A thickener control method and device based on non-deterministic latent space model

Abstract

The present invention discloses a thickener control method and device based on a non-deterministic latent space model, relating to the field of intelligent mining control technology. The method comprises: obtaining the current operating parameters of the thickener system, including the inlet and outlet flow rates and the inlet and outlet concentrations; inputting the current operating parameters into a trained non-deterministic discrete-time state-space model; obtaining the distribution of mud layer pressure changes in the thickener system based on the current operating parameters and the trained non-deterministic discrete-time state-space model; optimizing the input control sequence of the thickener system according to a cross-entropy optimization algorithm based on the results sampled from the mud layer pressure change distribution of the thickener system, obtaining the optimal input control sequence of the thickener system, and controlling the thickener system. The present invention can better represent the complex noise disturbances and non-determinism of the thickener system, so that the entire prediction and control method has better prediction accuracy and control accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of mining intelligent control, and in particular to a thickener control method and device based on a non-deterministic latent space model. Background Art

[0002] The optimization and control of complex process industrial systems has garnered widespread attention in the fields of industrial informatization and intelligent control. In modern mining technology, thickeners are important large-scale sedimentation tools. They create a high-concentration underflow of tailings particles under the influence of gravity, a certain level of mud pressure, and rake agitation, thereby reducing water and concentrating the tailings. When controlling thickeners, mud pressure is a core control metric. Controlling thickener pressure indirectly controls other key thickener variables, such as underflow concentration and mud pressure. Since mud layer pressure has complex nonlinear and time-delay relationships with other process monitoring variables such as feed flow rate, feed concentration, discharge flow rate, and mud layer height, and since thickener systems have high operating costs and low operational fault tolerance, the model-free online learning control method used in [Ban Xiaojuan; Yuan Zhaolin; Liu Ting; Li Jia; He Runzi; A method for online control of a thickener based on reinforcement learning: China, CN103454176[P / OL]] suffers from cold start and uncertain convergence time, and has certain limitations in real thickener system control applications.

[0003] With the development of industrial technology, model predictive control technology based on optimal control theory has been widely used. Using system offline data to build a thickener prediction or simulation model and using model-based control methods to achieve optimal control of thickener system operating parameters is a safer and more effective thickener control method. F., Langarica, S., Díaz, P., Torres, M., & Salas, JC (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 the Encoder-Decoder architecture was used to predict the future bottom flow concentration and mud layer pressure change sequence, and the particle swarm optimization algorithm was used to solve the control sequence. However, the deterministic time series prediction model used in this method did not take into account the non-deterministic characteristics of the thickener system itself, which resulted in poor prediction accuracy of the model, resulting in poor control accuracy of the system. Summary of the Invention

[0004] The present invention addresses the problem that the non-deterministic characteristics of the thickener system itself result in poor prediction accuracy of the model and poor control accuracy of the system.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0006] In one aspect, a thickener control method based on a non-deterministic latent space model is provided, the method being applied to an electronic device, the method comprising:

[0007] S1. Obtain the current operating parameters of the thickener system, including the inlet and outlet flow rates and the inlet and outlet concentrations.

[0008] S2. Input the inlet and outlet flow rates and inlet and outlet concentrations into the trained non-deterministic discrete-time state-space model.

[0009] S3. Based on the inlet and outlet flow rates, inlet and outlet concentrations, and the trained non-deterministic discrete-time state-space model, the mud layer pressure change distribution of the thickener system is obtained.

[0010] S4. Based on the results obtained from sampling the mud layer pressure change 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.

[0011] Optionally, the trained non-deterministic discrete-time state-space model in S2 includes:

[0012] S21. Obtain historical operating parameters of the thickener system; wherein the historical operating parameters include sample inlet and outlet flow rates, sample inlet and outlet concentrations, and sample mud layer pressure.

[0013] S22. Based on a deep neural network containing latent variables, a non-deterministic discrete-time state-space model of the thickener system is constructed. The sample inlet and outlet flow rates, sample inlet and outlet concentrations, and sample mud layer pressure are input into the non-deterministic discrete-time state-space model to obtain the reconstructed predicted mud layer pressure. The non-deterministic discrete-time state-space model is trained based on the reconstruction error of the predicted mud layer pressure and the sample mud layer pressure.

[0014] Optionally, the step of obtaining historical operating parameters of the thickener system in S21 includes:

[0015] Collect the original operating parameters monitored by each sensor of the thickener system.

[0016] The mean and variance of each parameter in the original operating parameters are calculated, and based on the statistical mean and variance of each parameter, the original operating parameters are normalized and scaled to obtain the historical operating parameters.

[0017] Optionally, the training of the non-deterministic discrete-time state-space model in S22 includes: estimating the gradient of the loss function with respect to the parameters of the non-deterministic discrete-time state-space model, and after obtaining each gradient, optimizing the training of the non-deterministic discrete-time state-space model using the stochastic gradient descent method.

[0018] Optionally, the non-deterministic discrete-time state-space model includes a posterior encoding module and a priori prediction module.

[0019] Among them, the posterior encoding module is used for latent variable reasoning to realize the encoding of historical operating data of the thickener system.

[0020] The prior prediction module is used to represent the prior distribution of latent variables and realize the prediction of the mud layer pressure of the thickener system.

[0021] Optionally, the non-deterministic discrete-time state-space model includes a posterior encoding module and a priori prediction module including:

[0022] Based on the variational autoencoder method, an approximate a posteriori inference model is constructed from the observed quantities of the thickener system to the latent variables of the thickener system. The variational evidence lower bound is used as the optimization target of the approximate a posteriori inference model, and the approximate a posteriori inference model is trained for the observation posterior encoding module and the prior prediction module.

[0023] Optionally, in S4, based on the results obtained by sampling the mud layer pressure change distribution of the thickener system, the input control sequence of the thickener system is optimized according to a cross entropy optimization algorithm to obtain an optimal input control sequence of the thickener system, and the thickener system is controlled based on the optimal input control sequence, including:

[0024] S41. Construct an optimal input control sequence distribution of the initial state, the optimal input control sequence distribution of the initial state obeys a Gaussian distribution, and sample to obtain the optimal input control sequence.

[0025] S42. Construct an evaluation function. Based on the optimal input control sequence and the trained non-deterministic discrete-time state-space model, obtain the error between the mud layer pressure of the thickener system and the set value, as well as the instability of the optimal input control sequence of the thickener. Re-estimate the distribution of the optimal input control sequence based on the evaluation function and the optimal input control sequence obtained by sampling.

[0026] S43. Repeat step S42. After a preset number of iterations, the mean value of the optimal input control sequence finally obtained is used as the system action of the thickener system at the next moment.

[0027] In another aspect, a thickener control device based on a non-deterministic latent space model is provided, the device being applied to an electronic device, the device comprising:

[0028] The data acquisition module is used to obtain the current operating parameters of the thickener system, including the inlet and outlet flow rates and the inlet and outlet concentrations.

[0029] The non-deterministic discrete-time state-space model prediction module is used to input current operating parameters into the trained non-deterministic discrete-time state-space model.

[0030] The non-deterministic discrete-time state-space model output module is used to obtain the mud layer pressure change distribution of the thickener system based on the current operating parameters and the trained non-deterministic discrete-time state-space model.

[0031] The optimal input control module is used to optimize the input control sequence of the thickener system based on the results obtained by sampling the mud layer pressure change distribution of the thickener system to obtain the optimal input control sequence of the thickener system.

[0032] Optionally, the non-deterministic discrete-time state-space model prediction module is further configured to:

[0033] Trained non-deterministic discrete-time state-space models, including:

[0034] S21. Obtain historical operating parameters of the thickener system; wherein the historical operating parameters include sample inlet and outlet flow rates, sample inlet and outlet concentrations, and sample mud layer pressure.

[0035] S22. Based on a deep neural network containing latent variables, a non-deterministic discrete-time state-space model of the thickener system is constructed. The sample inlet and outlet flow rates, sample inlet and outlet concentrations, and sample mud layer pressure are input into the non-deterministic discrete-time state-space model to obtain the reconstructed predicted mud layer pressure. The non-deterministic discrete-time state-space model is trained based on the reconstruction error of the predicted mud layer pressure and the sample mud layer pressure.

[0036] Optionally, the non-deterministic discrete-time state-space model prediction module is further configured to:

[0037] The historical operating parameters of the thickener system are obtained in S21, including:

[0038] Collect the original operating parameters monitored by each sensor of the thickener system.

[0039] The mean and variance of each parameter in the original operating parameters are calculated, and based on the statistical mean and variance of each parameter, the original operating parameters are normalized and scaled to obtain the historical operating parameters.

[0040] Optionally, the non-deterministic discrete-time state-space model prediction module is further configured to:

[0041] The training of the non-deterministic discrete-time state-space model in S22 includes: estimating the gradient of the loss function with respect to the parameters of the non-deterministic discrete-time state-space model, and after obtaining each gradient, optimizing the training of the non-deterministic discrete-time state-space model using the stochastic gradient descent method.

[0042] Optionally, the non-deterministic discrete-time state-space model prediction module is further configured to:

[0043] The non-deterministic discrete-time state-space model includes a posterior encoding module and a priori prediction module.

[0044] Among them, the posterior encoding module is used for latent variable reasoning to realize the encoding of historical operating data of the thickener system.

[0045] The prior prediction module is used to represent the prior distribution of latent variables and realize the prediction of the mud layer pressure of the thickener system.

[0046] Optionally, the non-deterministic discrete-time state-space model prediction module is further configured to:

[0047] The non-deterministic discrete-time state space model includes a posterior encoding module and a priori prediction module including:

[0048] Based on the variational autoencoder method, an approximate a posteriori inference model is constructed from the observed quantities of the thickener system to the latent variables of the thickener system. The variational evidence lower bound is used as the optimization target of the approximate a posteriori inference model, and the approximate a posteriori inference model is trained for the observation posterior encoding module and the prior prediction module.

[0049] Optionally, the optimal input control module is further configured to:

[0050] Based on the results obtained from sampling the distribution of mud layer pressure changes in the thickener system, the input control sequence of the thickener system is optimized using the cross entropy optimization algorithm to obtain the optimal input control sequence of the thickener system. The thickener system is controlled based on the optimal input control sequence, including:

[0051] S41. Construct an optimal input control sequence distribution of the initial state, the optimal input control sequence distribution of the initial state obeys a Gaussian distribution, and sample to obtain the optimal input control sequence.

[0052] S42. Construct an evaluation function. Based on the optimal input control sequence and the trained non-deterministic discrete-time state-space model, obtain the error between the mud layer pressure of the thickener system and the set value, as well as the instability of the optimal input control sequence of the thickener. Re-estimate the distribution of the optimal input control sequence based on the evaluation function and the optimal input control sequence obtained by sampling.

[0053] S43. Repeat step S42. After a preset number of iterations, the mean value of the optimal input control sequence finally obtained is used as the system action of the thickener system at the next moment.

[0054] On the one hand, an electronic device is provided, comprising a processor and a memory, wherein 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 thickener control method based on the non-deterministic latent space model.

[0055] In one aspect, a computer-readable storage medium is provided, wherein the storage medium stores at least one instruction, and the at least one instruction is loaded and executed by a processor to implement the above-mentioned thickener control method based on the non-deterministic latent space model.

[0056] The above technical solutions of the embodiments of the present invention have at least the following beneficial effects:

[0057] The present invention innovatively proposes to use a non-deterministic latent space model with randomness as a prediction model for the thickener system. This method constructs and learns a non-deterministic latent variable dynamic model of the dynamic change process of the mud layer pressure of the thickener system based on the thickener system operating data. In the control stage, the trained non-deterministic latent variable dynamic model is used to predict the mud layer pressure change of the system in a certain period of time in the future under a given control input, and the cross-entropy algorithm is used to calculate the optimal control input sequence of the future system. Compared with the traditional predictive control method of the thickener based on a deterministic model, the present invention uses a non-deterministic latent variable dynamic model with randomness as the prediction model of the thickener system. This method can better represent the complex noise disturbance and uncertainty of the thickener system. Therefore, the entire prediction and control method has better prediction accuracy and control accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0059] Figure 1 1 is a flow chart of a thickener control method based on a non-deterministic latent space model provided by an embodiment of the present invention;

[0060] Figure 2 is an overview diagram of a thickener device provided by an embodiment of the present invention;

[0061] Figure 3 This is a flowchart of a construction based on a non-deterministic latent space model provided by an embodiment of the present invention;

[0062] Figure 4 is a structural diagram of a non-deterministic latent space model provided by an embodiment of the present invention;

[0063] Figure 5 This is a model predictive control framework diagram provided by an embodiment of the present invention;

[0064] Figure 6 This is a flowchart of an optimal input control sequence optimization for a thickener system provided by an embodiment of the present invention;

[0065] Figure 7 This is a thickener control service diagram based on the Python Flask service framework HTTP protocol provided by an embodiment of the present invention;

[0066] Figure 8 1 is a flow chart of a thickener control device based on a non-deterministic latent space model provided by an embodiment of the present invention;

[0067] Figure 9 It is a structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0068] In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0069] like Figure 1 As shown, an embodiment of the present invention provides a thickener control method based on a non-deterministic latent space model, which is applied to an electronic device and includes:

[0070] S1. Obtain the current operating parameters of the thickener system, including the inlet and outlet flow rates and the inlet and outlet concentrations.

[0071] S2. Input the inlet and outlet flow rates and inlet and outlet concentrations into the trained non-deterministic discrete-time state-space model.

[0072] Optionally, the trained non-deterministic discrete-time state-space model in S2 includes:

[0073] S21. Obtain historical operating parameters of the thickener system; wherein the historical operating parameters include sample inlet and outlet flow rates, sample inlet and outlet concentrations, and sample mud layer pressure.

[0074] S22. Based on a deep neural network containing latent variables, a non-deterministic discrete-time state-space model of the thickener system is constructed. The sample inlet and outlet flow rates, sample inlet and outlet concentrations, and sample mud layer pressure are input into the non-deterministic discrete-time state-space model to obtain the reconstructed predicted mud layer pressure. The non-deterministic discrete-time state-space model is trained based on the reconstruction error of the predicted mud layer pressure and the sample mud layer pressure.

[0075] Optionally, obtaining historical operating parameters of the thickener system in S21 includes:

[0076] Collect the original operating parameters monitored by each sensor of the thickener system.

[0077] The mean and variance of each parameter in the original operating parameters are calculated, and based on the statistical mean and variance of each parameter, the original operating parameters are normalized and scaled to obtain the historical operating parameters.

[0078] Optionally, the training of the non-deterministic discrete-time state-space model in S22 includes: estimating the gradient of the loss function with respect to the parameters of the non-deterministic discrete-time state-space model, and after obtaining each gradient, optimizing the training of the non-deterministic discrete-time state-space model using a stochastic gradient descent method.

[0079] Optionally, the non-deterministic discrete-time state-space model includes a posterior encoding module and a priori prediction module.

[0080] Among them, the posterior encoding module is used for latent variable reasoning to realize the encoding of historical operating data of the thickener system.

[0081] The prior prediction module is used to represent the prior distribution of latent variables and realize the prediction of the mud layer pressure of the thickener system.

[0082] Optionally, the non-deterministic discrete-time state-space model includes a posterior encoding module and a priori prediction module including:

[0083] Based on the variational autoencoder method, an approximate a posteriori inference model is constructed from the observed quantities of the thickener system to the latent variables of the thickener system. The variational evidence lower bound is used as the optimization target of the approximate a posteriori inference model, and the approximate a posteriori inference model is trained for the observation posterior encoding module and the prior prediction module.

[0084] S3. Based on the inlet and outlet flow rates, inlet and outlet concentrations, and the trained non-deterministic discrete-time state-space model, the mud layer pressure change distribution of the thickener system is obtained.

[0085] S4. Based on the results obtained from sampling the mud layer pressure change 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.

[0086] Optionally, in S4, based on the results obtained by sampling the mud layer pressure change distribution of the thickener system, the input control sequence of the thickener system is optimized according to a cross entropy optimization algorithm to obtain an optimal input control sequence of the thickener system, and the thickener system is controlled based on the optimal input control sequence, including:

[0087] S41. Construct an optimal input control sequence distribution of the initial state, the optimal input control sequence distribution of the initial state obeys a Gaussian distribution, and sample to obtain the optimal input control sequence.

[0088] S42. Construct an evaluation function. Based on the optimal input control sequence and the trained non-deterministic discrete-time state-space model, obtain the error between the mud layer pressure of the thickener system and the set value, as well as the instability of the optimal input control sequence of the thickener. Re-estimate the distribution of the optimal input control sequence based on the evaluation function and the optimal input control sequence obtained by sampling.

[0089] S43. Repeat step S42. After a preset number of iterations, the mean value of the optimal input control sequence finally obtained is used as the system action of the thickener system at the next moment.

[0090] In a feasible implementation, an embodiment of the present invention proposes a thickener mud layer pressure modeling and predictive control method based on a non-deterministic latent space model in response to the non-deterministic characteristics of the thickener system. The method first establishes a deep time series network composed of a priori modules and a posteriori modules, and uses the offline operation data of the thickener system to obtain a non-deterministic latent space state space model for identifying the thickener system; the network is used to identify the state space model of the thickener system. Based on the state space model, an online control algorithm for the thickener bottom flow concentration based on model predictive control is proposed. By initializing the action distribution, establishing an evaluation function, and using an optimization algorithm for feedback correction and rolling optimization, the optimal control sequence of the thickener system is calculated. The thickener mud layer pressure modeling and control method based on non-deterministic latent space of the embodiment of the present invention have higher prediction and control accuracy than the traditional model predictive control method using a deterministic model, and are more suitable for the modeling and control of the thickener system in an actual production environment.

[0091] like Figure 2 The thickener equipment overview shown in the figure is a typical slurry sedimentation and separation tool, widely used in process industries such as metallurgy, mining, and petrochemicals. The upstream section produces low-concentration slurry with fluctuating concentration and flow rate. Due to the greater density of sediment particles than water and the flocculation effect of the flocculant, the sand particles continuously settle and form a high-concentration underflow at the bottom of the thickener. Under the pressure of the underflow pump, the sand particles are drawn into the conveying pipeline.

[0092] When controlling a thickener, underflow concentration is the core control parameter. This parameter has a complex coupling relationship with other process monitoring variables, such as feed flow rate, feed concentration, discharge flow rate, and sludge layer pressure. Relevant research has shown that among the multiple parameters affecting underflow concentration, changes in the thickener's sludge layer pressure can largely reflect changes in underflow concentration. In other words, stabilizing the thickener's sludge layer pressure can indirectly stabilize the thickener's underflow concentration. Therefore, for the system, sludge layer pressure is the core control target of the thickener's settling control process.

[0093] The performance of the thickening and settling process is evaluated, and its core control indicator is the mud layer pressure y1, which is affected by the control input, system state parameters, and other external noise. The control input is the underflow flow rate u(k), and the system state parameters are the mud layer height h(k), the rake speed c1(k), and the flocculant pump speed c2(k) as external noise inputs. Since the thickener feed flow rate and the slurry concentration at the inlet and outlet also have a corresponding relationship with the mud layer pressure, and in the actual thickener production process, the above variables cannot be manually controlled, so they are also used as prediction variables y2, y3, y4∈R 3 .

[0094] According to the above definition, y=[y1(k),y2(k),y3(k),y4(k)]∈R 4 is the system control quantity, u(k)∈R is the controllable input quantity, which is an important parameter to characterize the current thickener state, c(k)=[c1(k),c2(k)]∈R 2 is the controllable system noise, and h(k) is the system state variable. This parameter is an important parameter that characterizes the current thickener state. It can be indirectly controlled but is not used as a control target. In the industrial field, the particle size and composition of the thickener feed will affect the thickener underflow concentration and thus the mud layer pressure. However, since these variables cannot be directly observed and have small fluctuations, in order to simplify the problem, the thickener is modeled as the following system:

[0095]

[0096] in, is the sequence of external input variables of the system, including the underflow flow u(k). x is the sequence length; are the system state variables, including mud layer pressure, feed flow rate, feed concentration, and underflow concentration. Output y(t+1) is the system state value at the next moment.

[0097] In an embodiment of the present invention, it is innovatively proposed to use a non-deterministic latent space model with randomness as a prediction model for the thickener system. This method constructs and learns a non-deterministic latent variable dynamic model of the dynamic change process of the mud layer pressure of the thickener system based on the thickener system operating data. In the control stage, the trained non-deterministic latent variable dynamic model is used to predict the mud layer pressure change of the system in a certain period of time in the future under a given control input, and the cross-entropy algorithm is used to calculate the optimal control input sequence of the future system. Compared with the traditional predictive control method of the thickener based on a deterministic model, the present invention uses a non-deterministic latent variable dynamic model with randomness as the prediction model of the thickener system. This method can better represent the complex noise disturbance and uncertainty of the thickener system, so the whole set of prediction and control methods has better prediction accuracy and control accuracy.

[0098] like Figure 3 As shown, the embodiment of the present invention provides a construction flow chart based on a non-deterministic latent space model. The processing flow of the method may include:

[0099] S310, obtaining historical operating parameters of the thickener system; wherein the historical operating parameters include sample inlet and outlet flow rates, sample inlet and outlet concentrations, and sample mud layer pressure.

[0100] Optionally, obtaining historical operating parameters of the thickener system in S310 includes steps 3101-3102:

[0101] S3101. Collect original operating parameters monitored by various sensors of the thickener system.

[0102] In one feasible implementation, historical operating parameters of the thickener system, as monitored by various sensors in the system, are obtained. This embodiment utilizes OPC (Object Linking and Embedding for Process Control) technology to read data from an industrial distributed control system (DCS). A data reading service is developed using the OpenOPC toolkit and deployed on a computing terminal. This service is then connected to an OPC server in the DCS control room at the industrial site via a network cable. This enables real-time reading of DCS sensor data, which is then stored in a local MySQL database and subsequently transferred to a MongoDB database. The sensor data is archived every minute.

[0103] S3102. Calculate the mean and variance of each parameter in the original operating parameters, and based on the calculated mean and variance of each parameter, normalize and scale the original operating parameters to obtain historical operating parameters.

[0104] In one feasible implementation, the present embodiment exports thickener operation data from MongoDB and enters data from five monitoring points, including inlet and outlet flow rates, inlet and outlet concentrations, and mud layer pressure, into a CSV (Comma-Separated Values) file. Because the values of different physical quantities vary significantly, this can prevent the network from learning effectively and make hyperparameter setting difficult. Therefore, the present embodiment preprocesses the raw operating parameters using the numpy toolkit. Specifically, the mean and variance of the raw operating parameters are normalized. The specific formula is as follows:

[0105]

[0106] Among them, x is the original operating parameter, x scale is the historical operating parameter, x mean is the original operating parameter mean, x std is the original running parameter variance.

[0107] The historical operating parameter set obtained after preprocessing is divided into three parts: training set (60%), test set (20%) and validation set (20%).

[0108] In addition, it is necessary to count the extreme values of each parameter in the historical operating parameters of the thickener system. This extreme value will be used to constrain the optimal control input sequence obtained by subsequent calculations so that it meets the fault tolerance range of the thickener system.

[0109] S320. Based on a deep neural network containing latent variables, a non-deterministic discrete-time state space model of the thickener system is constructed.

[0110] Alternatively, as Figure 4 The non-deterministic discrete-time state-space model structure diagram shown in Figure 5 As shown in the model predictive control framework diagram, the non-deterministic discrete-time state space model includes a posterior encoding module and a priori prediction module, specifically including S3201-S3202-S3203.

[0111] Based on the variational autoencoder method, an approximate a posteriori inference model is constructed from the observed quantities of the thickener system to the latent variables of the thickener system. The variational evidence lower bound is used as the optimization target of the approximate a posteriori inference model, and the approximate a posteriori inference model is trained for the observation posterior encoding module and the prior prediction module.

[0112] S3201, the posterior encoding module is used for latent variable reasoning to realize the encoding of the historical operating data of the thickener system.

[0113] In one feasible implementation, a nondeterministic discrete-time state-space model is used to model the thickener system. This involves describing the thickener system using a state-space model containing latent variables. The transition process of these latent variables represents a discrete, nondeterministic model of the thickener system's operating parameters. Both the initial probability distribution and the transition condition probability distribution of the latent variables follow a Gaussian distribution, and a deep neural network is used to model the parameters of the probability transition process.

[0114] Specifically, RSSM (Recurrent State Space Model) is used for forward prediction in the latent space. This model can be regarded as a nonlinear Kalman filter or a sequence VAE (Variational Autoencoder). The model contains deterministic branches and stochastic branches. The state of the RSSM model is divided into the random part z t and the deterministic part h t , which depends on the stochastic and deterministic parts of the previous time step through the RNN (Recurrent Neural Network).

[0115] Furthermore, the thickener system is modeled using a non-deterministic discrete-time state space to construct an approximate posterior model from the system's historical operating parameters x, y to the latent variable z, and the system input sequence x within the thickener system time T is converted to 1:T and the system output sequence y 1:T Encoded as a latent variable sequence z 1:T ,In the historical operating parameter sequence collected from the system, x is the system input, y is the system output, and t is the system operating time.

[0116] Based on the given system input sequence x 1:T and the system output sequence y 1:T , get the given system input x t The system output y t The probability distribution of p(y t |x t ). In the embodiment of the present application, the hidden variable sequence z of the system is introduced. 1：T To represent the randomness of the system and the long-term impact of system input on the system.

[0117] In the embodiment of the present application, the thickener system dynamics is represented as p(y t |x t )=p(z t |z t-1 ,x t-1 )p(y t |z t ), where p(z t |z t-1 ,x t-1 ) describes the input x of a given system. t Next, the hidden variable z t The transfer conditional probability distribution, p(y t |z t ) describes the hidden variable z at a given time in the system t Next, the system outputs y t distribution.

[0118] In order to learn the parameterized model p(z t |z t-1 ,x t-1 ) and p(y t |z t ) to achieve a given system input sequence x 1:T And predict the system output sequence y 1:T , and realize the given system output sequence y 1:T and the input sequence x 1:T In this case, estimate the latent variable sequence z 1:TThe posterior distribution of the present application embodiment adopts the variational Bayesian method to introduce the approximate posterior distribution q(z 1:T |y 1:T ,x 1:T ) to approximate the true posterior distribution p(z 1:T |y 1:T ,x 1:T ), and minimize the KL divergence between the two distributions so that the approximate posterior distribution of the latent variable q(z 1:T |y 1:T ,x 1:T ) continuously approaches the true posterior distribution p(z 1:T |y 1:T ,x 1:T ). Specifically, the latent variable random part z t The approximate posterior distribution of is shown in formula (1):

[0119]

[0120] It means that the posterior distribution of latent variables is approximately inferred from the observation sequence and action sequence of the system at the past T time points; where q(z t |z t-1 ,x t-1 ,y t ) is a parameterized diagonal Gaussian distribution.

[0121] Since the thickener system has a large uncertainty and the prediction model used is nonlinear, the required state posteriors cannot be directly calculated using parameter learning. Therefore, an encoder is used to infer the approximate hidden state posteriors from past observations and actions, and taking into account the deterministic part h of the RSSM model t , latent variable deterministic part h t The approximate posterior distribution of is as shown in the following formula (2):

[0122]

[0123] where q(z t |h t-1 ,y t ) is a parameterized neural network and feedforward neural network with a mean of μ0 and a variance of σ 0,t The diagonal Gaussian distribution z t ~N(μ 0,t ,diag(σ 0,t )).

[0124] S3202. Input the sample inlet and outlet flow rates, sample inlet and outlet concentrations, and sample mud layer pressure into the non-deterministic discrete-time state-space model to obtain a reconstructed predicted mud layer pressure, and train the non-deterministic discrete-time state-space model based on the predicted mud layer pressure and the reconstruction error of the sample mud layer pressure.

[0125] Optionally, the training of the non-deterministic discrete-time state-space model in S3202 includes: estimating the gradient of the loss function with respect to the parameters of the non-deterministic discrete-time state-space model, and after obtaining each gradient, optimizing the training of the non-deterministic discrete-time state-space model using a stochastic gradient descent method.

[0126] In a feasible implementation, the variational Bayesian method is used to solve the ELBO (Evidence Lower Bound) of the likelihood of the observed data, and this is used as the optimization target of the model to train the model posterior encoding module and the system dynamics prior prediction module.

[0127] From this, the ELBO of the log-likelihood of the observed data can be established, as shown in formula (3):

[0128]

[0129] The above standard variational lower bound contains a reconstruction term for the observed variable and a KL divergence regularization term that approximates the posterior distribution of the latent variable and the prior distribution of the latent variable, which is generated by the network p(y t |z t ) computes the conditional generative distribution of x, whose log-likelihood lnp(y t |z t ) is larger, indicating that the reconstruction effect of the model is better. At the same time, in order to enable the model to generate new samples, it is necessary to ensure the approximate posterior distribution q(z t |y ≤t ,x ≤t ) with a known prior distribution p(z t |z t-1 ,x t-1 ) as close as possible, that is, the KL divergence regularization term of the second term of the formula. The two formulas serve as the optimization target of the prediction model. Because pure random transformation makes it difficult for the transformation model to reliably remember information of multiple time steps, the embodiment of the present application introduces a deterministic latent variable sequence in the model. The model transmits both deterministic latent variables and random latent variables at the same time. Equations (4) and (5) represent the deterministic latent variables h in the system. t , random hidden variable z t And the relationship between them:

[0130] h t =f(h t-1,z t-1 ,x t-1 ) (4)

[0131] z t ~p(z t |h t ) (5)

[0132] Where f is a basic recurrent neural network. Since the thickener system has a large uncertainty and the prediction model used is nonlinear, the required state posteriors cannot be directly calculated. Therefore, an encoder parameterization is used to approximate the state posteriors. In order to enable the gradient used to optimize the model parameters to be back-propagated to the multi-part prediction calculation graph during training, the gradient flows through p(z t |z t-1 ,x t-1 ) can be passed through z t-1 Continue to enter p(z t-1 |z t-2 ,x t-2 ) and the multiple z samples sampled before prediction t-i ,

[0133] During the training of the model, the input of this module is the thickener history system input x t , output by the history system y t and history system input x t The encoded latent variable z t As well as the hidden state h(t-1) output by the posterior encoding module at the previous moment, the above formula (4) is further explained by formula (6), where formula (6) is used to update the system hidden state and update the network parameters:

[0134]

[0135] where f θ For a basic RNN network, It can be regarded as the feature extractor of x and z, both of which are fully connected networks with a two-layer structure. The number of input nodes is 1, which is the thickener bottom flow rate, 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; The number of input nodes is 4, including thickener mud layer pressure, underflow concentration, feed concentration and feed flow rate, 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. Specifically, the activation function of the output layer of the network is the tanh function.

[0136] During the experiment of the embodiment of the present application, the training history data length of the posterior prediction module is N=160, and the RNN network f θThe number of input nodes is 96, the number of output nodes is 32, the number of layers is 1, and θ is the network parameter of the network.

[0137] Therefore, the embodiment of the present application generalizes the standard variational lower bound to the evidence lower bound of the multi-part prediction of the D step, as shown in the following formula (7):

[0138]

[0139] in is a weighting factor used to adjust whether to focus more on short-term or long-term predictions. D is the step size of the multi-step prediction, and d is the step size of each iteration during the training process (d gradually increases from 1 to D). During the experiment, the prediction performance of the model with D = 1, 3, and 5 was tested. The above model and the standard variational lower bound used for optimization were determined. Furthermore, multiple deep neural networks were used to parameterize the various probability functions and deterministic functions in the non-deterministic discrete-time state space model. The model was trained using the collected and preprocessed training data; the prediction effect of the model was verified using the validation set data.

[0140] In one feasible implementation, during the model validation phase, after updating the system hidden state, the observed variable state decoder is used to decode the system hidden state to obtain a reconstruction of the thickener observed variables. The error between this and the historical thickener operating parameters is then calculated to observe the model prediction effect. Specifically, the loss used in the model validation process is the root relative squared error (RRSE) between the true value and the predicted value of the thickener control target, which is calculated as shown in the following formula (8):

[0141]

[0142] Among them, P (ij) is the value predicted by model i for variable j (among n variables); T j is the target value of variable j; Then it is given by the following formula (9):

[0143]

[0144] S3203, the prior prediction module is used to represent the prior distribution of latent variables and realize the prediction of the mud layer pressure of the thickener system.

[0145] In a feasible embodiment, the prior prediction module is constructed as follows: the prior encoding module inputs the system input x of the thickener system in the future period according to the model parameters trained by the model posterior encoding module in the above steps. t+L , the hidden variable z obtained by encoding the system inputt+L And the hidden layer h of the thickener system at the current moment t , the output model predicts the future latent space state h of the L-length thickener system t+L , as shown in the following formula (10):

[0146]

[0147] The network part is basically the same as the above-mentioned posterior encoding module, and finally the observed variable decoder is used to convert the above-mentioned dense machine prediction hidden state h t+L Decoded into thickener operating parameters to obtain the model's predicted targets.

[0148] Furthermore, the trained model is used for thickener system prediction, which can be specifically expressed as the following two processes: (11) (12):

[0149]

[0150]

[0151] Among them, τ is the backward prediction time, and (T:T+τ) means predicting τ backward from time T. is the predicted result for the future time τ. Equation (11) indicates that the posterior encoding module is used to encode the system input and output data of the thickener system over the past period of time, and the initial state of the latent variables of the thickener system at time T is predicted and sampled. The transfer process of the latent variables in the latent space is discrete and non-deterministic. The probability distribution of the latent variable z0 at the initial position and the probability distribution of the transfer condition are both Gaussian distributions encoded using the historical operating parameter sequence of the thickener system.

[0152] Formula (12) represents the process of prediction using the system prior prediction module, and the input includes the hidden state of the system obtained by the encoder Thickener system input x T+τ The output is the system output under a given system input. The thickener system's current external input data is encoded to obtain the system's current latent variables. Combined with the decoder in the model, the system output under the current input is obtained. The system output includes mud layer pressure and underflow concentration.

[0153] Specifically, in this embodiment, training was performed with 800 epochs, a batch size of 1024, a learning rate of 0.0005, a decay rate of 0.98, and 10 decay steps. SGD (stochastic gradient descent) backpropagation was used for training. The model was implemented and trained using the PyTorch framework, and the trained model was saved as *.pth and *.pkl files.

[0154] In an embodiment of the present invention, it is innovatively proposed to use a non-deterministic latent space model with randomness as a prediction model for the thickener system. This method constructs and learns a non-deterministic latent variable dynamic model of the dynamic change process of the mud layer pressure of the thickener system based on the thickener system operating data. In the control stage, the trained non-deterministic latent variable dynamic model is used to predict the mud layer pressure change of the system in a certain period of time in the future under a given control input, and the cross-entropy algorithm is used to calculate the optimal control input sequence of the future system. Compared with the traditional predictive control method of the thickener based on a deterministic model, the present invention uses a non-deterministic latent variable dynamic model with randomness as the prediction model of the thickener system. This method can better represent the complex noise disturbance and uncertainty of the thickener system, so the whole set of prediction and control methods has better prediction accuracy and control accuracy.

[0155] like Figure 6 As shown, an embodiment of the present invention provides a flowchart for optimizing the optimal input control sequence of a thickener system. The processing flow of the method may include:

[0156] S610, determine the optimization function: according to the thickener control requirements, the control sequence of the thickener is made as stable as possible, and the underflow concentration is stable at the thickener process setting value. The optimization function designed here is shown in the following formula (13):

[0157]

[0158] in, is the estimated value output by the prediction model, is the artificial set value of the control target, and n represents the dimension of the state value; Indicates that the system output is as close as possible to the artificial setting value of the control target. k To optimize the target, it represents the amount of action change; It means that the change of action is minimized through optimization; It is a penalty term. In order to ensure that the action calculation amount is within the normal range, once it exceeds the range, the value of this term will become larger, thereby achieving the purpose of controlling the range of control system action changes.

[0159] In a feasible implementation, in terms of control, based on the trained non-deterministic latent space state space model, the mud layer pressure change of the system is predicted according to the current operating parameters. Based on the change prediction results, the input control sequence of the thickener system is optimized using CEM (Cross Entropy Method, cross entropy optimization algorithm) to obtain the optimal input control sequence of the thickener system.

[0160] Since it is very costly to conduct thickener control experiments in real industrial scenarios, the experiments conducted in the embodiments of the present application are all simulation experiments, and the effectiveness of the control algorithm is verified through simulation experiments.

[0161] S620, Optimization Algorithm Selection: For the currently used non-deterministic prediction model, gradient-based optimization algorithms and traditional PSO (Particle Swarm Optimization) algorithms will make the optimization time too long. Therefore, the cross-entropy optimization algorithm is selected. This optimization algorithm predicts the distribution of action sequences in the short term in the future. It is more suitable for cooperation with random prediction models and uses it as a transfer model to optimize and control the system. Therefore, it is also more suitable as an optimization algorithm for the thickener system.

[0162] S630, the optimization process includes: optimizing the input control sequence of the thickener system according to the cross entropy optimization algorithm based on the results obtained from the distribution of the mud layer pressure variation of the thickener system, obtaining the optimal input control sequence of the thickener system, and controlling the thickener system based on the optimal input control sequence, specifically including S6301-S6303:

[0163] S6301. Construct an optimal input control sequence distribution of the initial state, where the optimal input control sequence distribution of the initial state obeys a Gaussian distribution, and obtain the optimal input control sequence by sampling.

[0164] In one feasible embodiment, initializing the optimal input control sequence distribution includes constructing an optimal input control sequence distribution for the initial state from the currently solved optimal control sequence distribution (initially, each position in the sequence distribution follows a standard normal distribution), sampling a large number of control sequences from this distribution, and predicting the future distribution of key variables in the thickener system under each sampled control input sequence based on a non-deterministic discrete-time state-space model, thereby obtaining the optimal input control sequence for the system.

[0165] S6302. Construct an evaluation function. Based on the optimal input control sequence and the trained non-deterministic discrete-time state-space model, obtain the error between the thickener system mud layer pressure and the set value and the instability of the thickener optimal input control sequence. Re-estimate the optimal input control sequence distribution based on the evaluation function and the optimal input control sequence obtained by sampling.

[0166] S6303: Repeat step S6302. After a preset number of iterations, the mean of the optimal input control sequence distribution finally obtained is used as the optimal input control sequence of the thickener system.

[0167] In one feasible implementation, the input control sequence includes underflow flow rate, underflow concentration, feed flow rate, and feed concentration; the optimization algorithm is a cross entropy optimization algorithm. The algorithm execution process is as follows:

[0168] The optimal input control sequence distribution A(a) of the current thickener system that obeys Gaussian distribution is constructed using the mean and variance of the original thickener operating parameters. t:t+H ), initialize the action confidence matrix A(a t:t+H )←Normal(0,1), generate J action sequences based on the system action confidence distribution sampling Predict the system status in the future through the prediction model Sampling to obtain j control input sequences of length H

[0169] The non-deterministic discrete-time state space model is used to predict the system trajectory in H time corresponding to each control action sequence. As shown in formula (14), where o is the system observation.

[0170]

[0171] Among them, q(s t |o 1:t ,a 1:t-1 ) is the system state s at time t t The posterior represents the system observation from time 1:t and the system action a within 1:t-1 1:t-1 The system state s at time t is obtained t . Represents the system state s at time τ in the future τ The prior means using the system state s at the previous moment τ-1 and the system action in the jth action sequence at the previous moment Prediction s τ . Indicates that by multiplying s τ The a priori prediction of the system state of length H

[0172] The error between the system output predicted by the sampled action sequence and its manually set value is calculated by the evaluation function formula (15), and the evaluation values of multiple input control sequences are given. Then, they are sorted from small to large according to the evaluation values, and the top K control input sequences with the smallest evaluation values are used to update the optimal control input distribution (16) and (17).

[0173]

[0174]

[0175]

[0176] Among them, R (j) is the evaluation value (or reward) of the jth action sequence, r τ represents the evaluation value at time τ, represents the system state at the jth action sequence τ.

[0177] μ t:t+H is the mean of the first K control input sequences, σ t:t+H is its standard deviation. Indicates that the mean is μ t:t+H , with variance σ t:t+H Gaussian distribution, from which the control input sequence A(a t:t+H )

[0178] After I iterations, the mean μ of the optimal control input sequence obtained by sampling the action sequence distribution obtained by optimization is used as the system input to act on the thickener system.

[0179] Specifically, after I iterations, the mean μ of the first action sequence in the first K action sequences is output as the system action. In the simulation experiment of the thickener, the number of iterations is I=50 and K=10. Specifically, for the thickener system modeled using the non-deterministic discrete time state space model, the system's observed variable is s in the algorithm. t is the mud layer pressure, feed flow, feed concentration and discharge concentration, the system action variable a t is the discharge flow rate.

[0180] In an embodiment of the present invention, it is innovatively proposed to use a non-deterministic latent space model with randomness as a prediction model for the thickener system. This method constructs and learns a non-deterministic latent variable dynamic model of the dynamic change process of the mud layer pressure of the thickener system based on the thickener system operating data. In the control stage, the trained non-deterministic latent variable dynamic model is used to predict the mud layer pressure change of the system in a certain period of time in the future under a given control input, and the cross-entropy algorithm is used to calculate the optimal control input sequence of the future system. Compared with the traditional predictive control method of the thickener based on a deterministic model, the present invention uses a non-deterministic latent variable dynamic model with randomness as the prediction model of the thickener system. This method can better represent the complex noise disturbance and uncertainty of the thickener system, so the whole set of prediction and control methods has better prediction accuracy and control accuracy.

[0181] like Figure 7As shown, an embodiment of the present invention provides a thickener control service based on the Python Flask service framework HTTP protocol. The framework includes a thickener mud layer pressure controller and a Python-based flask_service server.

[0182] Specifically, the thickener mud layer pressure controller consists of a data transmitter, a data receiver, and a state controller. The data transmitter sends the current thickener system status to the Flask server via HTTP and sends a POST request to invoke the various functional modules of the flask_service service segment. The data receiver receives the updated thickener system data from the Flask server and the optimal control input for the thickener, derived from a thickener mud layer pressure modeling and control method based on a non-deterministic latent space model. The state controller controls the thickener mud layer pressure. The state controller acts as the thickener underflow flow controller.

[0183] Specifically, the Python-based flask_service server includes two modules: a state update module and an optimization algorithm module. It receives the module call instruction sent by the thickener mud layer pressure controller through the receiver and sends the return output of each module to the thickener mud layer pressure controller.

[0184] Specifically, the state update module updates the thickener system hidden state through the RSSM model posterior encoding module trained using the thickener historical operating parameters, and returns the updated thickener system hidden state and the system current data.

[0185] Specifically, the optimization algorithm module performs control optimization according to the thickener system hidden state and the current state of the system by using the optimization algorithm and returns the optimal control input for the thickener system at the next moment.

[0186] In an embodiment of the present invention, it is innovatively proposed to use a non-deterministic latent space model with randomness as a prediction model for the thickener system. This method constructs and learns a non-deterministic latent variable dynamic model of the dynamic change process of the mud layer pressure of the thickener system based on the thickener system operating data. In the control stage, the trained non-deterministic latent variable dynamic model is used to predict the mud layer pressure change of the system in a certain period of time in the future under a given control input, and the cross-entropy algorithm is used to calculate the optimal control input sequence of the future system. Compared with the traditional predictive control method of the thickener based on a deterministic model, the present invention uses a non-deterministic latent variable dynamic model with randomness as the prediction model of the thickener system. This method can better represent the complex noise disturbance and uncertainty of the thickener system, so the whole set of prediction and control methods has better prediction accuracy and control accuracy.

[0187] like Figure 8 As shown, an embodiment of the present invention provides a thickener control device 800 based on a non-deterministic latent space model. The device 800 is used to implement the above-mentioned thickener control method based on a non-deterministic latent space model. The device 800 includes:

[0188] The data acquisition module 810 is used to obtain the current operating parameters of the thickener system, including the inlet and outlet flow rates and the inlet and outlet concentrations;

[0189] Prediction module 820, for inputting current operating parameters into the trained non-deterministic discrete-time state-space model;

[0190] Output module 830 is used to obtain the distribution of mud layer pressure changes of the thickener system based on the current operating parameters and the trained non-deterministic discrete time state space model;

[0191] The optimal input control module 840 is configured to optimize the input control sequence of the thickener system based on the results obtained by sampling the distribution of the mud layer pressure variation of the thickener system to obtain the optimal input control sequence of the thickener system.

[0192] Optionally, the apparatus 800 further includes a training module 850;

[0193] The training module 850 is used to:

[0194] Obtain historical operating parameters of the thickener system; wherein the historical operating parameters include sample inlet and outlet flow rates, sample inlet and outlet concentrations, and sample mud layer pressure;

[0195] Based on a deep neural network containing latent variables, a non-deterministic discrete-time state-space model of the thickener system is constructed. The sample inlet and outlet flow rates, sample inlet and outlet concentrations, and sample mud layer pressure are input into the non-deterministic discrete-time state-space model to obtain the reconstructed predicted mud layer pressure. The non-deterministic discrete-time state-space model is trained based on the reconstruction error between the predicted mud layer pressure and the sample mud layer pressure.

[0196] Optionally, the prediction module 820 is further configured to:

[0197] Collect the original operating parameters monitored by various sensors of the thickener system;

[0198] The mean and variance of each parameter in the original operating parameters are calculated, and based on the statistical mean and variance of each parameter, the original operating parameters are normalized and scaled to obtain the historical operating parameters.

[0199] Optionally, the training module 850 is further configured to:

[0200] When training a non-deterministic discrete-time state-space model, the gradient of the loss function with respect to the parameters of the non-deterministic discrete-time state-space model is estimated. After obtaining each gradient, the stochastic gradient descent method is used to optimize the training of the non-deterministic discrete-time state-space model.

[0201] Optionally, the prediction module 820 is further configured to:

[0202] Based on the variational autoencoder method, an approximate a posteriori inference model is constructed from the observed quantities of the thickener system to the latent variables of the thickener system. The variational evidence lower bound is used as the optimization target of the approximate a posteriori inference model, and the approximate a posteriori inference model is trained for the observation posterior encoding module and the prior prediction module.

[0203] Optionally, the optimal input control module 840 is further configured to:

[0204] S81. Construct an optimal input control sequence distribution of the initial state, where the optimal input control sequence distribution of the initial state obeys a Gaussian distribution, and obtain the optimal input control sequence by sampling;

[0205] S82. Construct an evaluation function, and based on the optimal input control sequence and the trained non-deterministic discrete-time state-space model, obtain the error between the thickener system mud layer pressure and the set value and the instability of the thickener optimal input control sequence. Re-estimate the optimal input control sequence distribution based on the evaluation function and the optimal input control sequence obtained by sampling.

[0206] S83. Repeat step S82. After a certain number of iterations, the mean value of the optimal input control sequence finally obtained is used as the system action of the thickener system at the next moment.

[0207] In an embodiment of the present invention, it is innovatively proposed to use a non-deterministic latent space model with randomness as a prediction model for the thickener system. This method constructs and learns a non-deterministic latent variable dynamic model of the dynamic change process of the mud layer pressure of the thickener system based on the thickener system operating data. In the control stage, the trained non-deterministic latent variable dynamic model is used to predict the mud layer pressure change of the system in a certain period of time in the future under a given control input, and the cross-entropy algorithm is used to calculate the optimal control input sequence of the future system. Compared with the traditional predictive control method of the thickener based on a deterministic model, the present invention uses a non-deterministic latent variable dynamic model with randomness as the prediction model of the thickener system. This method can better represent the complex noise disturbance and uncertainty of the thickener system, so the whole set of prediction and control methods has better prediction accuracy and control accuracy.

[0208] like Figure 9As shown, an electronic device 900 according to an embodiment of the present invention is provided. The electronic device 900 may have relatively large differences due to different configurations or performances. The electronic device 900 may include one or more processors (central processing units, CPUs) 901 and one or more memories 902. The memories 902 store at least one instruction, which is loaded and executed by the processor 901 to implement the following steps of a thickener control method based on a non-deterministic latent space model:

[0209] Obtain the current operating parameters of the thickener system, including inlet and outlet flow rates and inlet and outlet concentrations;

[0210] Input the inlet and outlet flow rates and inlet and outlet concentrations into the trained non-deterministic discrete-time state space model;

[0211] Based on the inlet and outlet flow rates, inlet and outlet concentrations, and the trained non-deterministic discrete-time state-space model, the distribution of the mud layer pressure variation of the thickener system is obtained.

[0212] Based on the results obtained from sampling the mud layer pressure variation distribution of the thickener system, the input control sequence of the thickener system is optimized using the cross entropy optimization algorithm to obtain the optimal input control sequence of the thickener system. The thickener system is then controlled based on the optimal input control sequence.

[0213] Those skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware, or by a program to instruct the relevant hardware, and the program may be stored in a computer-readable storage medium, which may be a read-only memory, a disk, or an optical disk, etc.

[0214] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A thickener control method based on a non-deterministic latent space model, characterized in that: The method comprises: S1. Obtaining current operating parameters of the thickener system, wherein the current operating parameters include inlet and outlet flow rates and inlet and outlet concentrations; S2. Inputting the inlet and outlet flow rates and inlet and outlet concentrations into a trained non-deterministic discrete-time state-space model; S3. Based on the inlet and outlet flow rates, inlet and outlet concentrations, and the trained non-deterministic discrete-time state-space model, obtain a distribution of changes in the mud layer pressure of the thickener system; S4. Based on the results obtained by sampling the distribution of the mud layer pressure change of the thickener system, optimizing the input control sequence of the thickener system according to a cross entropy optimization algorithm to obtain an optimal input control sequence of the thickener system, and controlling the thickener system based on the optimal input control sequence; The trained non-deterministic discrete-time state-space model in S2 includes: S21. Construct a thickener system using the mud layer pressure, feed flow rate, feed concentration, and underflow concentration as system control variables, the underflow flow rate as a controllable input variable, the rake speed and flocculant pump speed as controllable system noise variables, and the mud layer height as a system state variable. Obtaining historical operating parameters of the thickener system; wherein the historical operating parameters include sample inlet and outlet flow rates, sample inlet and outlet concentrations, and sample mud layer pressure; S22. Construct a non-deterministic discrete-time state-space model of the thickener system based on a deep neural network including latent variables, input the sample inlet and outlet flow rates, sample inlet and outlet concentrations, and sample mud layer pressure into the non-deterministic discrete-time state-space model to obtain a reconstructed predicted mud layer pressure, and train the non-deterministic discrete-time state-space model based on the predicted mud layer pressure and the reconstruction error of the sample mud layer pressure; The non-deterministic discrete-time state space model includes a posterior encoding module and a priori prediction module; The posterior encoding module is used for latent variable reasoning to realize the encoding of the historical operation data of the thickener system; The posterior encoding module introduces a deterministic latent variable sequence, so that the deterministic latent variables and random latent variables are transmitted simultaneously in the model. Formulas (1) and (2) represent the deterministic latent variables h in the system. t , random hidden variable z t And the relationship between them: h t =f(h t-1 ,z t-1 ,x t-1 ) (1) z t ~p(z t |h t )(2) Where f is a basic recurrent neural network, x t is the system input, t is the system running time; The posterior encoding module generalizes the standard variational lower bound to the evidence lower bound of D-step multi-part prediction; The prior prediction module is used to represent the prior distribution of the latent variable and realize the prediction of the mud layer pressure of the thickener system; The step S4 includes optimizing the input control sequence of the thickener system according to a cross entropy optimization algorithm based on the results obtained by sampling the mud layer pressure change 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, including: S41, constructing an optimal input control sequence distribution of an initial state, wherein the optimal input control sequence distribution of the initial state obeys a Gaussian distribution, and sampling to obtain the optimal input control sequence; S42: constructing an evaluation function, and obtaining an error between the thickener system mud layer pressure and a set value and an 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 based on the evaluation function and the optimal input control sequence obtained by sampling; The evaluation function is shown in the following formula (3): Among them, R (j) is the evaluation value of the jth action sequence, H is the length of the control input sequence, r τ represents the evaluation value at time τ, represents the system state at the jth action sequence τ; S43 , repeating step S42 , and after a preset number of iterations, taking the mean value of the optimal input control sequence finally solved as the system action of the thickener system at the next moment.

2. The thickener control method based on the non-deterministic latent space model according to claim 1 is characterized in that: The step S21 of obtaining historical operating parameters of the thickener system includes: collecting original operating parameters monitored by various sensors of the thickener system; The mean and variance of each parameter in the original operating parameters are counted, and based on the counted mean and variance of each parameter, the original operating parameters are normalized and scaled to obtain the historical operating parameters.

3. The thickener control method based on the non-deterministic latent space model according to claim 1, characterized in that: The training of the non-deterministic discrete-time state-space model in S22 includes: estimating the gradient of the loss function with respect to the parameters of the non-deterministic discrete-time state-space model, and after obtaining each gradient, optimizing the training of the non-deterministic discrete-time state-space model using the stochastic gradient descent method.

4. The thickener control method based on the non-deterministic latent space model according to claim 1, characterized in that: The non-deterministic discrete-time state space model includes a posterior encoding module and a priori prediction module including: Based on the variational autoencoder method, an approximate a posteriori inference model is constructed from the observed quantities of the thickener system to the latent variables of the thickener system, and the variational evidence lower bound is used as the optimization target of the approximate a posteriori inference model. The approximate a posteriori inference model is trained to observe the posterior encoding module and the prior prediction module.

5. A thickener control device based on a non-deterministic latent space model, characterized in that: The device comprises: A data acquisition module, configured to acquire current operating parameters of the thickener system, including inlet and outlet flow rates and inlet and outlet concentrations; a non-deterministic discrete-time state-space model prediction module, configured to input the current operating parameters into a trained non-deterministic discrete-time state-space model; a non-deterministic discrete-time state-space model output module, configured to obtain a distribution of mud layer pressure changes of the thickener system based on the current operating parameters and the trained non-deterministic discrete-time state-space model; an optimal input control module for optimizing an input control sequence of the thickener system according to a cross entropy optimization algorithm based on a result sampled from a distribution of changes in mud layer pressure of the thickener system, to obtain an optimal input control sequence of the thickener system; The trained non-deterministic discrete-time state-space model includes: The thickener system is constructed by taking mud layer pressure, feed flow rate, feed concentration and underflow concentration as system control variables, underflow flow rate as controllable input variable, rake frame speed and flocculant pump speed as controllable system noise variables, and mud layer height as system state variable. Obtaining historical operating parameters of the thickener system; wherein the historical operating parameters include sample inlet and outlet flow rates, sample inlet and outlet concentrations, and sample mud layer pressure; Based on a deep neural network containing latent variables, a non-deterministic discrete-time state-space model of the thickener system is constructed, the sample inlet and outlet flow rates and the sample inlet and outlet concentrations are input into the non-deterministic discrete-time state-space model to obtain a predicted mud layer pressure, and the non-deterministic discrete-time state-space model is trained based on the predicted mud layer pressure and the sample mud layer pressure; The non-deterministic discrete-time state space model includes a posterior encoding module and a priori prediction module; The posterior encoding module is used for latent variable reasoning to realize the encoding of the historical operation data of the thickener system; The posterior encoding module introduces a deterministic latent variable sequence, so that the deterministic latent variables and random latent variables are transmitted simultaneously in the model. Formulas (1) and (2) represent the deterministic latent variables h in the system. t , random hidden variable z t And the relationship between them: h t =f(h t-1 ,z t-1 ,x t-1 ) (1) z t ~p(z t |h t )(2) Where f is a basic recurrent neural network, x t is the system input, t is the system running time; The posterior encoding module generalizes the standard variational lower bound to the evidence lower bound of D-step multi-part prediction; The prior prediction module is used to represent the prior distribution of the latent variable and realize the prediction of the mud layer pressure of the thickener system; The method includes optimizing an input control sequence of the thickener system according to a cross entropy optimization algorithm based on a result sampled from a distribution of changes in mud layer pressure 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, including: S41, constructing an optimal input control sequence distribution of an initial state, wherein the optimal input control sequence distribution of the initial state obeys a Gaussian distribution, and sampling to obtain the optimal input control sequence; S42: constructing an evaluation function, and obtaining an error between the thickener system mud layer pressure and a set value and an 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 based on the evaluation function and the optimal input control sequence obtained by sampling; The evaluation function is shown in the following formula (3): Among them, R (j) is the evaluation value of the jth action sequence, H is the length of the control input sequence, r τ represents the evaluation value at time τ, s τ (j) represents the system state at the jth action sequence τ; S43 , repeating step S42 , and after a preset number of iterations, taking the mean value of the optimal input control sequence finally solved as the system action of the thickener system at the next moment.

Images