Monitoring and / or control of plants via machine learning regressors

Abstract

An embodiment of a computer-implemented regressor (7) for simulating, monitoring and / or controlling a batch plant (1) is disclosed, the batch plant (1) receiving one or more educts (3, 5) having associated educt quality parameters (x1, x2) and a process that is associated with associated process parameters (y j ) to process said educts (3, 5) and output products (4, 5) with associated product quality parameters (Q1, Q2). The regressor (7) comprises at least two regressor units (9, 10) based on machine learning principles, each regressor unit (9, 10) having an input for receiving input data and an output for outputting output data. The first regressor unit (9) processes said educts (3, 5) with a i ) and outputting at least one educt influence parameter (R1). A second regressor unit (10) is implemented to compute the educt influence parameter (R1) and the process parameter (y j ) and at least one product quality parameter (Q j ) is implemented.

Description

[Technical field]

[0001] The invention relates in particular to a method including an artificial neural network system, a regressor and a controller for simulating, monitoring and / or controlling a plant. Furthermore, a method for training an artificial neural network system is provided. In particular, a model based on an artificial neural network (ANN) is suitable for simulating a batch process in which an educt is converted into a product by a batch processing unit and is suitable for monitoring and / or controlling a batch plant. [Background technology]

[0002] In a batch process or batch plant, the production of several products takes place in the same set of equipment or processing units, for example chemical or biological reactors. On the one hand, it is desirable to optimize a particular process involving a particular educt that is processed into a desired product. On the other hand, the scheduling of batch operations using a single processing unit can be improved if the process can be accurately modeled and simulated.

[0003] Traditionally, the simulation of chemical reactions occurring in batch processes has relied on rigorous models or so-called white-box models based on first principles, which typically require significant computational power and resources to model complex physicochemical systems that involve nonlinear equations.

[0004] Another approach is the so-called black-box models, developed in the past and relying on machine learning concepts. For example, neural networks can be used to predict certain properties of a product or its quality based on input data, including process characteristics and educt properties. Such neural networks need to be trained with multiple data sets, which are sometimes not available. Moreover, traditional black-box models based on artificial neural networks need to be retrained if the conditions in the batch plant being modeled, the desired product quality and / or the educt properties change, making the traditional approach inflexible.

[0005] A further approach is a hybrid model that combines white-box and black-box models. Document WO2020 / 227383A1 discloses a method and system for computer-based process modeling and simulation that combines first-principles and machine learning models to help when either model is missing. In one example, measured input values ​​are adjusted by first-principles techniques. A machine learning model of the target chemical process is trained with the adjusted values. In another example, the machine learning model represents the residual between the predictions of the first-principles model and the empirical data. The residual machine learning model modifies the physical phenomenon predictions in the first-principles model of the chemical process. In another example, the first-principles simulation model uses the process input data and the predicted values ​​of the machine learning model to generate a simulation result of the chemical process.

[0006] Document WO2020 / 058237A2 discloses a method and system capable of predicting the value of a product quality attribute of a compound or its formulation as a result of a multi-stage manufacturing process, where the entire process or process steps are characterized by process parameters. This is achieved by performing multivariate data analysis of the process data in a quality prediction model, which identifies or expresses the mathematical relationship between the quality attribute 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 neural network models combined with empirical process knowledge obtained over time. [Prior art documents] [Patent documents]

[0007] [Patent Document 1] WO2020 / 227383A1 [Patent Document 2] WO2020 / 058237A2 Summary of the Invention [Problem to be solved by the invention]

[0008] It is an object of the present disclosure to provide improved methods and systems for monitoring and / or controlling a plant. [Means for solving the problem]

[0009] The aspects of the independent claims solve this problem.

[0010] The present disclosure provides a computer-implemented controller for simulating, monitoring and / or controlling a plant. The plant may be implemented to receive one or more educts with associated educt quality parameters, process the educts, the process of processing having associated process parameters, and output a product with associated product quality parameters. The regressor comprises at least two regressor units, each regressor unit comprising an input for receiving input data and an output for outputting output data. Within the regressor units, 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 parameters and the process parameters and output at least one product quality parameter. The regressor and / or the regressor units are based on machine learning principles.

[0011] In an 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 with associated educt quality parameters, a process to process said educts with associated process parameters and output a product with associated product quality parameters. The ANN system has at least two ANNs, each ANN having an input section with an input node for receiving input data and an output section with an output node for outputting output data. A first ANN is implemented to receive said educt quality parameters and output at least one educt impact parameter. A second ANN is implemented to receive said educt impact parameters and said process parameters and output at least one product quality parameter.

[0012] Applicant has discovered that nested or concatenated ANNs are suitable regressors for solving the separable problems posed by batch plants, and therefore the disclosed regressors can be implemented as ANN systems that include ANNs.

[0013] In the proposed ANN system, data relating to the educts, i.e. educt quality parameters, and process data relating to the actual process carried out in the batch plant, i.e. process parameters, are considered separately. A first ANN models / simulates the influence of the educt properties in terms of the educt quality parameters, while a second ANN models / simulates e.g. chemical or biological processes, which are at least partly driven by process parameters that can be observed or set during the operation of the batch plant. The second ANN also receives input from the first ANN so that it can reliably predict the product quality parameters.

[0014] Each ANN is understood to include 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 hidden layers between the input layer and the output layer. Typically, each ANN is characterized by configuration data that includes at least bias and weight values ​​for each node in the ANN. This configuration data is obtained by training each ANN.

[0015] The trained ANN may also be referred to as a model of the chemical reactions occurring in the process units of a batch plant. The trained ANN may also be considered as a regressor for the basic problem of mapping input data (e.g., educt or process properties) to output or target data (e.g., product properties).

[0016] The proposed approach of separating the complex problem of mapping various educt quality parameters and process parameters to product quality parameters allows for efficient use of sparse data sets in training the ANN or model, respectively. An embodiment of the ANN system allows for sharing training data sets used for different products produced in the same or similar batch plant. In particular, a process model or a second ANN can benefit from a first ANN or educt model trained for similar or the same educt used in a different batch plant process. For example, the ANN system can change the educt model by updating the configuration data of the first ANN when the educt is changed, but the overall batch process modeled by the second ANN is essentially unchanged. Thus, the learned parts of the batch plant process ANN remain. 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 parameters and process parameters to target or model product quality parameters.

[0017] The applicant's research has shown that the resulting model or neural network system for simulating a batch plant is robust to noise in the training data. In an embodiment, the partial black-box model for the first ANN or educt model describes the influence of the educts independent of the plant or chemical process setup. Thus, the first ANN can be used in connection with an alternative batch plant or batch plant unit modeled by an alternative second ANN. Thus, in particular, the trained first ANN can be reused and only sparse training data for setting up the second ANN or process model for the alternative batch plant needs to be used.

[0018] In an embodiment, the output product quality parameters are generated in computer readable form, displayed, and / or used to control, schedule, or adapt the batch plant, particularly by a controller.

[0019] In an embodiment, the second ANN is based on a training data set that includes process parameters and product target variables that correspond to product quality parameters associated with the respective products. For example, the process parameters may include measurements observable during the chemical reaction, temperature values, maximum temperature values, time spans, reaction spans, storage times of the catalyst, number of free isocyanate groups, or other characteristics of the time series. Other process parameters that affect the chemical reaction in the batch reactor may also be considered.

[0020] Product quality parameters include viscosity value, hardness value, roughness value, drug interaction effects, pH value or solubility of the product. During chemical processes or reactions in batch plants, other product quality parameters can also be taken into account that characterize the obtained product or products.

[0021] In an embodiment of the ANN system, a first ANN is trained based on a training data set that includes educt quality parameters and residuals of a second ANN trained as target variables for educt impact parameters, both educt quality parameters and residuals corresponding to the respective product. The residuals or prediction errors of the second ANN may be due to sparse availability of a training data set, which is used to train the first ANN that models the educt quality. Educt quality parameters include viscosity values, hydroxyl numbers, concentration values, color parameters, etc. Other quality parameters are also conceivable.

[0022] The ANN system can be seen as a linked system of a first ANN and a second ANN. In an embodiment of the ANN system, the system comprises: a plurality of first ANNs, each of the first ANNs corresponding to an educt, implemented to receive a corresponding educt quality parameter, and implemented to output at least one corresponding educt impact parameter; and a plurality of second ANNs, each of the second ANNs corresponding to a process for manufacturing a product and implemented to receive a combination of a corresponding process parameter and an educt impact parameter from the first ANN and to output a corresponding product quality parameter;

[0023] It is also conceivable to associate a first ANN with different educts and a second ANN with a batch process with certain associated process parameters. The outputs from the first ANNs are combined, weighted and fed to the second ANN. For example, each second ANN can receive a linear combination of educt influence parameters from the first ANN. As a result, the ANN system can predict product quality parameters of products produced in a batch process by combining one or more educts, where the educts are characterized by educt quality parameters.

[0024] An "ANN corresponding to an action" is understood to be an ANN that models the action, and thus a "second ANN corresponding to a product" is an ANN that is trained and configured to output approximate product quality parameters in response to process parameters associated with the manufacturing process of the given product.

[0025] In an embodiment of the ANN system, at least one of the ANNs is a feed-forward ANN. In an embodiment, a Bayesian neural network can be used as the ANN. At least one ANN can be considered to further include hidden nodes between the input nodes and the output nodes.

[0026] According to one aspect of the present disclosure, a controller is proposed for controlling a batch plant. The batch plant is implemented to receive one or more educts with associated educt quality parameters, process said educts with associated process parameters, and output a product with one or more associated product quality parameters. The controller comprises an ANN system as disclosed above or below with respect to an embodiment, the controller being implemented to adapt the process as a function of the product quality parameters output from a second ANN in response to the adapted process parameters. For example, based on a simulation of the ANN system of a process in the batch plant, the controller alters the process and thus the associated process parameters to obtain a desired product quality.

[0027] In an embodiment of the controller, the controller comprises a computer processing device implemented to perform operations implementing the ANN system and to execute an optimization algorithm for adapting process parameters such that product quality parameters output from the second ANN correspond to a predetermined product quality.

[0028] Another aspect of the present disclosure provides a method for training a regressor, e.g., in view of an ANN system as disclosed above or below with respect to certain embodiments. The training method comprises at least one of the following steps: providing a plurality of training data sets for the first product and at least one second product, each training data set including product target variables corresponding to educt quality parameters, process parameters, and product quality parameters associated with the product; training a second ANN based on training data subsets including process parameters and product target variables corresponding to the first product, thereby obtaining a first residual for each training data subset; training a second ANN based on training data subsets including process parameters and product target variables corresponding to at least one additional product, thereby obtaining a second residual for each training data subset; and Training a first ANN based on a training data subset comprising residuals as target variables of educt quality parameters and educt impact parameters corresponding to the first product and based on a further training data subset comprising residuals as target variables of educt quality parameters and educt impact parameters corresponding to at least one further product.

[0029] The training method may comprise training a first ANN based on a training data set of all products containing a particular educt, targeting the quality parameters of this educt and residuals from previous (training) steps, preferably for each educt.

[0030] Aspects of the above training methodology reduce the need for a large number of training data subsets or sets because a first ANN is trained to provide residuals for a second ANN. In an embodiment of the training method, a step is performed of training a second ANN based on a training data subset including, for each product, the process parameters, the educt impact parameters output from a first ANN trained in response to educt quality parameters associated with educt use to manufacture the respective product, and product target variables corresponding to product quality parameters associated with the respective product. Thus, the second ANN receives additional input from the trained first ANN.

[0031] In an embodiment of the training method, steps of training a first ANN and training a second ANN are performed iteratively based on a training data subset including process parameters, educt impact parameters from a first ANN trained on educt quality parameters associated with educt use to manufacture the respective products, and product target variables corresponding to product quality parameters associated with the respective products.

[0032] In an embodiment, the training method further comprises: generating said training data set by operating a batch plant and measuring process parameters and product quality parameters; and / or Generating the training data set, simulating a batch plant process based on the quality parameters, and developing a white-box-numerical model for generating process parameters and product quality parameters.

[0033] According to one aspect of the disclosure, there is provided a method for simulating, monitoring and / or controlling a batch plant, the batch plant receiving one or more educts having associated educt quality parameters, a process being implemented to process said educts having associated process parameters and output a product having associated product quality parameters, the method comprising using an ANN system as described above or below with respect to certain embodiments, and the ANN system being trained according to a training method as disclosed above or below with respect to the embodiments.

[0034] It will be understood that the regressor unit is not necessarily implemented as an ANN, and the present disclosure encompasses other suitable configurations of regressor units. Thus, throughout this disclosure, the term "ANN" may be substituted with "regressor unit" to fully appreciate the scope of the invention. The regressor unit may be implemented, for example, as a computer-implemented regression method in terms of a software function or service.

[0035] According to a further aspect, the present disclosure relates to a computer program product including computer readable instructions that, in response to execution of the machine readable instructions, cause a computing system including one or more processing devices to perform the above-described methods and functions for simulating, monitoring and / or controlling a batch plant.

[0036] In an embodiment, the computer program product comprises program code for performing the above-mentioned methods and functions by the computerized control device when executed on at least one control computer. The computer program product, such as a computer program means, may be embodied as a memory card, a USB stick, a CD-ROM, a DVD or as a file that can be downloaded from a server in a network. For example, such a file may be provided by transferring the files constituting the computer program product from a wireless communication network.

[0037] Further possible embodiments or alternative solutions of the invention include combinations of features, not explicitly mentioned herein, described above or below with respect to the embodiments, and those skilled in the art may also add individual or isolated aspects and features to the invention in its most basic form.

[0038] Further embodiments, features and advantages of the present invention will become apparent from the following description and dependent claims, taken in conjunction with the accompanying drawings. [Brief description of the drawings]

[0039] [Figure 1] FIG. 1 is a schematic diagram illustrating one embodiment of a batch plant. [Diagram 2] FIG. 1 is a schematic diagram of a first embodiment of an ANN system. [Diagram 3] FIG. 2 is a schematic diagram of a second embodiment of an ANN system. [Figure 4] 1 is a flow chart including method steps for training an embodiment of an ANN system. [Diagram 5] FIG. 1 is a schematic diagram of a third embodiment of an ANN system. [Figure 6] 1 shows an algorithm for training an embodiment of an ANN system. [Figure 7] FIG. 1 is a schematic diagram of an embodiment of a control system for a batch plant. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[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 stirred tank reactor. The tank reactor 2 can be used to process various educts under specific process parameters to produce products. The products can be for example intermediate products of prepolymer synthesis, polyols, coating agents, or other chemical, pharmaceutical, or biological compositions.

[0041] In the example of FIG. 1, one input educt 3 is shown to be processed in reactor 2 into product 4. Specific educt quality parameters can be set for educt 3, indicated by label x1. Educt quality parameters include specific properties of educt 3, such as reactivity, hydroxyl number, specific isomers, etc. In reactor 2, a chemical reaction process takes place, which is characterized by a process parameter y. After the reaction process in reactor 2 is finished or stopped, a product 45 can be recovered having a specific quality parameter, indicated as Q1. For example, quality parameter Q1 may refer to the purity or concentration of a substance in product 4.

[0042] The batch processing device 2 can also be used for other products. For example, in dashed lines an alternative educt 5 is depicted having an educt quality parameter x2. The batch processing results in an alternative product 6 having a product quality parameter Q2. It is desired to predict the effect of the educts 3, 5 and the applied process parameter y on the products 4, 6, and in particular on their product quality parameter. The problem can be written as follows:

[0043]

number

[0044] where j represents the jth product of all potential products {1, 2, ... p} in this disclosure. A model or simulation of a batch process is a process that is characterized by the quality parameters x of the educt and the process parameters y j The combination of these is the product quality parameter Q j The input vector x describes the quality measurements of the educts, e.g. substance concentration, viscosity measurements, color parameters, etc. The input vector y j denotes the process parameters that can be measured during the production of the product in the reactor 2. Time series features such as maximum, minimum, average values, or physical observations such as temperature and pressure values ​​can be considered as process parameters. The output product quality parameter Q jis a scalar and represents the desired quality of product j.

[0045] Figure 2 is a schematic diagram of a first embodiment of an ANN system implemented to model or simulate a batch process, for example in the reactor 2 of Figure 1. The ANN system 7 is composed of a first ANN 9 and a second ANN 10. The underlying ANN models or regressors are 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 configuration of the ANNs 9, 10 is defined by configuration data (not shown) including bias and weight values ​​for each node of the respective ANNs.

[0046] The first ANN9 is a process for determining a predetermined product quality parameter Q j of the educt involved in the production of the desired product having the educt quality parameter x i The first ANN 9 receives the educt influence parameter R i to the input node of the second ANN 10. The second ANN 10 further outputs the process parameter y j The second ANN 10 receives a predicted quality parameter Q j Thus, an embodiment of the ANN system 7 outputs the educt quality parameter x i and the process parameter y j The product quality parameters Q j Assume that the problem depicted in Equation 1 can be written as

[0047]

number

[0048] Equation 2 expresses the product quality parameter Q j Assuming that the mapping or function of is separable, the function f is the pipe quality parameter x i The function g depends on the process parameter yj It depends on the educt quality parameter x i It would be advantageous if the ANN system 7 in FIG. 2 solves a regression problem that can be written as:

[0049]

number

number

number

[0050] In general, the ANN system 7 shown in FIG. 2 is capable of simulating a batch process or a reactor 2, respectively. The ANN system 7 is configured on the one hand to simulate the educt quality parameter x i (ANN1), on the other hand, to model the effect of the process parameter y j By configuring the first and second ANNs 9, 10 specialized for modeling the influence of (ANN 21), efficient learning and setting of configuration data for the ANNs 9, 10 is possible.

[0051] 3 is a schematic diagram of a second embodiment of an ANN system 8. The ANN system 8 includes a first ANN 9 and two second ANNs 10, 11, labelled ANN21, ANN22. The labels indicate the model implemented through the configuration data of the ANNs (biases, weights, number of nodes, topology). The first second ANN 10 receives the educt influence parameter R from the first ANN 9. iand receives process parameters y1 describing or characterizing a batch process leading to product j=1 having product quality parameter Q1. Another, second ANN11 receives alternative process parameters y2 characterizing a process leading to an alternative product j=2 having product quality parameter Q2. The models ANN1, ANN21 and ANN22 are implemented as shallow neural networks, for example Bayesian type neural networks, eliminating the need for validation of the data set when training.

[0052] Figure 4 shows a flow chart including steps for training the ANN systems 7, 8 shown in Figure 2 or Figure 3. In a first step S1, the ANN21, ANN22 models are trained separately from each other, for example using a Bayesian learning method. The respective training data set for ANN21 includes a process parameter y1 and a predefined product quality parameter Q i is included as a target variable. Multiple training data sets may be used for each product. In step S1, the same learning process is performed for model ANN22, with the process parameter y2 included or the measured product quality parameter Q2 placed as the designed target variable. For example, due to the limited amount of training data and the missing educt influence input, models ANN21 and ANN22 produce residuals or errors in their quality predictions.

[0053] In the next step S2, the residuals of the target variables, i.e. Q1 and Q2, are calculated.

[0054] Next, in step S3, the ANN model ANN1 calculates the educt quality parameter x i, and the in-pipe quality parameters as target variables for training the residuals obtained in step S2. The training of the ANN 1 in step S3 is performed for all available products j. Thus, within the ANN model architecture, the influence of the product quality parameters and the influence of the various process parameters are separated. The ANN that models or predicts the influence of the educt parameters, y i It is an advantage to be able to train with a larger data set resulting from the processing of a first product along y and a second product along y2. It is also conceivable to consider training data sets for the educt model ANN1 for further batch processes leading to further products.

[0055] In the next step S4, a second ANN implementing the models ANN21, ANN22 is retrained. The additional learning in step S3 uses additional inputs from the prediction residuals obtained from the first model ANN1. Thus, the prediction accuracy of the product quality parameters Q1, Q2 is further improved.

[0056] In an embodiment, steps S3 and S4 are performed iteratively. The architecture of the ANN system 7, 8 allows for sparse training data sets and still allows for robust black-box models for controlling batch plants. ANN configuration data may be reused or recycled, especially in multi-product batch plants where shared or frequently used educts are deployed.

[0057] Figure 7 is a schematic diagram of an embodiment of a control system for a batch plant using an ANN system 7, 8 or improvements thereof. Figure 7 shows a controller 13 coupled to a batch plant, for example as shown in Figure 1. Like or similar reference numerals are used and will not be explicitly described again. The controller 13 controls the educt quality parameter x i , and a given product quality parameter Q' pHere, the predetermined product quality parameters refer to the desired quality of the product produced by the reactor 2. The reactor 2 receives the process parameters y for a particular product p. p The controller 16 may be, for example, a computer-implemented device that performs operations to implement an ANN system as disclosed above or below. The controller 13 may further operate and / or characterize the product by measuring the observed quality parameter Q p But the product quality parameter Q' p An optimization algorithm is performed to match a given product quality according to Next, various aspects of the simulation / control of the plant shown in Figure 7 are described in detail.

[0058] Figure 5 shows several educts and their associated educt quality parameters X 1,2,3,4 Based on the above, an ANN system or setup or architecture capable of simulating various products having product quality parameters Q1, Q2, Q3 is shown. 1,2,3,4 refers to the ANN educt model, e.g., corresponding to ANN1 in Fig. 3, and g 1,2,3 refers to the process model implemented by the second ANN (e.g., ANN21, ANN22 in FIG. 3). 1,2,3 refers to a linear regressor that models a chemical reaction involving a sample of educts. For example, regressor h1 receives the output from educt models f1 and f2. One can generalize the architecture shown in Fig. 5 with multiple first ANNs, as well as four first ANNs, denoted by f1, f2, f1, f2, f3, f4, corresponding to educts characterized by educt quality parameters X1, X2, X3, X4, respectively. Then, multiple second ANNs, labeled by g1, g2, g3, receive combinations of weighted educt influence parameters R1, R2, R3, R4 and specific process parameters y1, y2, y3 corresponding to the batch process leading to products 1, 2, 3, respectively, with associated product quality parameters Q1, Q2, Q3. In general, one can consider j=1...p products and i=1...m educts.

[0059] A general algorithmic representation using an ANN-based regression model is shown in Figure 6. Algorithm 1 shown in Figure 6 provides a learning method for a generalized ANN architecture based specifically on ANN systems 7, 8, and 12. j Assume we have samples or datasets. The input parameters of the model are labeled as in Equation 4,

number

[0060]

number

[0061] The result of the learning algorithm, the regressor or ANN model

number

[0062]

number

[0063] In lines 1 to 6 of Algorithm 1 shown in Fig. 6, the ANN model

number

[0064]

number

number

[0065] Next, for every product p in lines 8 to 10 of Algorithm 1, we use the regression model

number

[0066] According to the applicant's research, a suitable ANN architecture for the first and second ANN models is a single hidden layer neural network trained by Bayesian control. Thus, an efficient neural network based simulation method and control capability of the batch plant is obtained. Software libraries for implementing the disclosed ANN models and their training are available in computer implemented form. For example, reference can be made to the MATLAB Deep Learning Toolbox for configuring and operating the ANNs disclosed herein.

[0067] An advantage of the disclosed method and system is that it allows predicting product quality even when only sparse data sets are available. The applicant's research shows that simulations of batch plants based on the disclosed approach reach an accuracy comparable to or better than that obtained from white-box models, when such models are available. Thus, a flexible and efficient tool is provided for the simulation, prediction and control of chemical processes deployed, for example, in batch processes.

[0068] Although the simulation of batch plant operations has been disclosed using an ANN in an ANN system, the invention is not limited to such a regressor. Alternative configurations of regression systems for performing regression can be envisioned, including, for example, Gaussian processes, linear regression, elastic net regularized models, and the like. For example, any of the ANNs disclosed herein can be replaced with an appropriate regression unit. It is understood that the regressor can be implemented as a software service, a hardware unit, or a distributed computer network. [Explanation of symbols]

[0069] 1. Batch plant 2. Reactor 3, 5 Educt 4, 6 Products 7, 8, 12 ANN system 9, 10, 11 ANN 13 Control device S1 Second ANN training Calculating S2 residuals S3 Training the first ANN S4 Retraining the second ANN

Claims

1. A computer-implemented regressor (7) for simulation, monitoring and / or control of a plant (1), wherein the plant (1) comprises: Receiving one or more educts (3, 5) having associated educt quality parameters (x 1 , x 2 ); a step of processing the educts (3, 5), the processing having associated process parameters (y); Output a product (4, 5) having a related product quality parameter (Q 1, Q 2 ), and is implemented to perform the steps of the regressor (7) comprises: at least two regressor units (9, 10), each regressor unit (9, 10) having: an input for receiving input data; an output for outputting output data; The first regressor unit (9) receives the educt quality parameter (x i ), and is implemented to output at least one educt influence parameter (R 1 ). 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) being based on machine learning principles, the regressor (7).

2. The first and second regressor units (9, 10) are artificial neural networks (ANNs) (9, 10), each comprising an input layer having input nodes for receiving input data and an output layer having output nodes for outputting output data, the second ANN (10) being trained based on a training data set comprising process parameters corresponding to product quality parameters associated with respective products and product target variables, the regressor according to claim 1.

3. The first ANN (9) is trained based on a training data set comprising the educt quality parameters and the residuals of the second ANN (10) trained as target variables of the educt influence parameters, both corresponding to respective products, the regressor according to claim 2.

4. A plurality of first ANNs, each ANN of the first ANNs corresponding to an educt and receiving a corresponding educt quality parameter (X i=1…4 ), and being implemented to output at least one corresponding educt influence parameter (R i=1…4 ); a plurality of first ANNs A plurality of second ANNs, each ANN of the second ANNs corresponding to a process for manufacturing a product, and corresponding process parameters (Y j=1…3 ) and a combination of educt influence parameters (R i=1…4 ) received from the first ANN, and implemented to output corresponding product quality parameters (Q j=1…3 ), a plurality of second ANNs, the regressor according to any one of claims 1 to 3, having.

5. The regressor according to claim 2 or 3, wherein at least one of the ANNs is a feedforward ANN, a Bayesian neural network, and / or at least one of the ANNs further comprises hidden nodes.

6. The regressor according to any one of claims 1 to 3, wherein the educt quality parameters comprise at least one of a viscosity value, a hydroxyl value, a concentration value, and a color parameter.

7. The process parameter (y j ) includes at least one of the measured observed value, temperature value, maximum temperature value, time span, reaction time, storage time of the catalyst, number of free isocyanate (NCO) groups, and characteristics of the time series, and the regressor according to any one of claims 1 to 3.

8. The product quality parameter (Q j ) includes at least one of a viscosity value, a hardness value, a roughness value, a drug interaction, a pH value, and a solubility, and the regressor according to any one of claims 1 to 3.

9. A control device (13) for controlling a plant (2), wherein the plant (2) receives one or more educts having associated educt quality parameters (x i ), and a step of processing the educts, the processing having associated process parameters (y j ), and a step of outputting a product having associated product quality parameters (Q j ), is implemented to perform. The control device (13) has a regressor (1, 7, 8, 12) according to any one of claims 1 to 3. The control device (13) is implemented to adapt a process of processing the educt in response to an adapted process parameter as a function of a product quality parameter (Q j ) output from the regressor unit (10).

10. The control device (13) executes an operation of implementing the regressor (1, 7, 8, 12), and adapts process parameters so that the product quality parameter (Q j ) output from the regressor unit (10) corresponds to a predetermined product quality, and comprises a computer processing device implemented to execute an optimization algorithm, the control device according to claim 9.

11. A method of training the regressor according to claim 1, wherein the regressor (9, 10) is a machine learning unit, the method comprising: For the first and at least one second product (j = 1... p), a plurality of n p Providing a training data set, each training data set including an educt quality parameter (x), a process parameter (y), and a product target variable (Q) corresponding to a product quality parameter associated with the product (j) Based on a training data subset including process parameters (y) and product target variables (Q) corresponding to the first product (j = 1), training the second regressor unit (10) and / or the third regressor unit, thereby obtaining a first residual (R) for each training data subset; Based on a training data subset including process parameters (y) and product target variables (Q) corresponding to at least one additional product (j ≠ 1), training the regressor unit (10), thereby obtaining a second residual (R) for each training data subset; and Training the first regressor unit (9) based on a training data subset including educt quality parameters (x) and the residual (R) as the target variable of the educt influence parameter (both corresponding to the first product (j = 1)), and a further training data subset including educt quality parameters (x) and the residual (R) as the target variable of the educt influence parameter (both corresponding to the at least one additional product (j ≠ 1)). A method comprising:

12. The method according to claim 11, further comprising training the second regressor unit (10) based on a training data subset including, for each product (j), 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 step of training the first regressor unit (9) and the step of training the second regressor unit (10) are repeatedly performed.

14. Operating the plant (1) and measuring process parameters and product quality parameters to generate the training data set, and / or The method according to any one of claims 11 to 13, further comprising generating the training data set, simulating a plant process based on educt quality parameters, and deploying a white-box - numerical model for generating process parameters and product quality parameters.

15. A method for simulation, monitoring and / or control of a plant (1), wherein the plant (1) is Receiving one or more educts (3, 5) having associated educt quality parameters (x i ); The step of processing the educt, the process being associated with process parameters (y j ), and Outputting a product (4, 5) having related product quality parameters (Q j ), and implementing to execute the steps The method is A method using the regressor (7, 8, 12) according to claim 1, wherein the regressor (7, 8, 12) is trained according to the method according to claim 11 or 12.

Images