Plant monitoring and / or control via machine learning regressors
A nested ANN system efficiently simulates and controls batch plants by separating educt and process influences, addressing the limitations of traditional models with sparse datasets and changing conditions, achieving accurate product quality predictions.
Patent Information
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- BASF SE
- Filing Date
- 2022-05-27
- Publication Date
- 2026-06-29
AI Technical Summary
Existing methods for simulating and controlling batch plants face challenges due to the high computational requirements of white-box models and the lack of flexibility and efficiency in traditional black-box models, particularly when dealing with sparse datasets and changes in educt characteristics or process conditions.
A nested artificial neural network (ANN) system is employed, comprising separate ANNs for educt quality parameters and process parameters, allowing for efficient training and prediction of product quality parameters, even with sparse datasets, by separating the influence of educts and process conditions.
The ANN system provides robust and flexible simulation and control of batch plants, reducing the need for extensive training data and enabling accurate prediction of product quality parameters, even with limited data availability.
Smart Images
Figure 0007881621000014 
Figure 0007881621000015 
Figure 0007881621000016
Abstract
Description
[Technical Field]
[0001] The present invention relates in particular to a method including an artificial neural network system, a regressor, and a control device for simulating, monitoring, and / or controlling a plant. Furthermore, a method for training an artificial neural network system is provided. In particular, models based on an artificial neural network (ANN) are suitable for simulating a batch process in which educt is converted into products by a batch processing unit, and are suitable for monitoring and / or controlling a batch plant. [Background technology]
[0002] In batch processing or batch plants, the production of multiple products is carried out using the same set of equipment or processing units, such as chemical or biological reactors. On the one hand, it is desirable to optimize a specific process that includes a particular eduction that is processed into the desired product. On the other hand, scheduling batch operations using a single processing unit can be improved if the process is accurately modeled and simulated.
[0003] Traditional approaches to simulating chemical reactions in batch processes have relied on rigorous models or so-called white-box models based on first principles. White-box models generally require significant computational power and resources to model complex physicochemical systems that involve nonlinear equations.
[0004] Another approach is the so-called black-box model, which relies on machine learning concepts developed in the past. For example, neural networks can be used to predict specific characteristics and quality of a product based on input data, including process and educt characteristics. Such neural networks sometimes need to be trained using multiple datasets that are not readily available. Furthermore, traditional black-box models based on artificial neural networks lack flexibility because they need to be retrained if the conditions in the batch plant being modeled, the desired product quality, and / or educt characteristics change.
[0005] A further approach involves hybrid models that combine white-box and black-box models. Reference WO2020 / 227383A1 discloses a computer-based process modeling and simulation method and system that combines first-principles models and machine learning models, useful when either model is lacking. In one example, measured input values are adjusted using first-principles techniques. A machine learning model of the target chemical process is trained on these adjusted values. In another example, the machine learning model represents the residual between the predictions of the first-principles model and empirical data. The residual machine learning model corrects the predictions of physical phenomena in the first-principles model of the chemical process. In yet another example, a first-principles simulation model uses process input data and the predicted values of the machine learning model to generate simulation results for the chemical process.
[0006] Reference WO2020 / 058237A2 discloses a method and system capable of predicting the values of product quality attributes of a compound or its formulation as a result of a multi-stage manufacturing process. Here, the entire process or process step is characterized by process parameters. This is achieved by performing a multivariate data analysis of the process data in a quality prediction model, which identifies or represents the mathematical relationship between quality attributes and the process parameters of the manufacturing process or its sub-processes. The quality prediction model is obtained by mathematical modeling of historical process data, most preferably using a neural network model combined with empirical process knowledge obtained over time. [Prior art documents] [Patent Documents]
[0007] [Patent Document 1] WO2020 / 227383A1 [Patent Document 2] WO2020 / 058237A2 [Overview of the project] [Problems that the invention aims to solve]
[0008] The purpose of this disclosure is to provide improved methods and systems for monitoring and / or controlling a plant. [Means for solving the problem]
[0009] The aspect of the independent claim solves this problem.
[0010] This disclosure provides a computer-implemented control device for simulating, monitoring, and / or controlling a plant. The plant may be implemented to receive one or more educts having relevant educt quality parameters, process the educts, and output a product having relevant process parameters and relevant product quality parameters. A regressor comprises at least two regressor units, each regressor unit comprising an input unit for receiving input data and an output unit for outputting output data. Within the regressor unit, a first regressor unit is implemented to receive the educt quality parameters and output at least one educt impact parameter, and a second regressor unit is implemented to receive the educt impact parameter and the process parameter and output at least one product quality parameter. The regressor and / or regressor units are based on machine learning principles.
[0011] In one embodiment, an artificial neural network (ANN) system for simulating, monitoring, and / or controlling a batch plant is disclosed. The batch plant is implemented to receive one or more educts having associated educt quality parameters, process the educts having associated process parameters, and output products having associated product quality parameters. The ANN system has at least two ANNs, each ANN having an input section having an input node for receiving input data and an output section having an output node for outputting output data. The first ANN is implemented to receive the educt quality parameters and output at least one educt impact parameter. The second ANN is implemented to receive the educt impact parameter and the process parameter and output at least one product quality parameter.
[0012] The applicant has discovered that nested or concatenated ANNs are suitable regressors for solving separable problems posed by batch plants. Thus, the disclosed regressor can be implemented as an ANN system that includes ANNs.
[0013] In the proposed ANN system, data regarding the educt, i.e., educt quality parameters, and process data regarding the actual process executed in the batch plant, i.e., process parameters, are considered separately. The first ANN models / simulates the influence of educt characteristics from the perspective of educt quality parameters, and the second ANN models / simulates, for example, chemical or biological processes that are at least partially driven by process parameters observable or set during the operation of the batch plant. The second ANN also receives inputs from the first ANN so as to reliably predict the quality parameters of the product.
[0014] It is understood that each ANN includes an input layer having input nodes for receiving input data and an output layer having output nodes for outputting output data. The ANN may also include a hidden layer between the input layer and the output layer. Generally, each ANN is characterized by configuration data that includes at least the bias values and weight values of each node within the ANN. This configuration data is obtained by training each ANN.
[0015] Also, the trained ANN may be referred to as a model of chemical reactions occurring in the process units of the batch plant. The trained ANN is also considered a regressor for basic problems of mapping input data (e.g., characteristics of the educt or process) to output data or target data (e.g., characteristics of the product).
[0016] The proposed approach, which separates the complex problem of mapping various educt quality and process parameters to product quality parameters, allows for the efficient use of sparse datasets when training the ANNs or models, respectively. Embodiments of the ANN system enable the sharing of training datasets used for different products produced in the same or similar batch plants. In particular, a process model or second ANN can benefit from a first ANN or educt model trained for similar or the same educt used in different batch plant processes. For example, the ANN system can modify the educt model by updating the configuration data of the first ANN, even if the educt changes, as long as the overall batch process modeled by the second ANN remains essentially unchanged. Thus, a trained partial ANN for the unchanged component of the batch plant process remains. Doing so reduces the amount of training data required compared to conventional artificial neural networks that need to be fully trained on a set of educt quality and process parameters to target or model product quality parameters.
[0017] According to the applicant's research, the resulting model or neural network system for simulating a batch plant is robust to noise in the training data. In embodiments, a partial black-box model relating to the first ANN or educt model describes the influence of the educt, which is independent of the plant or chemical process setup. Thus, the first ANN can be used in relation to an alternative batch plant or batch plant unit modeled by an alternative second ANN. In particular, the trained first ANN can be reused, and only sparse training data is needed to set up the second ANN or process model for the alternative batch plant.
[0018] In embodiments, output product quality parameters are generated and displayed in a computer-readable format and / or used to control, schedule, or adapt a batch plant, particularly by a control device.
[0019] In embodiments, the second ANN is based on a training dataset containing process parameters and product target variables corresponding to product quality parameters associated with each product. For example, process parameters may include observable measurements during a chemical reaction, temperature values, maximum temperature values, time span, reaction span, catalyst storage time, number of free isocyanate groups, or other characteristics of the time series. Other process parameters affecting the chemical reaction in a batch reactor may also be considered.
[0020] Product quality parameters include the viscosity, hardness, roughness, drug interaction effects, pH, or solubility of the product. During chemical processes or reactions in a batch plant, the resulting product or other product quality parameters that characterize the product may also be considered.
[0021] In embodiments of the ANN system, the first ANN is trained on a training dataset that includes educt quality parameters and residuals of a second ANN trained as target variables for educt influence parameters, where both educt quality parameters and residuals correspond to the respective products. The residuals or prediction errors of the second ANN may be due to the sparse availability of the training dataset and are used to train the first ANN that models educt quality. Educt quality parameters include viscosity values, hydroxyl value, concentration values, color parameters, etc. Other quality parameters are also possible.
[0022] An ANN system can be viewed as a linked system of a first ANN and a second ANN. In an embodiment of the ANN system, the system has the following: Multiple first ANNs, each of which corresponds to an educt and is implemented to receive a corresponding educt quality parameter and to output at least one corresponding educt impact parameter; and Multiple second ANNs, each second ANN corresponding to a process for manufacturing a product, and implemented to receive a combination of corresponding process parameters and eductive influence parameters from a first ANN, and to output corresponding product quality parameters.
[0023] It is also conceivable to associate the first ANN with different educts and the second ANN with a batch process having specific relevant process parameters. The outputs from the first ANN are combined, weighted, and fed to the second ANN. For example, each second ANN may receive a linear combination of educt influence parameters from the first ANN. As a result, the ANN system can predict the product quality parameters of products produced in a batch process by combining one or more educts, each characterized by educt quality parameters.
[0024] "An ANN corresponding to a certain action" is interpreted as an ANN that models that action. Therefore, "a second ANN corresponding to a certain product" is an ANN that has been trained and configured to output approximate product quality parameters in response to process parameters related to the manufacturing process of a given product.
[0025] In embodiments of the ANN system, at least one of the ANNs is a feedforward ANN. In embodiments, a Bayesian neural network can be used as the ANN. At least one ANN can be thought of to further include hidden nodes between the input and output nodes.
[0026] According to one aspect of the present disclosure, a control device for controlling a batch plant is proposed. The batch plant is implemented to receive one or more educts having associated educt quality parameters, process the educts having associated process parameters, and output products having one or more associated product quality parameters. The control device comprises an ANN system as disclosed above or below with respect to embodiments, and the control device is implemented to adapt the process as a function of product quality parameters output from a second ANN in response to adapted process parameters. For example, based on a simulation of the ANN system of the process in the batch plant, the control device modifies the process and thus associated process parameters to obtain a desired product quality.
[0027] In an embodiment of the control device, the control device includes a computer processing unit implemented to perform an operation to implement an ANN system and to execute an optimization algorithm to adapt process parameters so that the product quality parameters output from the second ANN correspond to a predetermined product quality.
[0028] Other aspects of this disclosure provide, for example, methods for training a regressor in the context of an ANN system as disclosed above or below with respect to a particular embodiment. The training method includes at least one of the following steps: A step of providing multiple training datasets for a first product and at least one second product. Each training dataset includes product target variables corresponding to educt quality parameters, process parameters, and product quality parameters related to the product; A step of training a second ANN based on a training data subset containing process parameters and product target variables corresponding to the first product. This yields the first residual for each training data subset; A step of training a second ANN based on a training data subset containing process parameters and product target variables corresponding to at least one further product; thereby obtaining a second residual for each training data subset; and A step of training a first ANN based on a subset of training data including residuals as target variables for educt quality parameters and educt impact parameters corresponding to a first product, and based on a further subset of training data including residuals as target variables for educt quality parameters and educt impact parameters corresponding to at least one further product.
[0029] The training method may include training a first ANN based on a training dataset of all products, including a specific educt, targeting the quality parameters of this educt and the residuals from previous (training) steps. This is preferably done per educt.
[0030] According to the aspects of the training method described above, the first ANN is trained to provide the residuals of the second ANN, thus reducing the need for a large number of training data subsets or sets. In an embodiment of the training method, for each product, a step is performed to train a second ANN based on a subset of training data that includes process parameters, educt influence parameters output from a first ANN trained in response to educt quality parameters associated with educt usage to manufacture each product, and product target variables corresponding to product quality parameters associated with each product. Thus, the second ANN receives additional input from the trained first ANN.
[0031] In an embodiment of the training method, the steps of training a first ANN and training a second ANN are repeatedly performed based on a subset of training data that includes process parameters, educt influence parameters from a first ANN trained for educt quality parameters related to the use of educts to manufacture each product, and product target variables corresponding to product quality parameters related to each product.
[0032] In one embodiment, the training method further includes: A step of generating the training dataset by operating a batch plant and measuring process parameters and product quality parameters; and / or The steps include generating the aforementioned training dataset, simulating a batch plant process based on the aforementioned quality parameters, and deploying a white-box numerical model to generate process parameters and product quality parameters.
[0033] According to one aspect of the present disclosure, a method is provided for simulating, monitoring, and / or controlling a batch plant, wherein the batch plant is configured to receive one or more educts having relevant educt quality parameters, a process to process the educts having relevant process parameters, and output products having relevant product quality parameters, the method comprising the step of using an ANN system as described above or below with respect to a particular embodiment, wherein the ANN system is trained according to a training method disclosed above or below with respect to an embodiment.
[0034] A regressor unit is not necessarily implemented as an ANN, and it should be understood that this disclosure also encompasses other appropriate configurations of regressor units. Therefore, throughout this disclosure, the term "ANN" may be replaced with "regressor unit" in order to fully understand the scope of the invention. A regressor unit can be implemented, for example, as a computer-implemented regression method in terms of software functions or services.
[0035] In a further aspect, the disclosure relates to a computer program product including computer-readable instructions that, in response to the execution of machine-readable instructions, causes a computing system including one or more processing devices to perform the methods and functions described above for simulating, monitoring, and / or controlling a batch plant.
[0036] In an embodiment, the computer program product has program code for a computerized control device to perform the methods and functions described above when executed on at least one control computer. The computer program product, such as the computer program means, can be embodied as a memory card, USB stick, CD-ROM, DVD, or as a file that can be downloaded from a server on a network. For example, such a file may be provided by transferring the files constituting the computer program product over a wireless communication network.
[0037] Further possible embodiments or alternative solutions of the present invention also include combinations of features described above or below with respect to embodiments—not expressly mentioned herein. Furthermore, those skilled in the art may add individual or isolated embodiments and features to the most basic forms of the present invention.
[0038] Further embodiments, features, and advantages of the present invention will become apparent from the following description and dependent claims, which are referenced in conjunction with the accompanying drawings. [Brief explanation of the drawing]
[0039] [Figure 1] This is a schematic diagram showing one embodiment of a batch plant. [Figure 2] This is a schematic diagram of the first embodiment of the ANN system. [Figure 3] This is a schematic diagram of a second embodiment of the ANN system. [Figure 4] This is a flowchart containing method steps for training an embodiment of an ANN system. [Figure 5] This is a schematic diagram of a third embodiment of the ANN system. [Figure 6] This document presents an algorithm for training an ANN system embodiment. [Figure 7] This is a schematic diagram of one embodiment of a batch plant control system. [Modes for carrying out the invention]
[0040] Figure 1 is a schematic diagram of one embodiment of a batch plant. The batch plant 1 has a processing unit 2, which is, for example, a continuous stirring tank reactor. The tank reactor 2 can be used to process various educts under specific process parameters to produce products. Products can be, for example, intermediate products of prepolymer synthesis, polyols, coatings, and other chemical, pharmaceutical, or biological compositions.
[0041] In the example in Figure 1, one input educt 3 is shown that is processed in reactor 2 to become product 4. Specific educt quality parameters can be set for educt 3, indicated by label x1. These educt quality parameters include specific properties of educt 3, such as reactivity, hydroxyl value, and specific isomers. In reactor 2, a chemical reaction process takes place, characterized by process parameter y. After the reaction process in reactor 2 is completed or stopped, product 45 with specific quality parameters, indicated as Q1, can be recovered. For example, quality parameter Q1 may refer to the purity or concentration of the substance in product 4.
[0042] The batch processing unit 2 can also be used for other products. For example, the dashed line shows an alternative educt 5 with educt quality parameter x2. Batch processing yields an alternative product 6 with product quality parameter Q2. It is desirable to predict the effect of educts 3 and 5 and the applied process parameter y on products 4 and 6, and especially on their product quality parameters. This problem can be written as follows:
[0043]
number
[0044] Here, j represents the j-th product of all potential products {1, 2, ..., p} in this disclosure. The batch process model or simulation uses the educt quality parameter x and process parameter y j The combination of each product j is the product quality parameter Q j It needs to be mapped to the input vector x, which describes the quality measurements of the educt, such as substance concentration, viscosity measurement, color parameter, etc. The input vector y j This indicates process parameters that can be measured during product manufacturing in reactor 2. Time-series characteristics such as maximum, minimum, and average values, as well as physical observations such as temperature and pressure values, can be considered process parameters. Output product quality parameter Q jis a scalar and represents the desired quality of product j.
[0045] FIG. 2 is a schematic diagram of a first embodiment of an ANN system implemented to model or simulate a batch process, and is implemented, for example, in reactor 2 of FIG. 1. The ANN system 7 is composed of a first ANN 9 and a second ANN 10. The underlying ANN model or regressor is denoted as ANN1 and ANN21. The ANNs 9, 10 are, for example, shallow ANNs having an input layer with input nodes for receiving input data and an output layer with output nodes for outputting output data. The configurations of the ANNs 9, 10 are defined by configuration data including bias values and weight values of each node of each ANN (not shown).
[0046] The first ANN 9 receives the educt quality parameter x of the educt involved in the production of the desired product having a predetermined product quality parameter Q j and outputs the educt influence parameter R i to the input nodes of the second ANN 10. The second ANN 10 further receives the process parameter y of product j in reactor 2 i and outputs the predicted quality parameter Q j . Thus, the embodiment of the ANN system 7 maps the educt quality parameter x j and the process parameter y i to the product quality parameter Q j . Assume that the problem depicted in Equation 1 can be written as follows. j
[0047]
Equation
[0048] Equation 2 assumes that the mapping or function of the product quality parameter Q j is separable, the function f depends on the pipeline quality parameter x i and the function g depends on the process parameter yj Depends on the Educt quality parameter x. i However, this is advantageous if it similarly affects the reaction kinetics depending on the process within the reactor. The ANN system 7 in Figure 2 solves a regression problem which can be written as follows:
[0049]
number
number
number
[0050] Generally, the ANN system 7 shown in Figure 2 can simulate either a batch process or a reactor 2, respectively. The ANN system 7 can simulate the quality parameter x of the educt. i Modeling the effect of (ANN1), on the other hand, process parameter y j By configuring the first and second ANN9 and 10 specifically to model the effects (ANN21), efficient training and configuration of the configuration data for ANN9 and 10 becomes possible.
[0051] Figure 3 is a schematic diagram of a second embodiment of ANN system 8. ANN system 8 includes a first ANN9 and two second ANNs 10 and 11 labeled ANN21 and ANN22. These labels indicate the model implemented through the ANN configuration data (bias, weights, number of nodes, topology). The first second ANN10 is derived from the first ANN9 by an eductive parameter R iIt then receives a process parameter y1 that describes or characterizes the batch process leading to product j=1 with product quality parameter Q1. Another second ANN11 receives an alternative process parameter y2 that characterizes the process leading to alternative product j=2 with product quality parameter Q2. Models ANN1, ANN21, and ANN22 are implemented as shallow neural networks, such as Basian-type neural networks, eliminating the need for dataset validation during training.
[0052] Figure 4 shows a flowchart including steps for training the ANN systems 7 and 8 shown in Figure 2 or Figure 3. In the first step S1, the ANN21 and ANN22 models are trained separately from each other, for example, using a Basian learning method. Each training dataset for ANN21 contains process parameters y1 and a predetermined product quality parameter Q. i This is included as a target variable. Multiple training datasets may be used for each product. In step S1, the same learning process is performed on model ANN22, and the process parameter y2 is included as the designed target variable, or the measured product quality parameter Q2 is placed therein. For example, because the amount of training data is limited and the eductive influence input is missing, models ANN21 and ANN22 will produce residuals or errors in their quality predictions.
[0053] In the next step, S2, the residuals of the target variable, namely Q1 and Q2, are calculated.
[0054] Next, in step S3, the ANN model ANN1 determines the educt quality parameter x i, and are trained based on a training dataset that includes pipe quality parameters as training target variables for the residuals obtained in step S2. Training of ANN1 in step S3 is performed for all available products j. Thus, within the ANN model architecture, the effects of product quality parameters and the effects of various process parameters are separated. An ANN that models or predicts the effects of educt parameters is y i An advantage is that it can be trained with a larger dataset derived from the processing of the first product along y2 and the processing of the second product along y2. It is also conceivable to consider training datasets for the educt model ANN1 for further batch processes leading to further products.
[0055] In the next step, S4, a second ANN implementing models ANN21 and ANN22 is retrained. The additional learning in step S3 uses additional input from the prediction residuals obtained from the first model ANN1. Therefore, the prediction accuracy of product quality parameters Q1 and Q2 is further improved.
[0056] In the embodiment, steps S3 and S4 are performed repeatedly. The architecture of ANN systems 7 and 8 allows for handling sparse training datasets and enables a robust black-box model for controlling batch plants. In particular, in batch plants of multiple products, where shared or frequently used educts are present, ANN configuration data may be reused or recycled.
[0057] Figure 7 is a schematic diagram of one embodiment of a batch plant control system using the ANN system 7, 8 or an improved version thereof. Figure 7 shows a control device 13 coupled to a batch plant, for example, as shown in Figure 1. Similar or analogous reference figures are used and will not be explicitly described again. The control device 13 controls the educt quality parameter x i , and predetermined product quality parameter Q' pThe reactor receives the following: Here, the specified product quality parameter means the desired quality of the product produced by reactor 2. Reactor 2 receives the process parameter y for a particular product p. p The control device 16 is, for example, a device implemented in a computer and performs operations that implement an ANN system as disclosed above or below. The control device 13 further measures the observed quality parameter Q of the product. p However, the product quality parameter Q' p An optimization algorithm is performed to match the predetermined product quality according to the specifications. Next, various aspects of the plant simulation / control shown in Figure 7 will be described in detail.
[0058] Figure 5 shows multiple educts and their associated educt quality parameters X. 1,2,3,4 Based on this, we present an ANN system, setup, or architecture that can simulate various products having product quality parameters Q1, Q2, and Q3. 1,2,3,4 This refers to the ANN educt model corresponding to ANN1 in Figure 3, for example, g 1,2,3 This refers to the process model implemented by the second ANN (e.g., ANN21 and ANN22 in Figure 3). 1,2,3 This refers to a linear regressor that models a chemical reaction involving a sample of educt. For example, regressor h1 receives outputs from educt models f1 and f2. We can generalize the architecture shown in Figure 5, which has multiple first ANNs, as well as four first ANNs corresponding to educts denoted by f1, f2, f1, f2, f3, and f4, each characterized by educt quality parameters X1, X2, X3, and X4, respectively. Next, multiple second ANNs labeled g1, g2, and g3 receive combinations of specific process parameters y1, y2, and y3 corresponding to batch processes leading to products 1, 2, and 3, respectively, with weighted educt influence parameters R1, R2, R3, and R4, and associated product quality parameters Q1, Q2, and Q3. In general, we can consider j=1...p products and i=1...m educts.
[0059] Figure 6 shows a representation of a general algorithm using an ANN-based regression model. Algorithm 1 shown in Figure 6 provides a learning method for a generalized ANN architecture, particularly based on ANN systems 7, 8, and 12. The n production process of product j. j Assume there are 1 sample or dataset. The model's input parameters are labeled as shown in Equation 4,
number
[0060]
number
[0061] The learning algorithm determines whether the model is a regressor or an ANN.
number
[0062]
number
[0063] In lines 1-6 of Algorithm 1 shown in Figure 6, the ANN model
number
[0064]
number
number
[0065] Next, for all product p in lines 8 through 10 of Algorithm 1, we apply the regression model given by Equation 4.
number
[0066] According to the applicant's research, the appropriate ANN architecture for the first and second ANN models is a single-hidden-layer neural network trained by Basian control. Thus, an efficient neural network-based simulation method and batch plant control capability are obtained. The disclosed ANN models and the software libraries for training them are available in computer implementation form. For example, MATLAB DeepLearningToolbox can be referenced for configuring and operating the ANNs disclosed herein.
[0067] The advantage of the disclosed method and system is that it can predict product quality even when only sparse datasets are available. According to the applicant's research, simulations of batch plants based on the disclosed approach achieve accuracy equal to or greater than that obtained from white-box models, when white-box models are available. Therefore, a flexible and efficient tool is provided for the simulation, prediction, and control of chemical processes, for example, those deployed in batch processing.
[0068] While the simulation of batch plant operation is disclosed using an ANN in an ANN system, the present invention is not limited to such a regressor. For example, alternative configurations of the regression system can be considered for performing regressions including Gaussian processes, linear regression, elastic net regularization models, etc. For example, any ANN in this disclosure can be replaced with an appropriate regression unit. It is understood that the regressor may be implemented as a software service, a hardware unit, or a distributed computer network. [Explanation of symbols]
[0069] 1 batch plant 2 Reactors 3, 5 Educt 4.6 Products 7, 8, 12 ANN system 9, 10, 11 ANN 13 Control device S1 2ANN Training S2 Calculation of residuals S3 First ANN training S4 2ANN Retraining
Claims
1. A computer-implemented regressor (7) for the simulation, monitoring and / or control of plant (1), wherein plant (1) is Related Educt Quality Parameters (x 1 , x 2 A step of receiving one or more educts (3, 5) having ) A step of processing the aforementioned educts (3, 5), wherein the processing includes a step having a related process parameter (y), Related product quality parameters (Q 1, Q 2 The system is implemented to perform the process of outputting products (4, 5) having the following characteristics: The aforementioned regressor (7) is It has at least two regressor units (9, 10), and each regressor unit (9, 10) is An input section for receiving input data, It has an output unit for outputting output data, The first regressor unit (9) controls the educt quality parameter (x i ) receives at least one educt effect parameter (R 1 It is implemented to output ), The second regressor unit (10) is implemented to receive the educt influence parameter (R 1 ) and the process parameter (y j ) and output at least one product quality parameter (Q j ). The first and second regressor units (9, 10) are based on machine learning principles, and the regressor (7).
2. The first and second regressor units (9, 10) are artificial neural networks (ANNs) (9, 10), each including an input layer having input nodes for receiving input data and an output layer having output nodes for outputting output data. The regressor according to claim 1, wherein the second ANN(10) is trained on a training dataset that includes process parameters and product target variables corresponding to product quality parameters associated with each product.
3. The regressor according to claim 2, wherein the first ANN(9) is trained on a training dataset that includes the educt quality parameters and the residuals of the second ANN(10) trained as target variables for the educt influence parameters, both corresponding to their respective products.
4. A plurality of first ANNs, each of the first ANNs corresponds to an educt and the corresponding educt quality parameter (X i=1…4 ) receives at least one corresponding educt effect parameter (R i=1…4 Multiple first ANNs are implemented to output ), A plurality of second ANNs, each of the second ANNs corresponds to a process for manufacturing a product and the corresponding process parameter (Y j=1…3 ) and the combination of the educt influence parameters from the first ANN (R i=1…4 ) receives the corresponding product quality parameter (Q j=1…3 A regressor according to any one of claims 1 to 3, comprising a plurality of second ANNs implemented to output ).
5. The regressor according to claim 2 or 3, wherein at least one of the ANNs is a feedforward ANN, a Basian neural network, and / or at least one of the ANNs further comprises hidden nodes.
6. The Reglessa according to any one of claims 1 to 3, wherein the educt quality parameter includes at least one of viscosity value, hydroxyl value, concentration value, and color parameter.
7. The aforementioned process parameter (y j The regressor according to any one of claims 1 to 3, wherein the regressor includes at least one of the measured observed values, temperature values, maximum temperature values, time span, reaction time, catalyst storage time, number of free isocyanate (NCO) groups, and time series characteristics.
8. The aforementioned product quality parameter (Q j The regressor according to any one of claims 1 to 3, wherein the regressor comprises at least one of viscosity value, hardness value, roughness value, drug interaction, pH value, and solubility.
9. A control device (13) for controlling a plant (2), wherein the plant (2) controls the associated educt quality parameters (x i A step of receiving one or more educts having ) and a step of processing the educts, wherein the processing is related to process parameters (y j A process having ) and related product quality parameters (Q j The system is implemented to perform the process of outputting a product having ) and, The control device (13) has a regressor (1, 7, 8, 12) according to any one of claims 1 to 3, The control device (13) receives the product quality parameter (Q) output from the regressor unit (10). j A control device, implemented as a function of ), to adapt a process for processing the educt in response to adapted process parameters.
10. The control device (13) performs the operation of mounting the regressors (1, 7, 8, 12), and outputs the product quality parameter (Q) from the regressor unit (10). j The control device according to claim 9, comprising a computer processing unit implemented to execute an optimization algorithm that adapts process parameters to correspond to a predetermined product quality.
11. A method for training a regressor according to claim 1, wherein the regressor (9, 10) is a machine learning unit, and the method is For the first and at least one second product (j=1...p), multiple n p A step of providing training datasets, wherein each training dataset includes a product target variable (Q) corresponding to an educt quality parameter (x), a process parameter (y), and a product quality parameter associated with the product (j), The process involves training the second regressor unit (10) and / or the third regressor unit based on a training data subset including process parameters (y) and product target variables (Q) corresponding to the first product (j=1), thereby obtaining a first residual (R) for each training data subset. The process involves training the regressor unit (10) based on a training data subset containing process parameters (y) and product target variables (Q) corresponding to at least one further product (j≠1), thereby obtaining a second residual (R) for each training data subset, and A method comprising the step of training the first regressor unit (9) based on a training data subset including an educt quality parameter (x) and the residual (R) as a target variable for the educt influence parameter (both corresponding to the first product (j=1)) and a further training data subset including the educt quality parameter (x) and the residual (R) as a target variable for the educt influence parameter (both corresponding to the at least one further product (j≠1)).
12. The method according to claim 11, further comprising training the second regressor unit (10) for each product (j) based on a subset of training data including process parameters (y), educt influence parameters output from a first ANN (9) trained in response to educt quality parameters (x) related to the educt used to manufacture each product, and product target variables (Q) corresponding to product quality parameters related to each product.
13. The method according to claim 12, wherein the steps of training the first regressor unit (9) and training the second regressor unit (10) are repeated.
14. The training dataset is generated by operating the plant (1) and measuring process parameters and product quality parameters, and / or The method according to any one of claims 11 to 13, further comprising generating the training dataset, simulating a plant process based on educt quality parameters, and developing a white-box numerical model for generating process parameters and product quality parameters.
15. A method for simulating, monitoring and / or controlling a plant (1), wherein the plant (1) is Related Educt Quality Parameters (x i A step of receiving one or more educts (3, 5) having ) A step of processing the aforementioned educt, wherein the process parameter (y) related to the processing j A process having, Related product quality parameters (Q j The system is implemented to perform the process of outputting a product (4, 5) having ) and The aforementioned method, A method comprising using the regressors (7, 8, 12) described in claim 1, wherein the regressors (7, 8, 12) are trained according to the method described in claim 11 or 12.