing a non-deterministic discrete time state space model of a thickener system based on a deep neural network containing hidden variables, inputting sample feeding and discharging flow, sample feeding and discharging concentration and sample mud layer pressure into the non-deterministic discrete time state space model to obtain reconstructed predicted mud layer pressure, and training the non-deterministic discrete time state space model according to the predicted mud layer pressure and reconstruction errors of the sample mud layer pressure.
Optionally, the prediction module 820 is further configured to:
collecting original operation parameters monitored by each sensor of a thickener system;
and counting the mean value and the variance of each parameter in the original operation parameters, and carrying out normalized scaling on the original operation parameters based on the counted mean value and variance of each parameter to obtain historical operation parameters.
Optionally, the training module 850 is further configured to:
when the non-deterministic discrete time state space model is trained, the gradient of the loss function to the parameters of the non-deterministic discrete time state space model is estimated, and after each gradient is obtained, the random gradient descent method is adopted to carry out optimization training on the non-deterministic discrete time state space model.
Optionally, the prediction module 820 is further configured to:
based on the variational self-encoder method, an approximate posterior inference model from the observed quantity of the thickener system to the hidden variable of the thickener system is constructed, the lower bound of the variational evidence is used as an optimization target of the approximate posterior inference model, and the approximate posterior inference model is trained and used for an observation posterior coding module and a prior prediction module.
Optionally, the optimal input control module 840, further operable,
s81, constructing optimal input control sequence distribution of an initial state, wherein the optimal input control sequence distribution of the initial state obeys Gaussian distribution, and sampling to obtain an optimal input control sequence;
s82, constructing an evaluation function, obtaining the error between the mud layer pressure of the thickener system and a set value and the instability degree of the thickener optimal input control sequence based on the optimal input control sequence and a trained non-deterministic discrete time state space model, and re-estimating the distribution of the optimal input control sequence according to the evaluation function and the optimal input control sequence obtained by sampling;
and S83, repeating the step S82, and after a certain iteration turns, taking the mean value of the optimal input control sequence obtained by final solution as the system action of the thickener system at the next moment.
In the embodiment of the invention, a non-deterministic hidden space model with randomness is creatively provided as a prediction model of a thickener system, and the method is used for constructing and learning the non-deterministic hidden variable dynamic model of the thickener system mud layer pressure dynamic change process based on the running data of the thickener system. In the control stage, the trained non-deterministic hidden variable dynamic model is used for predicting the mud layer pressure change of the system in a period of time in the future under the given control input, and the cross entropy algorithm is used for calculating the optimal control input sequence of the future system. Compared with the traditional prediction control method of the thickener based on the deterministic model, the method uses the non-deterministic hidden variable dynamic model with randomness as the prediction model of the thickener system, and can better represent the complex noise disturbance and the non-determinacy of the thickener system, so that the whole set of prediction and control method has better prediction precision and control precision.
As shown in fig. 9, a schematic structural diagram of an electronic device 900 according to an embodiment of the present invention, where the electronic device 900 may generate relatively large differences due to different configurations or performances, and may include one or more processors (CPUs) 901 and one or more memories 902, where the memory 902 stores at least one instruction, and the at least one instruction is loaded and executed by the processors 901 to implement the following steps of the non-deterministic hidden space model-based method for controlling a thickener:
acquiring current operating parameters of a thickener system, wherein the current operating parameters comprise feeding and discharging flow and feeding and discharging concentration;
inputting the flow and concentration of the feeding and discharging materials into a trained non-deterministic discrete time state space model;
obtaining the mud layer pressure variation distribution of the thickener system based on the feeding and discharging flow, the feeding and discharging concentration and the trained nondeterministic discrete time state space model;
optimizing an input control sequence of the thickener system according to a cross entropy optimization algorithm based on a result obtained by sampling in the mud layer pressure variation distribution of the thickener system to obtain an optimal input control sequence of the thickener system, and controlling the thickener system based on the optimal input control sequence.
It will be understood by those skilled in the art that all or part of the steps for implementing the above embodiments may be implemented by hardware, or may be implemented by a program instructing relevant hardware, where the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a magnetic disk or an optical disk, etc.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.

Claims (10)

1. A thickener control method based on a non-determinacy hidden space model is characterized by comprising the following steps:
s1, obtaining current operation parameters of the thickener system, wherein the current operation parameters comprise feeding and discharging flow and feeding and discharging concentration;
s2, inputting the feeding and discharging flow and the feeding and discharging concentration into a trained non-deterministic discrete time state space model;
s3, obtaining the mud layer pressure variation distribution of the thickener system based on the feeding and discharging flow, the feeding and discharging concentration and the trained non-deterministic discrete time state space model;
s4, optimizing the input control sequence of the thickener system according to a cross entropy optimization algorithm based on the result obtained by sampling from the mud bed pressure variation distribution of the thickener system to obtain the optimal input control sequence of the thickener system, and controlling the thickener system based on the optimal input control sequence.
2. The non-deterministic hidden space model based thickener control method according to claim 1, wherein the trained non-deterministic discrete time state space model in S2 comprises:
s21, acquiring historical operating parameters of the thickener system; the historical operation parameters comprise sample feeding and discharging flow, sample feeding and discharging concentration and sample mud layer pressure;
s22, constructing a non-deterministic discrete time state space model of the thickener system based on a deep neural network containing hidden variables, inputting the sample feeding and discharging flow, the sample feeding and discharging concentration and the sample mud layer pressure into the non-deterministic discrete time state space model to obtain a reconstructed predicted mud layer pressure, and training the non-deterministic discrete time state space model according to the predicted mud layer pressure and the reconstruction error of the sample mud layer pressure.
3. The non-deterministic hidden space model based thickener control method according to claim 2, wherein the obtaining of the historical operating parameters of the thickener system in S21 comprises:
collecting original operation parameters monitored by each sensor of the thickener system;
and counting the mean value and the variance of each parameter in the original operation parameters, and carrying out normalized scaling on the original operation parameters based on the counted mean value and variance of each parameter to obtain the historical operation parameters.
4. The non-deterministic hidden space model based thickener control method according to claim 2, wherein the training of the non-deterministic discrete time state space model in S22 comprises: estimating the gradient of the loss function to the nondeterministic discrete time state space model parameters, and after obtaining each gradient, performing optimization training on the nondeterministic discrete time state space model by adopting a random gradient descent method.
5. The non-deterministic hidden space model based thickener control method according to claim 1, wherein the non-deterministic discrete time state space model comprises a posterior coding module and a prior prediction module;
the posterior coding module is used for hidden variable reasoning to realize the coding of the historical operating data of the thickener system;
and the prior prediction module is used for representing the prior distribution of the hidden variable to realize the prediction of the mud layer pressure of the thickener system.
6. The non-deterministic hidden space model based thickener control method according to claim 5, wherein the non-deterministic discrete time state space model comprising a posteriori coding module and a priori prediction module comprises:
based on a variational self-encoder method, an approximate posterior inference model from the observed quantity of the thickener system to the hidden variable of the thickener system is constructed, and the lower bound of variational evidence is used as an optimization target of the approximate posterior inference model to train the approximate posterior inference model for observing the posterior encoding module and the prior prediction module.
7. The method for controlling the thickener based on the non-deterministic hidden space model according to claim 1, wherein the step S4 is to optimize the input control sequence of the thickener system according to a cross entropy optimization algorithm based on the result obtained by sampling the mud pressure variation distribution of the thickener system to obtain the optimal input control sequence of the thickener system, and the step S comprises the following steps:
s41, constructing optimal input control sequence distribution of an initial state, wherein the optimal input control sequence distribution of the initial state obeys Gaussian distribution, and sampling to obtain the optimal input control sequence;
s42, constructing an evaluation function, obtaining an error between the mud layer pressure of the thickener system and a set value and the instability degree of the thickener optimal input control sequence based on the optimal input control sequence and the trained non-deterministic discrete time state space model, and re-estimating the distribution of the optimal input control sequence according to the evaluation function and the optimal input control sequence obtained by sampling;
and S43, repeating the step S42, and after preset iteration turns, taking the average value of the optimal input control sequence obtained by final solution as the system action of the thickener system at the next moment.
8. A thickener control device based on a non-deterministic hidden space model, which is characterized by comprising:
the data acquisition module is used for acquiring current operation parameters of the thickener system, wherein the current operation parameters comprise feeding and discharging flow and feeding and discharging concentration;
the non-deterministic discrete time state space model prediction module is used for inputting the current operation parameters into a trained non-deterministic discrete time state space model;
a non-deterministic discrete time state space model output module for obtaining the mud layer pressure variation distribution of the thickener system based on the current operating parameters and the trained non-deterministic discrete time state space model;
and the optimal input control module is used for optimizing the input control sequence of the thickener system according to a cross entropy optimization algorithm based on a result obtained by sampling in the mud bed pressure variation distribution of the thickener system to obtain the optimal input control sequence of the thickener system.
9. The non-deterministic hidden space model based thickener control device according to claim 8, wherein the trained non-deterministic discrete time state space model comprises:
acquiring historical operating parameters of the thickener system; the historical operation parameters comprise sample feeding and discharging flow, sample feeding and discharging concentration and sample mud layer pressure;
constructing a non-deterministic discrete time state space model of a thickener system based on a deep neural network containing hidden variables, inputting the sample feeding and discharging flow and the sample feeding and discharging concentration into the non-deterministic discrete time state space model to obtain a predicted mud layer pressure, and training the non-deterministic discrete time state space model according to the predicted mud layer pressure and the sample mud layer pressure.
10. The non-deterministic hidden space model based thickener control device according to claim 8, wherein the non-deterministic discrete time state space model comprises a posterior coding module and an a priori prediction module;
the posterior coding module is used for hidden variable reasoning to realize the coding of the historical operating data of the thickener system;
and the prior prediction module is used for representing the prior distribution of the hidden variable to realize the prediction of the mud layer pressure of the thickener system.
CN202111227806.0A 2021-10-21 2021-10-21 Thickener control method and device based on non-deterministic hidden space model Pending CN114036821A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111227806.0A CN114036821A (en) 2021-10-21 2021-10-21 Thickener control method and device based on non-deterministic hidden space model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111227806.0A CN114036821A (en) 2021-10-21 2021-10-21 Thickener control method and device based on non-deterministic hidden space model

Publications (1)

Publication Number Publication Date
CN114036821A true CN114036821A (en) 2022-02-11

Family

ID=80141668

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111227806.0A Pending CN114036821A (en) 2021-10-21 2021-10-21 Thickener control method and device based on non-deterministic hidden space model

Country Status (1)

Country Link
CN (1) CN114036821A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115492082A (en) * 2022-09-28 2022-12-20 中交一公局第七工程有限公司 Composite foundation treatment method, equipment and application for deep soft foundation

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115492082A (en) * 2022-09-28 2022-12-20 中交一公局第七工程有限公司 Composite foundation treatment method, equipment and application for deep soft foundation
CN115492082B (en) * 2022-09-28 2023-09-19 中交一公局第七工程有限公司 Composite foundation treatment method, equipment and application for deep soft foundation

Similar Documents

Publication Publication Date Title
JP2753460B2 (en) Process operation support method and system
CN108764540B (en) Water supply network pressure prediction method based on parallel LSTM series DNN
CN113837356B (en) Intelligent sewage treatment prediction method based on fused neural network
CN112966432A (en) Method and device for predicting remaining effective life of lithium ion battery
CN111815053B (en) Prediction method and system for industrial time sequence data
CN110889085A (en) Intelligent wastewater monitoring method and system based on complex network multiple online regression
CN110309537B (en) Intelligent health prediction method and system for aircraft
CN110824914B (en) Intelligent wastewater treatment monitoring method based on PCA-LSTM network
CN113837364B (en) Sewage treatment soft measurement method and system based on residual network and attention mechanism
CN112735541A (en) Sewage treatment water quality prediction method based on simple circulation unit neural network
CN115935834A (en) History fitting method based on deep autoregressive network and continuous learning strategy
CN114036821A (en) Thickener control method and device based on non-deterministic hidden space model
Yuan et al. Continuous-time prediction of industrial paste thickener system with differential ODE-net
CN112784479B (en) Flood flow prediction method
CN114119277A (en) Artificial intelligent neural network-based flocculation dosing decision analysis method
KR20220150037A (en) System and method for predicting water quality of each process of water treatment facility
CN113111576A (en) Mixed coding particle swarm-long and short term memory neural network based soft measurement method for ammonia nitrogen in effluent
CN112445136B (en) Thickener prediction control method and system based on continuous time neural network
CN116736130A (en) Lithium battery residual service life prediction method and system
Oveisi et al. LSTM Encoder-Decoder Dropout Model in Software Reliability Prediction.
CN115482877A (en) Fermentation process soft measurement modeling method based on time sequence diagram network
Nayak et al. Water quality time-series modeling and forecasting techniques
Warrington et al. Coping with simulators that don’t always return
CN111626472A (en) Scene trend judgment index computing system and method based on deep hybrid cloud model
CN115879569B (en) Online learning method and system for IoT observation data

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination