Controlling production of a product

Abstract

A method (1) of providing a control dataset for controlling production of a product comprises: receiving (10), by a computing device (2), an input dataset comprising a first target attribute value related to a first product attribute of the product and a second target attribute value related to a second product attribute of the product; feeding (12), by the computing device (2), the input dataset into a model to obtain an output dataset from the model, the output dataset comprising a plurality of results; determining (14), by the computing device (2), for each of the plurality of results, an associated quality score; selecting (18), by the computing device (2), one or more of the plurality of results based on the associated quality scores; and deriving (19), by the computing device (2), the control dataset from the selected one or more results.

Description

Technical Field

[0001] This specification relates to embodiments of a method for providing a control dataset for controlling the production of a product (e.g., a chemical product), embodiments of a corresponding computing device, embodiments of a method for producing the product, and embodiments of a system for producing the product. Background Technology

[0002] The production of products (e.g., chemical products) may take place in safety-critical environments, meaning that the products should exhibit their designated target attributes with an extremely high probability, such as weight, purity, shelf life, size, taste, design, and therapeutic effects. Therefore, the control data upon which product production is based must be selected to ensure that this high probability is indeed achieved.

[0003] The abstract of M. Bortz et al.'s article, "Multi-criteria optimization in chemical process design and decision support by navigation on Pareto sets," Elsevier, Computer & Chemical Engineering, Vol. 60 (2014), pp. 354-363, states: "Designing a chemical process is a multi-criteria optimization problem with conflicting objectives. This problem can be solved efficiently using Pareto sets. These sets contain all solutions where improvement to any objective can only be achieved by accepting a deterioration of at least one other objective." The article presents an algorithm for determining Pareto sets in an advanced steady-state process simulator. A predefined precision approximation of the Pareto set is calculated; the Pareto set can be convex or non-convex. According to this article, decision-makers can interactively navigate the Pareto set and explore different optimal solutions. Therefore, their decisions will be rooted in an understanding of the entire Pareto set. Summary of the Invention

[0004] The independent claims present the subject matter. The dependent claims define the features of the exemplary embodiments.

[0005] According to a first aspect, a method for providing a control dataset for controlling product production is presented. The method includes: receiving an input dataset by a computing device. The input dataset includes: a first target attribute value related to a first product attribute of the product and a second target attribute value related to a second product attribute of the product; and data related to a first input variable and a second input variable, wherein a first dependency between the first input variable and the first product attribute and the second product attribute is different from a second dependency between the second input variable and the first product attribute and the second product attribute. The method further includes feeding the input dataset into a model by the computing device to obtain an output dataset from the model, the output dataset including a plurality of results. Each result indicates: a first result value of the first product attribute, a second result value of the second product attribute; and a third value of the first input variable and a fourth value of the second input variable, wherein the first value, second value, third value, and fourth value are correlated with each other. The method further includes: determining an associated quality score for each of the plurality of results by the computing device; selecting one or more results from the plurality of results by the computing device based on these associated quality scores; and obtaining the control dataset by the computing device from the selected one or more results.

[0006] According to the second aspect, a computing device is configured to perform the method according to the first aspect.

[0007] According to a third aspect, a method for producing a product is presented, wherein the production of the product is performed based on a control dataset provided according to the method of the first aspect. According to a fourth aspect, a system for producing a product is presented, wherein the system is configured to perform the method of the third aspect.

[0008] According to another aspect, this disclosure relates to a method (1) for providing a control dataset for monitoring and / or controlling the production and / or processing of a product, the method (1) comprising:

[0009] - The computing device (2) receives (10) an input dataset, which includes:

[0010] ○ The first target attribute value that is related to and / or associated with the first product attribute of the product and the second target attribute value that is related to and / or associated with the second product attribute of the product;

[0011] ○ Data associated with a first input variable and a second input variable, wherein one or more first dependencies between the first input variable and the first product attribute and the second product attribute are different from one or more second dependencies between the second input variable and the first product attribute and the second product attribute; and

[0012] - The input dataset is fed (12) into the model by the computing device (2) to obtain (13) an output dataset from the model, which includes multiple results, each indicating:

[0013] ○ The first result value of the first product attribute, the second result value of the second product attribute; and

[0014] ○ The third value of the first input variable and the fourth value of the second input variable; and

[0015] -The computing device (2) obtains, and in particular determines (14) the associated quality score for each of the plurality of results;

[0016] - The computing device (2) selects (18) one or more of the multiple results based on these associated quality scores; and

[0017] - The control dataset (19) is obtained by the computing device (2) from one or more selected results, and the control dataset is optionally provided, particularly for monitoring and / or controlling the production and / or processing of the product.

[0018] According to some embodiments, a method is proposed to determine the predictive uncertainty of an arbitrary surrogate model used to compute the Pareto front and incorporate this predictive uncertainty into the inverse prediction task. Therefore, not only is the prediction at any point determined, but also information about the uncertainty caused by factors such as data variance, model uncertainty, and Pareto front approximation uncertainty is identified. This allows for Pareto front browsing through a single descriptor that can be tailored to the user's risk tolerance.

[0019] Hereinafter, some exemplary features of the exemplary embodiments will be described. These optional features may be combined in any manner to obtain further embodiments, unless a particular combination is expressly excluded by the specification.

[0020] Computing devices can include a single computer or a distributed computing architecture, such as a server system, a computer / server farm, or any other computer cluster.

[0021] The product can be a chemical product, such as a product from one of the following product groups: paint, surfactant, coating, or cosmetic (such as cream).

[0022] The input dataset can be provided to the computing device via wired and / or wireless data paths. For example, the input dataset may be provided by a user using a user device (such as a terminal device).

[0023] As described above, the input dataset includes a first target attribute value related to a first product attribute of the product and a second target attribute value related to a second product attribute of the product. The input dataset further includes data related to a first input variable and a second input variable, wherein a first dependency between the first input variable and the first product attribute and the second product attribute is different from a second dependency between the second input variable and the first product attribute and the second product attribute.

[0024] For example, a product’s primary product attribute is a chemical, physical, or biological attribute, such as one of the following: weight, density, size, design, proportion of a certain material, purity, taste, flammability, toxicity, acidity, reactivity (of various types), heat of combustion, medical effects, etc.

[0025] The second product attribute can be different from the first product attribute. For example, the second product attribute in the product attributes of a product is a chemical attribute, physical attribute, or biological attribute, such as one of the following: weight, density, size, design, proportion of a certain material, purity, taste, flammability, toxicity, acidity, reactivity (of various types), heat of combustion, medical effects, etc.

[0026] In one embodiment, one or both of the target attribute values ​​are absolute values. In another embodiment, one or both of the target attribute values ​​define a range of values. In yet another embodiment, one or both of the target attribute values ​​define that the corresponding product attribute will be maximized, minimized, or otherwise optimized.

[0027] The first input variable may relate to, for example, the precipitate on which the product is based (e.g., a chemical compound, such as an acid, base, soap, or solvent). Additionally, the second input variable may relate to, for example, the precipitate on which the product is based. Alternatively, one or more of the input variables may relate to parameters of the production process used to produce the product (e.g., temperature, feed rate, control rate, or time). Alternatively, one or more of the input variables may relate to one or more intermediate products formed during the production process used to produce the product. For example, data related to more than two input variables may be received, such as data related to up to 50 input variables.

[0028] As explained above, the first dependency between the first input variable and the first and second product attributes differs from the second dependency between the second input variable and the first and second product attributes. Therefore, according to one or more embodiments, the first target attribute value, the second target attribute value, and the data associated with the first and second input variables can define a multi-objective optimization with conflicting objectives.

[0029] In the next step of the method, as described above, the input dataset is fed into the model to obtain the output dataset from the model. The output dataset includes multiple results, where each result indicates: a first result value for a first product attribute, a second result value for a second product attribute, a third value for a first input variable, and a fourth value for a second input variable. These four values ​​are related to each other because they belong to one of the multiple results.

[0030] For example, the output dataset includes at least a portion of the Pareto front (also known as the Pareto boundary or Pareto curve) already obtained based on the input dataset and the model. The Pareto front may include at least a portion of all Pareto efficient solutions.

[0031] In this embodiment, the model is generated based on the input dataset and / or selected from multiple available models. The model may be hosted by a computing device. Different input datasets may require different models to obtain appropriate output datasets. For example, the model used to obtain the output dataset may be a multi-objective optimization model. For example, the model includes one or more of the following: linear regression model, partial least squares (PLS) model, Gaussian mixture model, random forest model, neural network model (e.g., physical information neural network model, etc.). Also as described below, the determination of the quality score can be independent of the model selected to obtain the output dataset; that is, according to the embodiment, the determination of the quality score may be independent of the model applied to obtain the output dataset.

[0032] Simultaneously or subsequently, the computing device determines an associated quality score for each of the multiple results. In an embodiment, each quality score indicates the reliability of the associated result; for example, based on the prediction interval, the computing device can determine uncertainty. For example, the larger the interval, the greater the uncertainty associated with the result. That is, in an embodiment, determining the associated quality score includes uncertainty estimation. For example, a prediction interval estimation step is performed after the model has provided the output dataset (e.g., Pareto points). For example, based on each corresponding Pareto point, a prediction interval is calculated such that the final Pareto frontier has confidence limits. These confidence limits can be automatically evaluated and / or evaluated by the user.

[0033] For example, the quality score is determined based on at least one of a first uncertainty value associated with the input dataset and a second uncertainty value associated with the model. For this purpose, the input dataset may additionally include experimental data regarding the measured or simulated relationship between (a) at least one of a first and second input variables and (b) at least one of a first and second product attribute. For example, the first uncertainty value associated with the input data can be determined based on the experimental data.

[0034] In one embodiment, for each result, the quality score indicates the probability that the result is achieved within a deterministic range, where the deterministic range can be defined as a portion of the input dataset. For example, the input dataset includes, for instance, a user-defined deterministic range, such as defining a range of 40% to 95%, meaning that the result must be achieved within this range. In this case, the corresponding quality score indicates—for example, based on new observations / experiments performed through the step of determining the quality score—the probability that the corresponding result is indeed within the specified range. In another embodiment, for each result, the quality score indicates—based on experimental data—how much of the input variable's determined value (e.g., expressed as a percentage) actually produced a determined result value for the product attribute. That is, while the determination of the quality score can take into account both the input dataset and the output dataset (including the results), this determination can be performed independently of the model applied to obtain the output dataset. Therefore, the determination of the quality score can be done in a model-independent manner.

[0035] Generally, uncertainty is a commonly used term in statistical modeling. For example, two types of uncertainty may arise: random uncertainty and cognitive uncertainty. Random uncertainty (originating from data uncertainty) stems from randomness and is inherent in any system. It is uncontrollable and unpredictable, existing in all natural phenomena, such as weather, sea levels, and the stock market. Cognitive uncertainty (originating from model uncertainty) is caused by a lack of knowledge or information. It is more controllable and can be reduced through research and analysis.

[0036] The techniques described in this paper can address both types of uncertainty based on the provided input dataset. As explained above, providing the experimental data (e.g., additional information about the measurement methods) and / or providing a copy of the measurement allows for taking into account random uncertainties when determining the associated quality scores.

[0037] In an embodiment, determining the associated quality score includes applying a quantile regression method to the output dataset. For example, in an embodiment, determining the associated quality score does not include applying least squares. For example, least squares can estimate the conditional mean of the response variable (i.e., the first outcome value of the first product attribute and the second outcome value of the second product attribute) given the values ​​of the predictor variable (i.e., the third value of the first input variable and the fourth value of the second input variable), while quantile regression can estimate the conditional median (or other quantile) of the response variable.

[0038] In this embodiment, the applied quantile regression method is independent of the model used to obtain the output dataset. For example, the applied quantile regression method is model-agnostic. For instance, even if different models may have been used to obtain the output dataset that includes the results, the quantile regression method can still produce comparable quality scores.

[0039] For example, forecast intervals are used to provide an estimate of the uncertainty of a corresponding outcome among multiple outcomes (e.g., the probability that the corresponding outcome falls within a specified range, as described above). For instance, a forecast interval takes into account both the uncertainty of a point estimate and the surrounding data dispersion. A forecast interval can represent a range of values ​​that may contain the true values ​​of some response variables (product attributes) based on a single new observation of a specific value of one or more predictor variables.

[0040] For example, the prediction interval is determined based on the quantile regression method.

[0041] In the following, exemplary embodiments of methods including the application of quantile regression will be explained. Generally, the term "quantile regression" is used herein in the context that is commonly associated with those skilled in the art.

[0042] In the first step, an input dataset is received. In the following example, the input dataset includes the data associated with the first and second input variables, i.e., (i) data associated with the input variables. Furthermore, the input dataset includes (ii) the experimental data. Additionally, the input dataset includes (iii) quantile range data indicating quantile ranges (e.g., the deterministic range described above).

[0043] For example, the quantile range is predefined. In another embodiment, the quantile range is selected, for example, by a computing device and / or by a user device, for example, based on corresponding input from a user of the user device.

[0044] For example, quantile range data specifies one or more quantiles, such as two quantiles, for example, 5% and 95% or 10% and 80%. For example, the corresponding quantiles define the probability of the desired outcome. For example, quantile range data (e.g., user-defined) defines a range from 40% to 95%, meaning that each outcome must be achieved within a range of 40% to 95% certainty. In this case, the corresponding quality score indicates—for example, based on new observations / experiments performed through steps that determine the quality score—the probability that the corresponding outcome does indeed fall within the specified range.

[0045] In the next step, a quantile regression model is generated. For example, the generated quantile regression model models the relationship between input variables (e.g., independent predictors) and specified quantiles (also called percentiles) of the target attribute (e.g., dependent variable).

[0046] The quantile regression model can be generated according to the following equation:

[0047] ,in,

[0048] It is a response variable / target attribute The quantiles,

[0049] It's a quantile.

[0050] It is the regression parameter / the coefficient to be estimated, and

[0051] , …, It is the first Predicted / input variables for each observation.

[0052] Furthermore, generating a quantile regression model can include reducing the mean absolute deviation (MAD) of the quantile regression according to the following equation until the MAD is below a predefined value:

[0053] ,in,

[0054] It is a function based on the following equation:

[0055] ,in,

[0056] It is the error of a single data point, and

[0057] max() returns the maximum value specified in parentheses, for example, u,0 or -u,0.

[0058] After generating the quantile regression model, the quantile range data and data related to the input variables are fed into the quantile regression model.

[0059] Then, a quality score is obtained. As explained above, each result can be associated with at least one quality score.

[0060] In an embodiment, the input dataset may additionally include constraint data indicating one or more unacceptable values ​​associated with at least one of the first and second input variables. For example, a constraint may define that the total share of chemicals in the formulation should be 1 (a). In an embodiment, the output dataset does not present any results whose values ​​correspond to unacceptable values ​​according to the constraints(s).

[0061] The method further includes a selection step, for example, performed by a computing device. Specifically, one or more results from a plurality of results are selected based on associated quality scores. Based on the selected results(s), a control dataset for controlling product production is obtained.

[0062] The selection of one or more results can be done entirely automatically, or it can be done by taking into account additional input from the user.

[0063] In embodiments of the method, the method further includes being prompted by a computing device to present a representation of multiple results and associated quality scores. For example, the computing device itself includes a display for displaying the multiple results and associated quality scores.

[0064] Alternatively, the presentation of multiple results and associated quality scores may be provided by the user's device. For example, the selection may be influenced by user input provided by the user when faced with a presentation of multiple results and associated quality scores. The presentation may include at least one of visual presentation (e.g., including error bars for visualizing quality scores) and / or auditory or tactile presentation.

[0065] In a further embodiment, after obtaining the quality score, the computing device can determine that the associated quality score, for which no results are found, falls within a required range. The computing device can then automatically select another model to obtain the output dataset. Based on this other model, the steps of obtaining the output dataset and determining the associated quality score can be repeated. For example, this iteration can be repeated until the determined quality score falls within the required range. In another embodiment, the model used to obtain the output dataset remains unchanged. For example, if the model cannot achieve the desired result within a specified range of determinism, then assuming the applied model accurately represents reality, the result may not be achievable in reality. On the other hand, if new observations contradict the model, the model can be updated or replaced with a more suitable model.

[0066] For example, using quantile regression to determine quality scores allows for a reliable and accurate evaluation of results, which can ultimately improve formulation performance and save resources that would otherwise be used for experiments.

[0067] Based on the selection of one or more results according to associated quality scores, the product manufacturing process can be implemented such that the product exhibits a first target attribute value and a second target attribute value with defined determinism. Therefore, the technique for obtaining a control dataset for controlling product manufacturing, as described above, can automatically reflect, for example, a customer's request that the product must exhibit attribute values ​​with predefined probabilities. According to one or more embodiments, based on associated quality scores, only those results that meet the customer's request are selected. Therefore, the technique for obtaining the control dataset described above allows for efficient and automatic adjustment of product manufacturing control according to the customer's risk tolerance.

[0068] In this embodiment, data related to... can refer to data associated with... Specifically, data related to and / or associated with... may include data representing... and / or data including...

[0069] In embodiments, one or more first and / or second dependencies may be represented by one or more mathematical equations and / or may include one or more mathematical equations. One or more first and / or second dependencies may describe the relationship between a first input variable and a first product attribute and a second product attribute, and / or between a second input variable and the first product attribute and the second product attribute. Different first and / or second dependencies may include different relationships between a first input variable and the first product attribute and the second product attribute, and / or between a second input variable and the first product attribute and the second product attribute. Specifically, different first and / or second dependencies may include different mathematical equations relating to the first input variable and the first product attribute and the second product attribute, and / or between the second input variable and the first product attribute and the second product attribute.

[0070] In embodiments, the model can be configured to provide an output dataset including multiple results in response to receiving an input dataset. Data associated with a first input variable and / or a second input variable may include one or more first suggested values ​​associated with the first input variable and / or one or more second suggested values ​​associated with the second input variable. Further, data associated with the first input variable and / or the second input variable may include one or more first suggested attribute values ​​and / or one or more second suggested attribute values. One or more first suggested input values ​​and / or one or more second suggested input values ​​may be associated with one or more first and / or second suggested attribute values ​​of a first and / or second product attribute, particularly via one or more first and / or second dependencies. One or more first and / or second suggested input values ​​and / or one or more suggested attribute values ​​may be associated with each other, i.e., with a product, preferably with a composition of the product. One or more first and / or second result values, as well as third and fourth values, may be associated with a target product (i.e., a product other than the product associated with one or more first and / or second suggested input values). The model can be configured to obtain first and second result values, as well as third and fourth values, based on the data distribution indicated by the input dataset, particularly the data associated with the first and second input variables. Alternatively, the model can be configured to obtain first and second result values ​​that are closer to the first and second target attribute values ​​than one or more first and / or second suggested attribute values.

[0071] In embodiments, the association of the first, second, third, and fourth values ​​with each other may include the first, second, third, and fourth values ​​being associated with a product, particularly a target product, and / or a composition of the product. Values ​​associated with each other may be associated with the same product and / or the same composition of the product.

[0072] In embodiments, obtaining, and particularly determining, a quality score may include obtaining the quality score from a quality and / or reliability model. The quality and / or reliability model may be configured to provide quality scores associated with multiple results in response to receiving multiple results. The quality and / or reliability model may be obtained based on at least a portion of the input dataset, particularly based on data associated with a first input variable and / or a second input variable.

[0073] In embodiments, obtaining a control dataset may include extracting at least a portion of the selected results into the control dataset. Providing control data may include providing at least a portion of the selected results. Selecting one or more results from a plurality of results may include obtaining, in particular receiving, a selection of one or more results from a plurality of results via a user interface. In response to providing, in particular displaying, multiple results and(s) of associated quality scores via a user interface, a selection of one or more results from a plurality of results may be obtained, in particular received. Displaying multiple results and(s) of associated quality scores may include displaying at least a portion of these results and displaying(s) of quality scores associated with at least a portion of these results.

[0074] In an embodiment, obtaining a quality score may include: obtaining a predefined quantile range and / or receiving quantile range data indicating the quantile range, for example, by specifying two or more quantiles (e.g., 5% and 95% or a narrower band where applicable); and providing at least a portion of the output dataset (particularly the first and second result values) to a quality and / or reliability model, and receiving a quality score associated with these results from the quality and / or reliability model.

[0075] This document also presents a computing device. This computing device is configured to perform methods according to one or more embodiments described above. Therefore, regarding embodiments of the computing device, refer to the foregoing.

[0076] This article also presents a method for producing a product, wherein the production of the product is performed based on a control dataset provided by a method according to one or more embodiments of the above embodiments.

[0077] This paper further presents a system for producing products, wherein the system is configured to execute the production method based on a control dataset.

[0078] Those skilled in the art will recognize the additional features and advantages after reading the following detailed description and reviewing the accompanying drawings. Attached Figure Description

[0079] The parts in the accompanying drawings are not necessarily drawn to scale; rather, the focus is on illustrating the principles of the invention. Furthermore, in the drawings, the same reference numerals denote corresponding parts / steps. In the drawings:

[0080] Figures 1A to 1E This illustration schematically shows a portion of an exemplary method for providing a control dataset for controlling product production, according to one or more embodiments;

[0081] Figure 2This illustration schematically shows a portion of an exemplary method for providing a control dataset for controlling product production, according to one or more embodiments;

[0082] Figure 3 The Pareto front with confidence bounds and the Pareto front without confidence bounds are illustrated schematically and exemplary.

[0083] Figure 4 This illustration schematically shows a portion of an exemplary method for providing a control dataset for controlling product production, according to one or more embodiments;

[0084] Figure 5 A computing device and a user device according to one or more embodiments are illustrated schematically and exemplaryly.

[0085] Figure 6 A system for producing a product according to one or more embodiments is illustrated schematically and exemplaryly; and

[0086] Figure 7 This illustration schematically shows a portion of an exemplary method for determining a quality score according to one or more embodiments. Detailed Implementation

[0087] Figures 1A to 1E The illustration schematically shows a portion of an exemplary method for providing a control dataset for controlling product production according to one or more embodiments, namely, the steps of receiving an input dataset by a computing device and providing an output dataset by the computing device.

[0088] In this example, the input dataset is provided by the user using their device. Figure 5 The diagram schematically illustrates both computing device 2 and user device 8. Both devices 2 and 8 include their respective memories 21 / 81 (stored instructions from their respective processors 22 / 82) and their respective transceivers 23 / 83 (for receiving and sending data). The exact configuration of these devices 2 and 8 is not particularly important at this time. For example, user device 8 could be a terminal device such as a personal computer, smartphone, tablet, laptop, etc. Computing device 2 could be, for example, a server or server farm or part thereof. The processor 22 of computing device 2 could include spatially distributed storage and processing devices operatively interconnected to execute instructions stored in memory 21.

[0089] refer to Figures 1A to 1B The graphical user interface (“Qritos”) provided by user device 2 allows users to provide input datasets. For example, after creating a new project on computing device 2, a user can add data, for instance, by uploading a file containing the input dataset.

[0090] according to Figures 1A to 1E The non-limiting example shown is intended to obtain a control dataset for controlling cake production. The process is the same for any other product.

[0091] The input dataset includes a first target attribute value related to a first product attribute of the product and a second target attribute value related to a second product attribute of the product. The input dataset further includes data related to a first input variable and a second input variable, wherein a first dependency between the first input variable and the first and second product attributes differs from a second dependency between the second input variable and the first and second product attributes. Refer to the cake example; see [link to cake example]. Figure 1C The input variables are "wheat flour," "spelt flour," "sugar," "chocolate," "nuts," and "carrots." For these input variables, the user inputs data, i.e., the corresponding allowed value ranges, see, for example, for "sugar," the allowed value range is from 0.15 to 0.35, for example, indicating the allowed weight share of sugar in a cake. In the example, the user further defines three target attributes, namely, target values ​​related to "calories," "texture," and "caramelization." The input variables "wheat flour," "spelt flour," "sugar," "chocolate," "nuts," and "carrots" have different effects on the target attributes "calories," "texture," and "caramelization." In other words, the dependencies of the input variables "wheat flour," "spelt flour," "sugar," "chocolate," "nuts," and "carrots" on each of the target attributes "calories," "texture," and "caramelization" are different from each other.

[0092] refer to Figure 1D The input dataset in this example further includes constraint data indicating one or more unacceptable values ​​associated with at least one of the first and second input variables. For example, refer to... Figure 1D The larger circle displayed allows users to indicate mixing constraints associated with the six ingredients (i.e., the input variables "wheat flour", "spelt flour", "sugar", "chocolate", "nuts", and "carrots").

[0093] After receiving the input dataset, computing device 2 feeds the input dataset into the model to obtain the output dataset from the model. The output dataset includes multiple results, where each result indicates the result value of the product attribute ("calories", "taste", and "caramelization") and the associated value of the input variables ("wheat flour", "spelt flour", "sugar", "chocolate", "nuts", and "carrots").

[0094] For example, the output dataset includes at least a portion of the Pareto front (also known as the Pareto boundary or Pareto curve) already obtained based on the input dataset and the model. The Pareto front may include at least a portion of all Pareto-efficient solutions. (Reference) Figure 1E The results are presented as a range of values ​​at user device 8, where, Figure 1E The lower part displays the values ​​of the input variables (referred to here as "inputs": "wheat flour," "spelt flour," "sugar," "chocolate," "nuts," and "carrots"), while the result values ​​for the product attributes ("calories," "texture," and "caramelization") are displayed in the lower part. Figure 1E The upper part is presented (here referred to as the "target"). Each result (i.e., a combination of three values ​​of the target attribute associated with "calories," "texture," and "caramelization") is associated with a set of corresponding values ​​for the input variables "wheat flour," "spelt flour," "sugar," "chocolate," "nuts," and "carrots." The user or computing device 2 (or computing device 2 itself) can select one of the results deemed most suitable. After selection, the corresponding dataset for the selected result can be saved.

[0095] Based on the selected results, computing device 2 can obtain a control dataset for controlling the production of the product (in this example: cake).

[0096] Figure 2 This demonstrates another method for providing control datasets for controlling product production. Here, it can be seen that... Figures 1A to 1D The input data is received in an exemplary or alternative manner (step 10). After receiving the input data, it is fed into the model (step 12) to obtain (step 13) the output dataset. Prior to step 12, the model may generate and / or select from multiple available models based on the input dataset. Different input datasets may require different models to obtain appropriate output datasets. For example, the model used to obtain the output dataset may be a multi-objective optimization model, such as a linear regression model, a partial least squares (PLS) model, a Gaussian mixture model, a random forest model, a neural network model (e.g., a physical information neural network model, etc.).

[0097] One of the problems with using multi-objective optimization models is the presence of multiple outcomes. Sometimes, several similar outcomes may be presented, making it difficult for both the computing device and the user to decide which outcome to select, and thus, based on this, to generate a control dataset for controlling product production.

[0098] Figure 2 The embodiment shown addresses this problem. Here, simultaneously with or after obtaining the output dataset (step 13), computing device 2 determines an associated quality score for each of the multiple results in step 14.

[0099] In one embodiment, for each result, a quality score indicates the probability that the result is achieved within a range of determinism, where the range of determinism can be defined as a portion of the input dataset. For example, the input dataset includes, for instance, a user-defined range of determinism, such as defining a range of 40% to 95%, meaning that the result must be achieved within this range of 40% to 95%. In this case, the corresponding quality score indicates—for example, based on new observations / experiments performed through the step of determining the quality score—the probability that the corresponding result is indeed within the specified range. In another embodiment, each quality score indicates the reliability of the associated result, for example, uncertainty can be determined based on the prediction interval. For example, the larger the interval, the greater the uncertainty associated with the result. For example, determining the associated quality score includes uncertainty estimation. For example, a prediction interval estimation step is performed after the model has provided the output dataset (e.g., Pareto points). For example, based on each corresponding Pareto point, a prediction interval is calculated such that the final Pareto front has confidence limits. These confidence limits can be automatically evaluated by computing device 2 and / or evaluated by the user.

[0100] Based on the additional quality score, transparent selection criteria can be established. For example, it can be defined as only selecting results whose corresponding associated quality score is above a threshold level.

[0101] For example, the quality score is determined based on at least one of a first uncertainty value associated with the input dataset and a second uncertainty value associated with the model. For this purpose, the input dataset may additionally include experimental data regarding the measured or simulated relationship between (a) at least one of a first and second input variables and (b) at least one of a first and second product attribute. For example, the first uncertainty value associated with the input data can be determined based on the experimental data.

[0102] For example, refer to Figures 1A to 1E For example, input data may include one or more experimental results that have been achieved in previous experiments, such as a specific combination of the six ingredients producing a specific set of values ​​for the three properties ("calories", "taste" and "caramelization") under specific conditions.

[0103] exist Figures 2 to 4 In the examples, determining the associated quality score may include, for instance, applying a quantile regression method to the output and input datasets. For example, in one embodiment, determining the associated quality score does not involve applying least squares.

[0104] In this embodiment, the associated quality score is determined independently of the model used to obtain the output dataset. For example, the associated quality score is determined to be model-independent. For example, even if different models may have been used to obtain the output dataset, comparable quality scores can still be produced after the step of determining the associated quality score.

[0105] Figure 7 A portion of an exemplary method for determining (step 14) a quality score according to one or more embodiments is illustrated schematically. In this embodiment, the quantile regression method is applied to determine the quality score.

[0106] In step 140, an input dataset is received (e.g., as part of step 10 above). In this example, the input dataset includes the data associated with the first and second input variables, i.e., (i) data associated with the input variables. Furthermore, the input dataset includes (ii) the experimental data. Additionally, the input dataset includes (iii) quantile range data indicating quantile ranges (e.g., the deterministic range described above).

[0107] For example, the quantile range is predefined. In another embodiment, the quantile range is selected, for example, by computing device 2 and / or by user device 8, for example, based on corresponding input from a user of user device 8.

[0108] For example, quantile range data specifies one or more quantiles, such as two quantiles, for example, 5% and 95% or 10% and 80%. For example, the corresponding quantiles define the probability of the desired outcome. For example, quantile range data (e.g., user-defined) defines a range from 40% to 95%, meaning that each outcome must be achieved within a range of 40% to 95% certainty. In this case, the corresponding quality score indicates—for example, based on new observations / experiments performed through steps that determine the quality score—the probability that the corresponding outcome does indeed fall within the specified range.

[0109] In step 142, a quantile regression model is generated. For example, the generated quantile regression model models the relationship between input variables (e.g., independent predictors) and specified quantiles (also called percentiles) of the target attribute (e.g., dependent variable).

[0110] In step 142, the quantile regression model can be generated according to the following equation:

[0111] ,in,

[0112] It is a response variable / target attribute The quantiles,

[0113] It's a quantile.

[0114] It is the regression parameter / the coefficient to be estimated, and

[0115] , …, It is the first Predicted / input variables for each observation.

[0116] Furthermore, generating a quantile regression model can include reducing the mean absolute deviation (MAD) of the quantile regression according to the following equation until the MAD is below a predefined value:

[0117] ,in,

[0118] It is a function based on the following equation:

[0119] ,in,

[0120] It is the error of a single data point, and

[0121] max() returns the maximum value specified in parentheses, for example, u,0 or -u,0.

[0122] After generating the quantile regression model in step 142, the quantile range data and data related to the input variables are fed into the quantile regression model. See [link to relevant documentation]. Figure 7 Step 144 in the process.

[0123] In this embodiment, the results are also fed into the quantile regression model.

[0124] In step 146, a quality score is obtained. As explained above, each result can be associated with at least one quality score.

[0125] In a further embodiment, more than one method is applied to determine the quality score. For example, in addition to applying a quantile regression method, a Gaussian process is also applied to determine the quantile regression method. For example, according to an embodiment, different quality scores obtained by applying different methods are considered to determine the final quality score for each result.

[0126] Return to Figure 2 The model is applied to the input dataset to obtain an output dataset including the results (step 13), as exemplarily described above. Further, for each of these results, an associated quality score is determined (step 14).

[0127] Figure 2The method further includes, prompted by the computing device 2 (step 15), presenting a representation of multiple results and associated quality scores. This presentation may include visual presentation, such as... Figure 1E The example shown is as follows. However, in addition to presenting the results, the associated quality scores are also represented, for example, based on error bars or confidence limits. For example, in Figure 3 In the diagram, the left figure shows the complete Pareto front without confidence bounds for two input variables, y1 and y2, while the right figure shows the same Pareto front with confidence bounds (dashed lines above and below the Pareto front). The greater the distance between the confidence bounds, or the larger the corresponding prediction interval, the lower the quality of the corresponding Pareto point.

[0128] As explained above, the selection of results is based on quality scores. This selection can be performed automatically by computing device 2 or by the user. In the latter case, the results and associated quality scores are preferably presented to the user graphically, for example, via user device 8. Based on the graphical representation, navigation services can be provided to the user, allowing them to scroll through the results and associated quality scores.

[0129] Figure 4 A further embodiment of method 1 for providing a control dataset for controlling product production is shown. This method can be summarized as follows:

[0130] Step 10 includes receiving an input dataset by computing device 2, wherein the input dataset can be configured as exemplarily described above.

[0131] Step 12 includes feeding an input dataset into a model by computing device 2 to obtain an output dataset 13 from the model, the output dataset including multiple results, wherein these results can be configured as exemplarily described above.

[0132] Step 14 includes determining an associated quality score for each of the multiple results by computing device 2. This determination can be performed as described above. For example, determining the associated quality score can be based on the input dataset. Further, determining the associated quality score may include performing at least one of (i) a linear regression method, (ii) a quantile regression method, and (iii) a prediction interval generation model.

[0133] In optional step 17, the quality score is checked for acceptability, for example, by performing a threshold test. If the quality score is determined to be unacceptable, the method may include step 11, whereby computing device 2 selects another model from a plurality of available models. Alternatively, a model deemed suitable may be generated. Steps 12 and 14 are then repeated until the quality score is determined to be acceptable. In another embodiment, the model used to obtain the output dataset (step 12) remains unchanged. For example, if the model fails to achieve the desired result within a specified range of determinism, then assuming the applied model accurately represents reality, the result may not be achievable in reality. On the other hand, if new observations contradict the model, the model may be updated or replaced with a more suitable model.

[0134] Step 18 involves the computing device selecting one or more results from a plurality of results based on associated quality scores. As mentioned above, this selection step can be performed automatically by the computing device and / or based on user input.

[0135] Finally, step 19 includes the computing device 2 obtaining the control dataset described in step 19 from one or more selected results. The obtained control dataset is configured to control the production of the product. For example, the obtained control dataset includes process parameters related to the production process and / or indications of the amount or other characteristics of the segregants used to produce the product.

[0136] Figure 6 A system 3 for producing product P according to one or more embodiments is illustrated schematically and exemplary. Of course, the details of system 3 depend on the product P to be produced and will not be elaborated here.

[0137] In this embodiment, system 3 is a chemical production plant.

[0138] Figure 6 System 3 is demonstrated to produce product P based on N quantities of segregates ED_1 to ED_N. Production is governed by controlled system parameters SP, which define, for example, the processing of segregates ED_1 to ED_N by one or more system processing units based on System 3. Figure 7 In the illustrated embodiment, system 3 is also controlled based on a control dataset CDS obtained according to the above embodiments. For example, the control dataset CDS defines at least one processing parameter for at least one system processing unit in system 3, which is used to process at least one of N segregants ED_1 to ED_N.

Claims

1. A method (1) for providing a control dataset for monitoring and / or controlling the production and / or processing of a product, the method (1) comprising: - The computing device (2) receives (10) an input dataset, which includes: ○ The first target attribute value related to the first product attribute of the product and the second target attribute value related to the second product attribute of the product; ○ Data associated with a first input variable and a second input variable, wherein one or more first dependencies between the first input variable and the first product attribute and the second product attribute are different from one or more second dependencies between the second input variable and the first product attribute and the second product attribute; and - The input dataset is fed (12) into the model by the computing device (2) to obtain (13) an output dataset from the model, which includes multiple results, each indicating: ○ The first result value of the first product attribute, the second result value of the second product attribute; and ○ The third value of the first input variable and the fourth value of the second input variable; and -The computing device (2) obtains, and in particular determines (14) the associated quality score for each of the plurality of results; - The computing device (2) selects (18) one or more of the multiple results based on these associated quality scores; and -The control dataset (19) is obtained by the computing device (2) from one or more selected results. Optionally, the control dataset may be provided, particularly for monitoring and / or controlling the production and / or processing of the product.

2. The method (1) as described in claim 1, wherein, The steps to determine (14) these associated quality scores include applying quantile regression methods.

3. The method (1) as claimed in claim 1 or 2, wherein, (14) These quality scores are determined independently of the model used to obtain the output dataset, and in particular independently of the model type associated with the model used to obtain the output dataset.

4. The method (1) as described in any of the preceding claims, wherein, (14) These quality scores are determined based on at least one of a first uncertainty value that is related to and / or associated with the input dataset and a second uncertainty value that is related to and / or associated with the model.

5. The method (1) as described in any of the preceding claims, wherein, The quality score indicates the reliability of the corresponding result, and / or the quality score includes at least one prediction interval.

6. The method (1) as described in any of the preceding claims, wherein, The input dataset further includes constraint data indicating one or more unacceptable values ​​related to and / or associated with at least one of the first and second input variables.

7. The method (1) as described in any of the preceding claims, wherein, The input dataset also includes experimental data on the measured or simulated relationship between (a) at least one of the first and second input variables and (b) at least one of the first and second product attributes.

8. The method (1) as described in any of the preceding claims, wherein, The method further includes: -The computing device (2) selects (11) the model from a number of available models.

9. The method (1) of claim 8, wherein, The selection of the model (11) was performed based on these quality scores.

10. The method (1) as described in any of the preceding claims, wherein, The model is hosted by the computing device (2).

11. The method (1) as described in any of the preceding claims, wherein, This model is a multi-objective optimization model.

12. The method (1) as described in any of the preceding claims, further comprising: - The computing device (2) prompts (15) to present the multiple results and the representation of these associated quality scores; And optionally, wherein: - The presentation includes at least one of visual presentation—for example, including error bars—and auditory presentation; and / or -The presentation is provided by the user equipment (8).

13. A computing device (2), wherein, The computing device (2) is configured to perform the method (1) according to any one of the preceding claims.

14. A method for producing a product, wherein, The production of the product is performed based on the control dataset provided by the method (1) according to any one of the preceding claims 1 to 12.

15. A system (3) for producing a product, wherein, The system (4) is configured to perform the method (3) as described in the preceding claim.

Images