Inverse determination of fuel compositions based on explainable artificial intelligence

Abstract

Computer-implemented method for generating a trained combustion property determination machine learning model (24) for determining at least one combustion parameter (18) of a given fuel using analysis data sets which are each characteristic of a composition of a given fuel, comprising the steps - Providing a trainable combustion property determination machine learning model (22) which includes a set of trainable parameters and which, based on an analysis data set acquired for a given fuel as input (2), generates at least one combustion parameter (18) as output (6); - Generating a training dataset comprising a variety of acquired analytical datasets on training fuels, which are used to generate the training dataset, as well as a variety of combustion parameters (18); - Training the combustion property determination machine learning model (22) based on the training data set; - Determine at least one influencing factor (28) by applying a method to explain the trained combustion property determination machine learning model (24).

Description

[0001] The present invention relates to a method and a device for generating a trained combustion property determination model of machine learning for determining at least one combustion parameter. Furthermore, the invention relates to a method and a device for the particularly predictive determination of a particularly novel fuel composition using the trained combustion property determination model.

[0002] The present invention relates to a method for predicting fuel properties and in particular for predicting properties of a fuel mixture during combustion in an engine and / or a resulting exhaust gas composition.

[0003] It is generally known from the prior art to examine fuel compositions on engine test benches and thereby determine the combustion properties of the fuels. Such combustion properties include, for example, exhaust gas composition, pollutant load in the exhaust gases (e.g., nitrogen oxides), in particular particulate matter or soot particle load in the exhaust gases, efficiency, combustion duration, and the like. Such a manual examination on the test bench must be carried out anew for each new fuel composition.

[0004] In recent years, methods have become increasingly available that allow the calculation or prediction of fuel properties based on a given fuel composition. Typically, artificial intelligence is used for this purpose. A machine learning model (e.g., a neural network) can be used to predict fuel properties, trained with analytical datasets relating to individual fuel components or the resulting mixtures.

[0005] Such a method is known, for example, from DE 10 2022 207 017 A1 of the applicant. By applying the method described therein, it is possible, for example, to predict a physicochemical property of a fuel mixture, e.g., its vapor pressure or ignitability. However, this method does not allow for predictions about the exhaust gas composition after combustion in an engine.

[0006] A comprehensive simulation of exhaust gas composition is not satisfactorily possible with currently available methods such as CFD (Computational Fluid Dynamics) simulation methods. Calculating the entire spectrum of exhaust emissions would be far too complex, which is why, as a rule, calculations are only performed in part.

[0007] It is known from the state of the art to use complex AI models, such as neural networks, for prediction. However, these complex AI models are so-called black-box models, which are not easy to interpret. In other words, it is not possible for a human to understand how the AI ​​model's predictions are composed in detail and how reliable these statements are. In the state of the art, this problem is circumvented by resorting to approximation solutions or by greatly simplifying the models used, which entails corresponding inaccuracies in the predictions.

[0008] A method for managing vehicle fleets is known from US Patent 2023 / 0260342 A1. The method comprises obtaining data relating to the condition and operation of each of a large number of vehicles of a specific type in a vehicle fleet; detecting an anomalous average fuel consumption over a specific period in at least one initial vehicle of the fleet based on the analysis of the obtained data; determining and considering the cause of the detected anomalous average fuel consumption by implementing an explainable artificial intelligence algorithm that considers various parameters of the initial vehicle, including parameters relating to the initial vehicle's driving behavior over a specific period, the initial vehicle's condition, and a range of meteorological and environmental elements, and deriving the influence of each of the parameters.

[0009] From KR 2023 0165 495 A a method for selecting key features for the construction of a model for predicting nitrogen oxide emissions of a diesel engine and a model for predicting nitrogen oxide emissions of a diesel engine, which is constructed with the selected key features, is known.A procedure for selecting key features for constructing a nitrogen oxide emission prediction model of a diesel engine comprises the following steps: constructing a base model with all features influencing nitrogen oxide emissions as inputs and nitrogen oxide emissions as outputs; and analyzing the importance of all features; selecting some features from the total features based on the importance of the total features and constructing at least one comparison model with the selected features; evaluating the accuracy of the at least one comparison model relative to the reference model; and determining a key feature based on the results of the accuracy evaluation.

[0010] For the future use of renewable or synthetic fuels, there is a need for a method to predict the combustion properties of new fuels, and in particular, to predict the exhaust gas composition during their combustion, without having to measure each individual fuel mixture on a test bench. Furthermore, there is a need for a method to determine new fuel mixtures that reduce emissions or increase efficiency. Finally, it is necessary for those skilled in the art to be able to understand the result of such a prediction.

[0011] The present invention is therefore based on the objective of overcoming the disadvantages known from the prior art and providing a method and a device for generating a trained combustion property determination model of machine learning for determining at least one combustion parameter and a method and a device for the particularly predictive determination of a new fuel composition using the trained combustion property determination model, which offer a possibility to reliably and comprehensibly predict combustion properties of fuels and to determine new fuel compositions on this basis.

[0012] The object of the invention is achieved by the subject matter of the independent claims. Advantageous embodiments and further developments of the invention are the subject matter of the dependent claims.

[0013] In an inventive, in particular computer-implemented, method for generating a trained combustion property determination model of machine learning for determining at least one combustion parameter of a given fuel mixture using analysis data sets, each of which is characteristic for a composition of a given fuel mixture, wherein each analysis data set for the respective given fuel mixture comprises at least one analysis parameter and a proportion parameter assigned to the analysis parameter, wherein the analysis parameter is characteristic for a group of compounds from the given fuel mixture which have at least one given functionality, and wherein the proportion parameter is characteristic for a quantity fraction of the compounds in the respective fuel mixture which are grouped in the group characterized by the analysis parameter,In one step, it comprises providing a trainable combustion property determination model of machine learning, which includes a set of trainable parameters and which, based on an analysis data set acquired for a given fuel mixture or data derived therefrom as input, generates at least one combustion parameter as output, wherein the combustion parameter is characteristic of at least one combustion property of the fuel mixture.

[0014] In a further step of the inventive method, a training dataset is generated, comprising a plurality of acquired and / or determined analysis datasets for training fuel mixtures, which are used to generate the training dataset, as well as a plurality of combustion parameters, wherein the combustion parameters are each characteristic of at least one combustion property of the respective training fuel mixture. In a further step of the inventive method, the combustion property determination machine learning model is trained on the basis of the training dataset.

[0015] In a further step of the inventive method, at least one influencing factor is determined by applying a method for explaining the trained combustion property determination model of machine learning, wherein the at least one influencing factor is characteristic of an influence of the at least one proportion factor and / or the at least one analysis factor on the combustion factor.

[0016] Preferably, the specified fuel mixture is a fuel mixture of at least two different fuels. Preferably, the fuel mixture comprises a plurality of fuels, wherein preferably more than two, preferably more than three, preferably more than five, preferably more than ten, preferably more than twenty, and particularly preferably more than fifty (especially pairwise different) fuels are mixed together in the fuel mixture.

[0017] Preferably, the fuel mixture contains the fuels in a predetermined and / or known and / or measured mixing ratio. Preferably, the mixing ratio is characteristic of the, in particular, volumetric and / or mass fractions and / or quantity fractions of the mixed fuels relative to each other.

[0018] Preferably, the fuel mixture can be used as fuel for an internal combustion engine, particularly in a vehicle. Preferably, the fuel mixture is gasoline. It would also be conceivable to apply the present invention to diesel fuel with appropriate modifications.

[0019] As mentioned above, in a preferred method a machine learning model (trainable combustion property determination machine learning model) is provided and trained using training fuel mixtures or data sets relating to these training fuel mixtures.

[0020] Preferably, the trainable combustion property determination machine learning model is based on an (artificial) neural network. Preferably, the neural network is selected from a group of neural networks that includes Recurrent Neural Networks (RNNs), Long Short-Term Memory Networks (LSTMs), Gated Recurrent Units (GRUs), Temporal Convolutional Networks (TCNs), Transformer Networks, Convolutional Neural Networks (CNNs), Autoencoders, Time Delay Neural Networks (TDNNs), Echo State Networks (ESNs), Graph Neural Networks (GNNs), or the like.

[0021] Preferably, the neural network is a Recurrent Neural Network (RNN) and / or a Long Short-Term Memory Network (LSTM) and / or a Temporal Convolutional Network (TCN) and / or an Echo State Network (ESN) and / or a Graph Neural Network (GNN).

[0022] Preferably, the trained model is a machine learning model, particularly a trainable one, which includes a set of parameters, particularly trainable ones, that are set to values ​​learned as a result of a training process.

[0023] Preferably, the training process or training method is selected from a group of training methods that includes supervised learning, unsupervised learning, semi-supervised learning, reinforcement learning, transfer learning, ensemble learning, cross-validation, Bayesian methods, online learning, anomaly detection, and the like. The training method is particularly preferred to be supervised learning and / or reinforcement learning and / or cross-validation.

[0024] In a preferred method, a training dataset is generated which contains a multitude of recorded and / or determined analysis datasets of training fuel mixtures, which are used to generate the training dataset. In a preferred method, such an analysis dataset is characteristic of a composition and / or property of the respective training fuel mixture.

[0025] In a preferred method, an analysis data set is generated and / or determined for a training fuel mixture, and preferably for each training fuel mixture. In a preferred method, the analysis data set for each training fuel mixture comprises at least one analytical parameter and at least one proportion parameter assigned to that analytical parameter. Preferably, the analytical parameter is characteristic of a group of compounds from the training fuel mixture that exhibit at least one predetermined functionality. Preferably, the proportion parameter is characteristic of a quantity fraction of the compounds in the training fuel mixture that are grouped together in the group characterized by the analytical parameter.

[0026] Preferably, the analytical parameter and its associated proportion refer to a (predefined and / or predefined) component of the respective fuel mixture, for example, a predefined group of compounds in the respective fuel mixture. It would also be conceivable that the analytical parameter refers to a specific chemical compound and the associated proportion to the proportion of this specific chemical compound in the fuel mixture.

[0027] The analytical parameter specifies the type of (predefined and / or predefined) component; it thus identifies the type of component. In other words, the analytical parameter serves as an identifier for the component of the fuel mixture that is to be specifically examined.

[0028] The proportion is particularly characteristic of the proportion of the component defined, characterized, or identified by the analytical parameter in the respective fuel mixture. In other words, the proportion represents a measure of the proportion of the component identified or characterized by the analytical parameter (e.g., a predefined group of compounds) in the respective fuel. Preferably, the proportion is characteristic of a volumetric proportion (e.g., at a predefined and / or predefinable temperature, such as 15°C, and / or pressure, and / or other parameters). However, it is also conceivable that the proportion is characteristic of a quantitative proportion and / or a mass proportion.

[0029] Preferably, each analysis data set (for the respective fuel mixture) comprises a plurality of (especially pairwise) analytical parameters and corresponding proportions assigned to each analytical parameter. Preferably, each analytical parameter is assigned exactly one proportion. The analytical parameters are preferably those described above, which relate to distinct components (such as the specified group of compounds), particularly preferably in pairwise configurations.

[0030] The analytical parameter is characteristic of a group of compounds (preferably hydrocarbon compounds and / or oxygen-containing compounds) that exhibit at least one predefined functionality and / or a predefined carbon number (or several predefined carbon numbers). In other words, by specifying the analytical parameter, it is possible to identify or specify the, or preferably all, compounds with at least one predefined functionality (which can, in particular, commonly or generally occur in fuels). Preferably, the analytical data set of a fuel mixture includes information on the hydrocarbons or hydrocarbon groups and / or on oxygenates (or oxygen-containing compounds) in the fuel mixture.

[0031] For example, the analytical parameter could be characteristic of a group of hydrocarbons which includes (in particular, approximately all hydrocarbons occurring in a fuel) with aromatic functionality and hydrocarbons with olefin (cyclic) functionality.

[0032] The proportion is characteristic of a quantity fraction of the compounds in the fuel mixture that are grouped in the group characterized (or identified or identifiable) by the analytical parameter. The quantity fraction preferably refers to a ratio (in particular a volumetric ratio and / or a mass ratio) of the compounds in a given total quantity of the respective fuel mixture, grouped in the group characterized by the analytical parameter, to the total quantity of the respective fuel. In particular, the analytical data set (preferably through the values ​​of the proportions assigned to the respective analytical parameters) is characteristic of the composition of the respective fuel mixture.

[0033] The term "acquisition of an analysis data set" refers in particular to the receipt of the analysis data set transmitted via a data transmission device and / or the retrieval of the analysis data set from a storage device, in particular a non-volatile one (such as an external storage device like an external (backend) server), by a device, in particular a processor-based and preferably described in more detail below, for generating a trained combustion property determination model of machine learning for determining at least one combustion parameter, such that (by acquiring) an analysis data set is made available for data processing of the analysis data set.

[0034] In addition, according to another embodiment, "acquisition of an analysis data set" can also be understood to mean, additionally or alternatively, the determination and / or generation of the analysis data set. This could be achieved through computational determination, for example by interpolation and / or extrapolation of predefined data values ​​(such as at least one or more similar fuels), and / or by experimental generation or experimental determination and / or experimental measurement of the analysis data set.

[0035] In an advantageous method, at least one analytical data set, and preferably a plurality of analytical data sets, and particularly preferably each analytical data set, comprises at least one fuel parameter characteristic of at least one fuel property, wherein the at least one fuel property is preferably selected from a group of fuel properties including ignition quality, energy content, density, vapor pressure, boiling point, viscosity, knock resistance, cetane number (CN), characteristic values ​​such as points of a distillation curve, yield sooting index (YSI), research octane number (RON), motor octane number (MON), front octane number (FOZ), road octane number (SOZ), and the like, as well as combinations thereof.In particular, the Yield Sooting Index (YSI), especially the soot yield index, is characteristic of the amount of soot formed by a fuel when injected at low concentration into a methane-air base flame.

[0036] Preferably, the analytical data sets are obtained from a reformulyzer analysis and / or a spectroscopic analysis. In a further preferred method, the acquired and / or to-be-acquired analytical data sets are generated by means of a gas chromatographic (experimental) measurement method, in particular a multidimensional one, for determining the hydrocarbon groups and / or the oxygen-containing compounds of the respective fuel.

[0037] Preferably, the method for determining hydrocarbon groups and oxygen-containing compounds in gasoline and ethanol fuel (E85) as defined in DIN EN ISO 22854 (multidimensional gas chromatographic method ISO / DIS 22854:2020) is used. With regard to a disclosure of this measurement method according to DIN EN ISO 22854 (in particular for determining and / or recording at least one analytical data set and / or determining at least one fraction, and preferably all analytical data sets or fractions), reference is made to ISO / DIS 22854:2020, the content of which is hereby incorporated into this application.

[0038] Preferably, the determination of fuel properties or the determination of data relating to at least one fuel property should be based on the data determined in ISO 22854 (for the analysis data sets of the respective fuels). The special feature is the classification of the fuel components into functional groups and the subdivision of these according to the number of carbon atoms or the precise naming of the relevant oxygenates.

[0039] In a further preferred method, the analytical parameter is characteristic of a group of hydrocarbons. Preferably, at least one predefined functionality is selected (or the several predefined functionalities are selected) from a group comprising a paraffin (n- / -iso) functionality, a naphthene functionality, an olefin (n- / -iso) functionality, an olefin (cyclic) functionality, and / or an aromatic functionality. The group of hydrocarbons with several predefined functionalities is to be understood, in particular, as including all hydrocarbons that exhibit at least one of the several predefined functionalities.

[0040] In other words, the analytical data set preferably contains information (in particular their (especially volume) fraction in the fuel) of the hydrocarbons with respect to a functionality of the hydrocarbons, wherein the at least one specified functionality is preferably selected from a group comprising a paraffin(n- / -iso) functionality, a naphthene functionality; an olefin(n- / -iso) functionality, an olefin(cyclic) functionality and / or an aromatic functionality.

[0041] In a further preferred method, the analytical parameter is characteristic of a group of oxygenates, wherein the at least one functionality is an ether functionality and / or an alcohol functionality and / or wherein the analytical parameter is characteristic of a carbon chain structure and / or at least and preferably exactly one type of alcohol.

[0042] Preferably, (in particular) one analytical parameter is characteristic of a group of oxygenates with an ether functionality. Additionally or alternatively, preferably (in particular) one analytical parameter is characteristic of a group of oxygenates with an alcohol functionality.

[0043] In a further preferred method, at least one, and preferably the, analytical parameter (more preferably several analytical parameters) is characteristic of a group of compounds, in particular hydrocarbon compounds, which have a predetermined number of carbon atoms. In particular, two or more carbon atoms within a functional group are preferably grouped together (to form a common group of compounds for which the analytical parameter is characteristic), preferably consecutive ones, such as paraffin C3 + C4. This offers the advantage that the number of data records in the analytical data set decreases.

[0044] Preferably, each combination of functionality and carbon number (C number) (or several predefined carbon numbers) represents a separate analytical parameter or identifier; together, they preferably form the analytical data set with their respective proportions (or proportion sizes). Preferably, each identifier is assigned its proportion in the mixture (in particular via the proportion size).

[0045] Preferably, each recorded analytical data set contains at most twenty, preferably at most 15, preferably at most 13 (pairwise distinct) analytical parameters for oxygenates. Preferably, each recorded analytical data set contains at most 15, preferably at most ten, preferably at most nine different analytical parameters, and particularly preferably exactly one analytical parameter for alcohols. Preferably, each recorded analytical data set contains at most five, preferably at most four (pairwise distinct) analytical parameters, and particularly preferably exactly one analytical parameter for ethers.

[0046] Preferably, each recorded analytical data set contains at most 15, preferably at most 9, and most preferably at most 6 different analytical parameters for paraffins (n- / -iso). Preferably, each recorded analytical data set contains at most 6, and more preferably exactly one, analytical parameter(s) for naphthenes. Preferably, each recorded analytical data set contains at most 8, preferably at most 7, and more preferably at most, and more preferably exactly 5 different analytical parameters for olefins (n- / iso). Preferably, each recorded analytical data set contains at most, and more preferably exactly 6, different analytical parameters for cyclic olefins. Preferably, each recorded analytical data set contains at most, and more preferably exactly 6, different analytical parameters for aromatics.

[0047] The aforementioned maximum numbers for the respective analytical parameters each offer the advantage (individually or in combination) that, firstly, typical data which are usually determined during the production of the fuels can be used and no further measurement data have to be generated or determined for recording the analytical data set, and at the same time, the number of analytical data sets can be reduced as much as possible.

[0048] Preferably, the analytical dataset includes a (metric and a proportion assigned to it) that is characteristic of (all) paraffins with carbon numbers 6 or 7. This formulation (and analogously in subsequent formulations concerning other functionalities / carbon numbers) is to be understood in particular as meaning that the analytical dataset contains (only) one common metric for paraffins with carbon number 6 and for paraffins with carbon number 7. The analytical dataset therefore does not, in particular, include a (separate) metric (and its assigned proportion) that is characteristic of paraffins with only carbon number 6. Furthermore, the analytical dataset also does not, in particular, include a (separate) metric (and its assigned proportion) that is characteristic of paraffins with only carbon number 7.In particular, it is therefore not possible to distinguish between paraffins with carbon number 6 and paraffins with carbon number 7 (e.g., their proportion) based (only) on the analysis data set.

[0049] Additionally or alternatively, the analytical data set preferably includes a (analytical parameter and a parameter assigned to this analytical parameter) proportion that is characteristic of (all) paraffins with carbon numbers of 8, 9, or 10. Additionally or alternatively, the analytical data set preferably includes a (analytical parameter and a parameter assigned to this analytical parameter) proportion that is characteristic of (all) paraffins with carbon numbers of at least 11 carbon atoms. Additionally or alternatively, the analytical data set preferably includes a separate (analytical parameter and a parameter assigned to this analytical parameter) proportion for each of the individual carbon numbers 3, 4, and 5.

[0050] Additionally or alternatively, the analysis data set preferably includes a (analysis parameter and a proportion assigned to this analysis parameter) for naphthenes with carbon numbers between 5 and 10, so that advantageously all carbon numbers are grouped together.

[0051] Additionally or alternatively, the analytical data set preferably includes a fraction (analytical parameter and a fraction associated with this analytical parameter) for olefins (n- / iso) with carbon numbers 4 or 5. Additionally or alternatively, the analytical data set preferably includes a fraction (analytical parameter and a fraction associated with this analytical parameter) for olefins (n- / iso) with carbon numbers 7 or 8. Additionally or alternatively, the analytical data set preferably includes a separate fraction (analytical parameter and a fraction associated with this analytical parameter) for olefins (n- / iso) with the respective individual carbon numbers 3 and 6.

[0052] Through further aggregation of the data, for example through the merging of C as described here n and C n+1 By combining compounds with consecutive carbon numbers (within a single functionality) into a new (common) group (characterized by a single analytical parameter), a further reduction in data effort is advantageously achieved.

[0053] Preferably, all olefins (cyclic) are grouped together into a single group characterized by exactly one analytical parameter. Preferably, the analytical dataset does not differentiate between olefins (cyclic) with different carbon numbers based on the analytical parameters (and the associated proportions).

[0054] Additionally or alternatively, carbon numbers within the olefins (cyclic) and / or aromatics are preferably not grouped together (into a common group characterized by a single analytical parameter). It is conceivable that the analytical dataset contains no analytical parameter(s) characteristic of olefins (cyclic) and / or aromatics, thus allowing the dataset to be further reduced. It has been found that these compounds have only a minor influence on fuel properties.

[0055] It would also be conceivable that the analytical parameter, and preferably its associated proportion, is characteristic of a single chemical compound. In other words, the analytical dataset could contain a separate analytical parameter and an associated proportion for each chemical compound present in the fuel mixture. In this case, the analytical datasets would contain a much larger amount of data, and determining the analytical parameters and proportions would be more complex, but this would yield a maximum amount of information that can be used to train the machine learning combustion property determination model.

[0056] In a preferred method, combustion measurement data and / or at least one combustion parameter derived therefrom are recorded and / or determined for each training fuel mixture (in addition to the analysis data set). In a preferred method, this data is recorded using an (engine) test bench. Here, the training fuel mixtures are burned in a test engine (under real-world conditions), and corresponding combustion data are recorded. Preferably, the recorded and / or determined combustion data is an exhaust gas composition of the exhaust gases produced during the combustion of the training fuel mixture. Preferably, the combustion data also includes further data relating to particulate matter and / or soot particle emissions, or data relating to pollutant emissions (e.g., regarding the presence of nitrogen oxides).Preferably, the combustion parameter is derived from the combustion measurement data. This combustion parameter is preferably characteristic of an exhaust gas composition, an exhaust gas characteristic, a proportion of nitrogen oxides in the exhaust gas, a pollutant load, a particulate matter load, a soot particle load, an efficiency, a combustion duration and / or an efficiency (of the engine).

[0057] It would be conceivable that the combustion measurement data (and / or the combustion parameter derived from it) could be retrieved analogously to the analysis data sets from a storage device, particularly a non-volatile one (such as an external storage device like an external (backend) server). This offers the advantage that no further measurements need to be taken for training fuel mixtures that have already been measured on the test bench.

[0058] In a preferred method, a training data set is generated based on the analysis data sets of a plurality of training fuel mixtures, in particular comprising at least one analysis parameter, at least one corresponding proportion parameter, and / or at least one fuel parameter. Preferably, this training data set is used to train the aforementioned machine learning model (trainable combustion property determination model of machine learning).

[0059] In an advantageous method, at least one analysis data set, and preferably a plurality and particularly preferably all analysis data sets (in addition to at least one analysis parameter, a proportion parameter and / or a fuel parameter), comprises at least one component parameter, wherein the component parameter is preferably characteristic of at least one property of an engine and more preferably of a property of a component of an engine that was used to acquire at least one combustion parameter associated with the at least one analysis data set. Preferably, the component parameter is characteristic of a hardware and / or software setting of the engine or the engine control unit.Preferably, the component size is characteristic of an engine setting and in particular of a compression ratio, for compression ratios, for one and preferably for a plurality of valve settings, for an engine displacement, for injection settings, for an air-fuel ratio (mixture richness), for a turbocharger or the like.

[0060] In a preferred method, the training data set comprises at least one and preferably a plurality of data sets, wherein each data set is preferably characteristic of a training fuel mixture.

[0061] In a preferred method, at least one data set, and preferably each data set of the training data set, comprises at least one analysis parameter and a proportion parameter associated with the analysis parameter, and preferably a plurality of analysis parameters and proportion parameters associated with the analysis parameter, at least one component parameter and preferably a plurality of component parameters, at least one combustion parameter and preferably a plurality of combustion parameters, and / or at least one fuel parameter and preferably a plurality of fuel parameters.

[0062] In other words, an exemplary data set for a training fuel mixture (of the training data set) includes data relating to a fuel composition (e.g., three components X, Y, Z in proportions of 40 vol.%, 40 vol.%, and 20 vol.%), data relating to a fuel property (e.g., vapor pressure of the fuel mixture), data relating to a combustion parameter (e.g., exhaust gas composition), and data relating to component influences (e.g., compression ratios).

[0063] Preferably, the acquired and / or determined analysis datasets (analysis parameters, proportion parameters, component parameters, and / or fuel parameters) are used as inputs for training the machine learning combustion property determination model. In a preferred method, the acquired (and / or determined combustion measurement data and / or the combustion parameters derived therefrom) are used as outputs for training the machine learning model. In other words, data relating to the training fuel mixture itself (data relating to the individual fuel components as well as data relating to the physicochemical properties of the fuel mixture) and / or component data are used as inputs for the machine learning model, while, for example, an exhaust gas composition serves as the model's output.

[0064] In other words, the machine learning combustion property determination model is trained to predict exhaust gas composition based on any given fuel mixture. Preferably, in this state, the trained machine learning model is a black-box model that cannot be easily interpreted or explained.

[0065] In a further step of the inventive method, the combustion property determination model is trained on the basis of the training data set. In a further step of the inventive method, at least one and preferably a plurality of influencing variables are determined by applying a machine learning method for explaining the trained combustion property determination model, wherein the at least one influencing variable is characteristic of an influence of at least one analysis variable and / or at least one proportion variable on the combustion variable.

[0066] In other words, for an analysis parameter and / or an associated proportion, a quantity or factor is determined that indicates how strongly a variation of the analysis parameter and / or the associated proportion affects the resulting combustion parameter, i.e., it is a measure of the influence of this parameter on the combustion parameter. For example, it can be determined that an analysis parameter X has no significant influence on the exhaust gas composition, but an analysis parameter Y significantly influences the exhaust gas composition.

[0067] In a preferred method, an influencing factor is determined for at least one, preferably for a plurality, and particularly preferably for each analytical parameter. In a preferred method, an influencing factor is determined for at least one, preferably for a plurality, and particularly preferably for each proportion assigned to the analytical parameter.

[0068] In an advantageous method, an influencing factor is determined for at least one component size, preferably for a plurality, and particularly preferably for each component size, wherein the influencing factor is characteristic of the influence of the at least one component size on the combustion parameter. In an advantageous method, an influencing factor is determined for at least one fuel parameter, preferably for a plurality, and particularly preferably for each fuel parameter, wherein the influencing factor is characteristic of the influence of the at least one fuel parameter on the combustion parameter. This offers the advantage that it is possible to determine, for example, how strongly a vapor pressure of the fuel mixture or a compression ratio affects an exhaust gas composition.

[0069] One preferred method for explaining the trained machine learning combustion property determination model is the LIME or SHAP approach. LIME stands for "Locally Interpretable Model-agnostic Explanation," while SHAP stands for "Shapley Additive Explanations." Both methods make it possible to explain a machine learning model's decision to a user and to determine the extent to which individual variables influence the prediction.

[0070] As mentioned above, the analysis parameters, the proportion parameters, the fuel parameters and / or the component parameters serve as input parameters for the machine learning model, while the combustion parameter is obtained as the output parameter of the model.

[0071] In a preferred method, the input quantity comprises a multitude of variables, and preferably an influencing factor is determined for each of these variables. These variables are preferably the analysis parameters, the proportion parameters, the fuel parameters, and / or the component parameters.

[0072] For clarity, the following explanations will refer to variables, where a variable is always understood to be an analysis parameter, a proportion, a fuel parameter, or a component parameter. In other words, the term "variable" can be replaced by "analysis parameter, proportion, component parameter, or fuel parameter."

[0073] Therefore, the preferred approach is to determine how each variable affects the combustion parameters. In other words, by applying a method to explain the trained machine learning combustion property determination model, the influence of each variable is determined. For example, this allows us to determine the influence of the fuel mixture vapor pressure, the volume fraction of oxygenates, the engine injection setting, etc., on the resulting exhaust gas composition, or how these variables affect efficiency.

[0074] One preferred method for explaining the trained machine learning combustion property determination model is a model-agnostic XAI (explainable artificial intelligence) method. Model-agnostic in this context means that the method is capable of explaining the decisions of virtually any machine learning model.

[0075] In an advantageous embodiment of the method according to the invention, the at least one determined influencing variable and preferably the plurality of determined influencing variables is characteristic for a global interpretability and / or for a local interpretability of the combustion property determination model (the output variable or the decision of the machine learning model).

[0076] In this context, global interpretability refers to how a variable generally (across all datasets used) affects the output variable, or what influence this variable has. For example, it could be determined how the vapor pressure of the fuel mixture generally affects the output (across all available datasets), or it could be determined that, for example, a specific component on the engine generally has no significant influence on the exhaust gas composition.

[0077] Local interpretability, in this context, refers to how a variable manifests itself within a single dataset used to generate a single output variable. For example, it could be determined what influence a component of the given training fuel mixture has on the exhaust gas composition, i.e., how strongly a proportion of hexane in a specific fuel formulation affects the measured exhaust gas composition.

[0078] One preferred method for explaining the trained combustion property prediction model of machine learning is the LIME approach. LIME (Local Interpretable Model-agnostic Explanations) is a technique introduced by Ribeiro et al. (2016) that provides explanations for individual predictions generated by black-box machine learning models. The key idea behind LIME is to approximate a complex model with a simpler, interpretable model centered around a specific instance. In this context, an instance is understood as a set or data set comprising an input and an output variable. In other words, an instance includes a multitude of variables (analysis parameters, proportion parameters, fuel parameters, component parameters) and at least one combustion parameter.

[0079] In an advantageous method, the training data set comprises a plurality of input variables and a plurality of output variables associated with the input variables, wherein the plurality of input variables each comprises at least one variable and preferably a plurality of variables, wherein the at least one variable and preferably the plurality of variables is selected from a group of variables comprising an analysis variable, a proportion variable, a fuel variable and a component variable, and wherein the output variable is a combustion variable.

[0080] In an advantageous method, the determination of at least one influencing variable by applying a method for explaining the trained combustion property determination model comprises, in one step, the generation of one, and preferably a plurality, of permuted input variables based on an input variable by permuting at least one variable of this input variable, wherein the permuted input variables differ from each other pairwise by a value of this variable. Preferably, at least one variable (of this input variable) is permuted. For this purpose, the value of this variable is preferably set to different values, which are based on the original value of this variable in the input variable. In other words, the value of one variable is permuted, while all other variables of the input variable are held constant. This step could be called "perturbation".

[0081] Preferably, the at least one permuted input variable is one in which a variable is permuted or changed. Preferably, a plurality of permuted input variables are generated in this way, each differing from the other by a permutation of exactly one variable. For example, a variable X (e.g., the proportion of hexane in the fuel mixture) could be permuted in 1% increments, such that the permuted input variables contain the values ​​for this variable X from, for example, (X - 20%·X) to (X+20%·X) in 1% increments.

[0082] Preferably, the extent of the permutations depends on the respective variable and, in particular, on the value of that variable. It would also be possible to repeat and / or adjust such a permutation depending on a subsequently determined influencing factor. This is useful, for example, if it is found in the further procedure (see below) that the 1% steps were too large and the results obtained exhibit such a scatter that a reliable conclusion cannot be drawn. In this case, the permutation could then be repeated with smaller steps, for example, with 0.1% steps. In a preferred method, a multitude of permuted input variables are generated for each variable of the input quantity.

[0083] In an advantageous embodiment, a further step involves determining a multitude of permuted combustion parameters (permuted output parameters) based on the multitude of permuted input parameters by applying the trained combustion property determination model. In other words, the generated permuted input parameters (permuted variables) are used as input for the trained combustion property determination model described above, and the permuted combustion parameters are obtained as output. This step could be referred to as "prediction."

[0084] In an advantageous method, a simple model (machine learning) is fitted to the permuted input variables and the permuted combustion variables (output variables) in a further step, whereby the permuted combustion variables (output variables) are weighted based on their similarity to the original combustion variable (the combustion variable assigned to the non-permuted input variable).

[0085] Preferably, the weighting is characteristic of a difference between the permuted combustion parameter and the non-permuted combustion parameter. In other words, it is determined how far the permuted combustion parameter deviates from the non-permuted combustion parameter. In a preferred method, the weighting is performed by determining a metric difference between the two combustion parameters, for example, by determining a Euclidean difference. This step could be referred to as "weighting".

[0086] In other words, a simple (interpretable) model is trained based on the weighted permuted input variables and permuted combustion variables to approximate the behavior of the black-box model locally around these input variables and combustion variables, respectively.

[0087] Preferably, the simple model is a linear regression, a decision tree, or something similar. This step could be called "fitting." For example, the values ​​of the permuted variable could be plotted against the values ​​of the permuted combustion parameter on the x-axis. Preferably, the output of the simple model is a fit function between these points (permuted variable as the x-coordinate, permuted combustion parameter as the y-variable), for example, in the simplest case, in the form of a regression line (linear regression).

[0088] In an advantageous embodiment, the influencing factor is determined in a further step of the method based on the adapted simple model or on the basis of the regression function determined by the adapted simple model. Preferably, the influencing factor is determined based on a coefficient of the regression function. In other words, the influencing factor is characteristic of the influence of the variable, which was permuted to generate the permuted input variables, on the value of the combustion variable.

[0089] For example, in the case of linear regression, the slope of the regression line is characteristic of the influence of the variable. In a preferred method, the influencing factor is characteristic of the simple model and serves to explain the combustion property determination model (black-box model) around this dataset.

[0090] In other words, it is proposed to reduce an instance of a complex and inexplicable model (black-box model) to an interpretable level and to determine an influencing factor for a single variable by permuting it according to the procedure described above. In a preferred method, the procedure described above is performed for each variable, with an influencing factor being determined for each variable.

[0091] One preferred or alternative method for explaining the trained machine learning combustion property prediction model is the SHAP approach. SHAP (Shapley Additive Explanation) is a method based on the use of Shapley values ​​from game theory. It compares how a model behaves when a specific feature is included in a feature set or not. Features not included in a feature set are filled with randomly selected values ​​from a background dataset. This process is repeated for each feature in the dataset for all possible feature combinations. Finally, a weighted average of all possible combinations is calculated. This calculated value is the Shapley value for that specific feature. The higher the value of the Shapley value, the greater the variable's influence on the prediction.Values ​​around zero therefore reflect only a minor influence.

[0092] Preferably, an input variable comprises a plurality of variables, preferably n variables. In an advantageous method, one variable is selected in a single step. Preferably, this is the variable whose influence is to be determined, i.e., whose influencing factor is to be ascertained. In an advantageous method, the determination of the at least one influencing factor in a single step comprises the determination of a plurality of coalitions and, in particular, the determination of (n-1) 2 Coalitions for an input variable, where the input variable preferably comprises n variables. A coalition is understood here to be a group of variables.

[0093] Advantageously, when determining the multitude of coalitions, the remaining, especially unselected, variables are taken into account. In other words, all variables except the selected variable are considered. Preferably, the number of coalitions depends on a number of variables that comprise the corresponding input variable. Preferably, the determination of the multitude of coalitions includes (n-1) 2 Coalitions allow for the selection of a single variable in one step. Preferably, this is the variable whose influence (SHAP value) is to be determined. In a preferred method, a multitude of coalitions (groups of variables) are formed based on the remaining (n-1) variables. In a preferred method, based on the (n-1) variables (n-1) 2 Coalitions are created. The (n-1) coalitions are preferred. 2Coalitions are obtained by first forming a first generation of coalitions that includes all (n-1) variables.

[0094] Preferably, at least one, and preferably a plurality, of second generations of coalitions are formed, each comprising (n-2) of the original (n-1) variables, wherein the second generations preferably differ from each other pairwise by one variable. In a preferred method, further generations are generated according to the number of variables. Preferably, a (n-1) penultimate generation is generated, each comprising only one of the original (n-1) variables, and preferably, a final generation is generated, which comprises no variables.

[0095] In a preferred method, the value of each variable that is no longer to be considered in the respective coalition is set to a value that occurs in another input variable of the training dataset.

[0096] In an advantageous method, for the plurality of coalitions and preferably for each of the (n-1) 2 Coalitions each determine a first and a second combustion parameter using the (above-described) trained combustion property determination model, whereby in determining the first combustion parameter the plurality of coalitions and preferably each of the (n-1) 2 Coalitions are each used in combination with the selected variable with its original value as input, and in determining the second combustion quantity, the plurality of coalitions and preferably each of the (n-1) 2Coalitions are used in combination with the selected variable, each with a modified value as input. Preferably, the modified value of the selected variable is a value that occurs in a different input (another dataset) of the training dataset.

[0097] Preferably, when determining the second combustion parameters, a value is assigned to the selected variable as occurs for another input parameter in the training dataset. In an advantageous method, for each of the plurality of coalitions, and preferably for each of the (n-1) 2 Coalitions determine a marginal amount of the selected variable by calculating the difference between the first and second combustion quantities. Preferably, this is done for each of the (n-1) 2Each coalition identifies a marginal contribution. In an advantageous method, an influencing factor for the selected variable is determined by calculating the average of the numerous marginal contributions.

[0098] In a preferred method, the selection is based on the multitude of marginal contributions and preferably on the (n-1) 2 The marginal contributions of the selected variable are used to determine an influencing factor for the selected variable, whereby the mean of the marginal contributions (of the selected variable) is calculated, which is a SHAP value (of the selected variable). Preferably, the SHAP value is a local influencing factor. In this context, "local" means that the SHAP value is characteristic of the influence of the selected variable on the input variable used to determine the value.

[0099] This procedure is illustrated by the following example. An input quantity consisting of four variables A, X, Y, Z is used, where the influence of variable A on the output quantity (combustion quantity) is to be determined. For the remaining three variables X, Y, Z, 2 3So there are 8 possible coalitions. In the first coalition, all three variables are contained in their original form (X, Y, Z). For the second level, there are several second coalitions. These are formed by replacing the value of one variable with another value, such as that found in a different input variable of the training dataset (marked with * below, so that the possible second coalitions are (X*,Y,Z), (X,Y*,Z), and (X,Y,Z*). For the third level, there are several third coalitions, which are formed analogously, replacing a second variable. The possible coalitions here are (X,Y*,Z*), (X*,Y,Z*), and (X*,Y*,Z). The last coalition is formed by replacing three variables (X*,Y*,Z*).

[0100] For all eight coalitions, a combustion parameter is determined by applying the combustion property determination model. This is done once for each of the eight coalitions in conjunction with the selected variable in its original form (e.g., (A,X,Y,Z)) and once for each of these eight coalitions in conjunction with the selected variable in a modified form (e.g., (A*,X,Y,Z)). For each coalition, a deviation between the first and second combustion parameters is determined (corresponding to the marginal contribution of variable A). Based on these marginal contributions, an average marginal contribution of variable A can be calculated. This average marginal contribution is the desired SHAP value and thus the influencing parameter to be determined. This is a local influencing parameter.In other words, this local influence refers to an amount of variable A on the specific single instance comprising variables A, X, Y, and Z, as well as the reference combustion quantity.

[0101] In a preferred method, a local influencing factor (SHAP value) is determined for a plurality, and preferably for each of the n variables, in the manner described above. The procedure described so far for determining the SHAP values ​​involves a determination starting from only one instance and thus corresponds to the determination of a local influencing factor. Preferably, the local influencing factor is characteristic of an influence or contribution of a single variable on the combustion parameter within that single instance. In a preferred method, local influencing factors (SHAP values) are determined for each of the n variables for a plurality of instances. Preferably, an averaged, and preferably global, influencing factor is determined for each of the n variables based on the plurality of local influencing factors.Preferably, this global influencing factor is characteristic of the influence of the respective variable across all instances used for determination. Preferably, this global influencing factor is an absolute SHAP value or a global SHAP value. The use of a global influencing factor offers the advantage that overarching influencing factors can be determined for individual variables. For example, an overarching influence of a component on the combustion parameter could be determined, or it could be determined whether the presence or absence of a specific fuel component has no or a significant influence on the resulting exhaust gas composition.

[0102] The LIME and SHAP methods described above each represent an approach to making a complex machine learning model interpretable or explainable, specifically in the form of local interpretability. The SHAP approach offers the additional advantage that (as described above) the machine learning model can also be made globally interpretable.

[0103] In a preferred method, at least one determined influencing factor, and preferably the plurality of determined influencing factors, is made available for output to a user. It would be possible for the determined influencing factors to be displayed graphically.

[0104] Preferably, a local influencing factor, determined using the SHAP approach, is represented in the form of a waterfall plot (see figure description). Such a waterfall plot preferably depicts the explanation for a single prediction (single instance). A waterfall plot is a representation in which a normalized prediction variable is plotted on the x-axis; for example, the determined combustion variable here corresponds to f(x) = 1. The x-axis can be started at the origin with the expected value E[f(x)]. This is the average value of all predictions determined for the single instance (see above regarding the determination of SHAP values). On the y-axis, the individual variables are arranged in ascending order according to their SHAP values. In addition to these values, their values ​​as they exist in the underlying dataset (instance) can also be plotted.

[0105] Starting with the expected value E[f(x)], the positive and negative SHAP values ​​for each variable are plotted as bars (with the SHAP value indicated), such that positive bars point to the right and negative bars to the left. Ideally, all these values ​​are arranged in such a way that a function value of f(x) = 1 is achieved. From such a waterfall plot, the user can see at a glance which variable has the strongest influence on the prediction in a single instance.

[0106] In a preferred method, the local influencing factors determined using the LIME approach could be represented in a similar form, for example in the form of a waterfall plot.

[0107] Preferably, the determined global influencing factors (absolute SHAP values ​​averaged across all instances) are displayed in the form of a dot plot. In a dot plot, the SHAP values ​​are preferably plotted on the x-axis, while the variables are arranged on the y-axis according to their influence on the prediction, with the variables being arranged according to their global influence (based on their absolute SHAP values). For each variable, all specified SHAP values ​​are plotted, with the individual data points color-coded according to the variable's value. For example, high values ​​of the variable are plotted in red and low values ​​in blue.

[0108] Such a dot plot offers the advantage that it is immediately apparent (to the user) which variable has the strongest global influence and how the local influence of the variable depends on a value of the variable.

[0109] The graphical representation methods described above represent only a small selection of possible representation methods and are listed only as examples, so the subject matter of the present invention is not limited to these representation possibilities. It would also be conceivable to depict only a portion of the determined local and / or global influencing factors.

[0110] The method described so far offers the advantage that a previously incomprehensible black-box model, used to predict a combustion property or quantity, can be made understandable to the user, allowing them to grasp how individual variables affect the prediction. This provides experts with a simple way to directly adjust the relevant parameters when optimizing a fuel composition, for example, to minimize exhaust emissions or pollutant levels, instead of having to conduct numerous experiments to determine the influence of individual variables.

[0111] The generated, trained, explainable machine learning model (combustion property determination model) can now be used for a variety of different processes. For example, knowing the influencing factors makes it easier to determine a new fuel composition. The proposed method also offers the advantage that component influences can be directly considered when predicting combustion properties (combustion parameters), which is not currently possible in this form with existing technology. A possible objective for applying the trained model could be to determine a new fuel composition to further reduce pollutant emissions or to increase the efficiency of the fuel mixture. Likewise, the trained model could be used to determine a new fuel composition based on a known fuel composition, e.g.,The model can be used to predict combustion properties such as exhaust gas composition, as obtained from a corresponding analysis of synthetic fuels. Furthermore, the trained model could also be applied to optimize engine components and investigate their influence on exhaust gas composition.

[0112] In a preferred method, a database is generated based on the determined influencing factors for the multitude of analytical parameters, proportions, fuel parameters, and / or component parameters, and is preferably made available for output to the user. This database is preferably used for a method for determining new fuel compositions or for a method for optimizing engine components, as described further below.

[0113] The method described above makes it possible to determine, for all components of the fuel mixture and their proportions, for the physicochemical properties of the fuel mixture, as well as for components of the engine used for combustion and its software and / or hardware settings, how strongly these individual factors affect the resulting exhaust gas composition and / or engine performance. This knowledge can then be used to optimize existing engines and engine settings, known fuel mixtures, etc., in such a way that emissions, especially pollutants (particulate matter, soot, nitrogen oxides, etc.), can be reduced, or engine efficiency can be increased (e.g., by increasing the combustion efficiency of the fuel mixture in the engine).

[0114] As described above, the combustion parameters (e.g., exhaust gas composition) are influenced by a multitude of variables, such as the parameters of the analysis, the proportions, the fuel parameters, or the component sizes. Thus, the combustion parameters could be considered, in a sense, as a function of these variables. One objective of the present invention is to minimize emissions or maximize the efficiency of a fuel composition.

[0115] Therefore, the present invention is further directed to a method for the particularly predictive determination of a particularly new fuel composition using the (above described) trained explainable combustion property determination model, which in one step comprises the determination of a particularly local extreme point of a combustion parameter as a function of at least one variable and preferably as a function of a plurality of variables, wherein the at least one variable and preferably the plurality of variables are selected from a group of variables comprising an analysis parameter, a proportion parameter, a fuel parameter and a component parameter.

[0116] In a method according to the invention, the determined influencing factors are taken into account when determining the particularly local extreme point. In a further step, a new fuel composition is determined based on the determined particularly local extreme point, wherein the determination of the new fuel composition comprises the determination of at least one and preferably a plurality of analytical parameters and the determination of at least one and preferably a plurality of proportion parameters assigned to the analytical parameters.

[0117] Taking the influencing factors into account offers the advantage that the focus when determining a new fuel composition can be placed on significantly fewer variables than would be necessary with state-of-the-art methods.

[0118] Preferably, the determination of the combustion parameter, particularly its local extreme point, involves identifying a local minimum or a local maximum. This is preferably dependent on the nature of the combustion parameter. In a preferred method, the combustion parameter is an exhaust gas composition, and preferably a proportion of pollutants in the exhaust gases. For example, in such a case, the combustion parameter represents a proportion of nitrogen oxides in the exhaust gases or a value for particulate matter or soot pollution. In these cases, the local extreme point is a local minimum at which the combustion parameter reaches a minimum value. The aim is thus to determine a new fuel composition with respect to minimized emissions.

[0119] On the other hand, the combustion parameter can also be a measure of combustion efficiency, for example, a coefficient of performance (COP). In this case, the local extreme point is a local maximum. The aim here is to determine a new fuel composition that will increase the COP or COP.

[0120] In this context, a local extremum is understood to be determined as a function of one or more variables, but not as a function of all variables present in the system that were used as inputs for training the model, and especially not as a function of a single step. Specifically, this involves a complex system with a multitude of variables that are not independent of each other and, in fact, influence one another.

[0121] In a preferred method, the particularly local extreme point is determined as a function of selected variables, the variables being chosen according to their influence. In another preferred method, the unselected variables are not considered when determining the particularly local extreme point and are preferably assumed to be constant. This offers the advantage that variables which have no or hardly any influence on the resulting exhaust gas composition do not need to be taken into account.

[0122] In a preferred method, at least one analytical parameter and at least one associated proportion are determined based on the identified local minimum, and preferably a plurality of analytical parameters and associated proportions are determined. Preferably, the plurality of analytical parameters and associated proportions are characteristic of a composition of the new fuel. Preferably, the component dimensions can be assumed to be constant when determining a new fuel composition.

[0123] In a preferred method, it would also be possible to determine the local extreme point as a function of one variable or several variables simultaneously. Preferably, the determination of the local extreme point is carried out as a function of the variables selected (based on the influencing factors). Preferably, a new fuel composition is determined based on the determined local extreme point.

[0124] The use of the trained combustion property determination model offers not only the possibility of determining new fuel compositions but also the potential for optimizing components. Analogous to the procedure for determining a new fuel composition, the trained combustion property determination model could be used to identify the local extreme point as a function of one or more component parameters, while keeping all variables associated with a fuel (analysis parameter, proportion, fuel property) constant.

[0125] Therefore, the present invention is also directed to a method for the particularly predictive determination of an optimized component size using the (above described) trained combustion property determination model, which in one step comprises the determination of a particularly local extreme point of a combustion parameter as a function of at least one component size and preferably as a function of a plurality of component sizes.

[0126] In an advantageous method, the determined influencing factors are taken into account when determining the particularly local extreme point. In a further step, an optimized component size is determined based on the identified particularly local extreme point.

[0127] Preferably, the optimized component size is characteristic of at least one component of the engine or of at least one software or hardware setting of the engine. For example, the optimized component size is an optimal valve setting or an optimal compression ratio.

[0128] In a preferred method, the results of the procedure for determining a new fuel composition and / or the results of the procedure for determining an optimized component size are validated with regard to the reliability of the results. This can be achieved, firstly, by using computer-aided methods known from the prior art. Secondly, it would be possible to measure a real sample of the determined new fuel composition on the engine test bench and subsequently compare the measurement results with the result of the combustion property determination model.

[0129] The present invention further relates to a device for generating a trained combustion property determination model of machine learning for determining at least one combustion parameter of a given fuel using analysis data sets, each of which is characteristic of a composition of a given fuel, wherein each analysis data set for the respective given fuel comprises at least one analysis parameter and a proportion parameter assigned to the analysis parameter, wherein the analysis parameter is characteristic of a group of compounds from the given fuel which have at least one given functionality, and wherein the proportion parameter is characteristic of a quantity fraction of the compounds in the respective fuel which are grouped in the group characterized by the analysis parameter.

[0130] Furthermore, the device is designed to provide a trainable combustion property determination model of machine learning, which includes a set of trainable parameters and which, based on an analysis data set acquired for a given fuel or data derived therefrom as input, generates at least one combustion parameter as output, wherein the combustion parameter is characteristic of at least one combustion property of the fuel.Furthermore, the device is configured to generate a training dataset comprising a multitude of acquired and / or determined analysis datasets for training fuels, which are used to generate the training dataset, as well as a multitude of (combustion measurement data and / or combustion parameters derived therefrom), wherein (the combustion measurement data and / or the combustion parameter derived therefrom) are each characteristic of at least one combustion property of the respective training fuel. Furthermore, the device is configured to train the combustion property determination model based on the training dataset.

[0131] According to the invention, the device is configured to determine at least one influencing variable by applying a method for explaining the trained combustion property determination model of machine learning, wherein the at least one influencing variable is characteristic of an influence of the at least one proportion variable and / or the at least one analysis variable on the combustion variable.

[0132] In a preferred embodiment, the device includes a data transmission device suitable and designed to retrieve the multitude of acquired and / or determined analysis data sets from an external source. Preferably, the data transmission device is configured to retrieve the multitude of acquired and / or determined analysis data sets from a storage device, particularly a non-volatile one (such as an external storage device like an external (backend) server). It would also be conceivable for the data transmission device to retrieve the multitude of acquired and / or determined analysis data sets (directly) from a measuring device used to determine the analysis data sets. Preferably, the measuring device is configured to use the method specified in DIN EN ISO 22854.to apply standardized procedures for the determination of hydrocarbon groups and oxygen-containing compounds in gasoline and ethanol fuel (E85) (multidimensional gas chromatographic method ISO / DIS 22854:2020).

[0133] Preferably, the data transmission device is suitable and designed to retrieve and / or receive at least one and preferably a plurality of analysis parameters, at least one and preferably a plurality of proportion parameters, at least one and preferably a plurality of fuel parameters and / or at least one and preferably a plurality of component parameters.

[0134] Preferably, the data transmission device is configured to retrieve combustion measurement data and / or a combustion parameter derived therefrom from another measuring device. This measuring device is preferably located downstream of an engine test bench or is part of the engine test bench. This other measuring device is preferably configured to determine exhaust gas composition, pollutant levels, particulate matter or soot particles, nitrogen oxide content, and the like. Preferably, the data transmission device is suitable and designed to retrieve at least one, and preferably a plurality, of component parameters from a storage device, particularly a non-volatile one (such as an external storage device like an external (backend) server). It would also be conceivable to retrieve the at least one, and preferably a plurality, of component parameters from an engine test bench.

[0135] In a preferred embodiment, the device is configured to provide the at least one determined influencing factor and preferably the plurality of determined influencing factors, in particular for output to a user via an output device or for determining a new fuel composition (as described above).

[0136] In a preferred method, the device has a data storage device on which the combustion property determination model to be trained or the trained combustion property determination model is stored.

[0137] Preferably, the device for generating a trained combustion property determination model of machine learning for determining at least one combustion parameter is configured, suitable, and / or designed to execute the above-described method for generating a trained combustion property determination model of machine learning for determining at least one combustion parameter, as well as all method steps already described above in connection with the method (or a preferred embodiment of the method), individually or in combination. Conversely, the method can be equipped with all features described within the device for generating a trained combustion property determination model of machine learning for determining at least one combustion parameter, individually or in combination.

[0138] The present invention further relates to a device for determining a new fuel composition using a trained combustion property determination model, which was trained according to a method according to a previously described embodiment, wherein the device is configured to determine a particularly local extreme point of a combustion parameter as a function of at least one variable and preferably as a function of a plurality of variables, wherein the at least one variable and preferably the plurality of variables are selected from a group of variables comprising an analysis parameter, a proportion parameter, a fuel parameter and a component parameter.

[0139] Furthermore, the device is designed to take into account the determined influencing factors when determining the particularly local extreme point and to determine a new fuel composition on the basis of the determined particularly local extreme point, wherein the determination of the new fuel composition includes the determination of at least one and preferably a plurality of analytical parameters and the determination of at least one and preferably a plurality of proportion parameters assigned to the analytical parameters.

[0140] Preferably, the device for the predictive determination of a new fuel composition is configured, suitable, and / or designed to carry out the above-described method for the predictive determination of a new fuel composition, as well as all process steps already described above in connection with the method (or a preferred embodiment of the method), individually or in combination. Conversely, the method can be equipped with all features described within the scope of the device for the predictive determination of a new fuel composition, individually or in combination.

[0141] The present invention further relates to a computer program or computer program product, comprising program means, in particular a program code, which represents or encodes at least some and preferably all of the process steps of the methods according to the invention (method for generating a trained combustion property determination model of machine learning for determining at least one combustion parameter; method for the particularly predictive determination of a particularly new fuel composition) and preferably one of the described preferred embodiments and is designed for execution by a processor device.

[0142] The present invention further relates to a data storage device on which at least one embodiment of the computer program according to the invention or a preferred embodiment of the computer program is stored.

[0143] Further advantages and embodiments can be seen from the attached drawings: It shows: Fig. 1 a black box model according to the state of the art, Fig. 2 a schematic representation of an embodiment of a method according to the invention for generating a trained combustion property determination model of machine learning, Fig. 3 a graphical representation for determining an influencing factor (LIME approach), Fig. 4 a graphical representation to illustrate an influencing factor (SHAP approach), Fig. 5 a schematic representation of an embodiment of a method according to the invention for determining a new fuel composition.

[0144] In Fig. Figure 1 schematically illustrates the functionality of a machine learning model. When using state-of-the-art machine learning models, such as neural networks, users face challenges in understanding the model's decisions and decision-making process. For example, a trained model (previously trained with training data) processes an input variable 2 and produces an output variable 6. Because the model's decision is neither comprehensible nor explainable to an external user, it is referred to as a "black box" model.

[0145] In Fig. Figure 2 shows a schematic representation of an embodiment of a method according to the invention for generating a trained combustion property determination model 26. First, an untrained (trainable) combustion property determination model 22 is provided and trained with training data 20. The training data 20 comprises a multitude of variables, for example, an analysis parameter 10, a proportion parameter 12, a fuel parameter 14, and / or a component parameter 16. These variables or parameters 10 to 16 can be determined, for example, by means of measurements (e.g., complete chemical analysis) or retrieved from a database. It would be conceivable that the analysis parameters 10, the proportion parameters 12, and the fuel parameters 14 could be obtained from a chemical analysis of a fuel mixture.

[0146] An example of an analysis parameter 10 represents the group of oxygenates present in the fuel mixture at a proportion of 20 vol% (proportion parameter 12). An example of a fuel parameter 14 represents an octane rating of 95. The training dataset can also include at least one component parameter 16. This is a parameter that is characteristic of an engine or a component of the engine, or of a software or hardware setting of the engine or engine control unit. An example of a component parameter 16 represents a compression ratio of 10:1. Parameters 10 to 16 serve as input parameters for the combustion property determination model.

[0147] The training dataset 20 also includes at least one combustion parameter 18. Such a combustion parameter 18 is characteristic of at least one property of the fuel mixture during combustion, e.g., on an engine test bench. An exemplary combustion parameter 18 is characteristic of an exhaust gas composition and includes, for example, a proportion of particulate matter in the exhaust gases or a proportion of nitrogen oxides in the exhaust gases. The combustion parameter 18 serves here as the output parameter for the combustion property determination model. In other words, the training dataset 20 contains a multitude of input parameters (parameters 10, 12, 14, 16) and output parameters (combustion parameters 18).

[0148] The training dataset 20 is used to train an untrained combustion property determination model 22, resulting in a trained combustion property determination model 24 (step S1). At this stage, this model is a black-box model. In a further step S2, a method for explaining the trained combustion property determination model (explainable model 26) is applied. This can be done, for example, using the LIME approach, the SHAP approach, or another prior art approach for explaining a trained machine learning model. Preferably, the application of the method for explaining the trained combustion property determination model 26 in a step S3 includes the determination of influencing variables 28. Preferably, an influencing variable 28 is determined for each variable (analysis variable 10, proportion variable 12, fuel variable 14, component variable 16).

[0149] In Fig. Figure 3 shows a schematic graphical representation of a plot that could occur when determining the influencing factors according to the LIME approach. In this procedure, based on a dataset of training data 20, which comprises a multitude of variables, one variable is first selected whose influence factor 28 is to be determined. The variables used here can be the analysis parameters 10, the proportion parameters 12, the fuel parameters 14, and / or the component parameters 16. In addition, this dataset includes a combustion parameter 18 (initial combustion parameter v0).

[0150] As described above, the variables represent the input for the combustion property determination model. Based on the selected variable, a multitude of permuted inputs are generated. This is achieved by permuting the selected variable. In other words, a multitude of data sets are created, which differ from each other in the value of the selected variable.

[0151] For example, a variable such as proportion 12, which has a value of 10 vol.%, could be selected (output variable x0). This numerical value is then permuted between 8 and 12 vol.%, generating a corresponding number of data sets or input values ​​(variables x1 to x6, reference 36). Using these permuted input values ​​and the trained combustion property determination model, a combustion parameter (permuted combustion parameters) is then determined as an output (combustion parameters v1 to v6, reference 42).

[0152] In a further step, the permuted input variables and the permuted combustion variables are adjusted, or more precisely, fitted, using a simple, interpretable model. In other words, a regression function is determined for the permuted data. Fig. Figure 3 illustrates this situation for the simple case where the simple model is a linear regression and the regression function is a straight line. Graph 30 is shown, where the variable x is plotted on the x-axis 32 and the combustion quantity v on the y-axis 34. The point with the coordinates (x0 / v0) represents the actual measured value used as the basis for the permutation above. The reference symbol 44 denotes the regression function in the form of a regression line. In this case, the slope of this line is characteristic of the influence of the variable x on the combustion quantity. From this, the influence quantity 28 to be determined can be derived.

[0153] In Fig. Figure 4 shows a so-called waterfall plot (50) that can be used to visualize determined influencing factors in the form of SHAP values. Here, the influencing factors (28) are determined using the SHAP approach. In the present example, an input quantity comprising four variables y1, y2, y3, and y4 is used. The combustion quantity v0 is assigned to this input quantity as the output quantity. First, the influencing factor for variable y1 is to be determined. For this purpose, variable y1 is first extracted from the dataset. For the remaining three variables y2, y3, and y4, 2 3Eight (variable) coalitions were generated. In this context, a coalition is understood to be a group of variables. For this purpose, the values ​​of the individual variables are successively replaced by other values ​​for that variable. These values ​​can be taken from a different dataset (different input variable) that was also used to train the combustion property determination model. These modified variables are subsequently marked with an asterisk (*). The following eight coalitions result for the three variables y2, y3, and y4: Erste Generation (y2, y3, y4) Zweite Generation (y2*, y3, y4), (y2, y3*, y4), (y2, y3, y4*) Dritte Generation (y2*, y3*, y4), (y2, y3*, y4*), (y2*, y3, y4*) Vierte Generation (y2*, y3*, y4*)

[0154] For each of these eight coalitions, a combustion parameter is determined by applying the trained combustion property determination model. Specifically, a first combustion parameter is generated for each coalition, using one of the aforementioned coalitions in conjunction with the variable y1 as the input. Furthermore, a second combustion parameter is generated for each of the eight coalitions, using one of the aforementioned coalitions in conjunction with the modified variable y1* (analogous to the above).

[0155] For each of the eight coalitions, the difference between the first and second combustion parameters is determined and weighted. This represents the marginal contribution of this variable to the result (determined combustion parameter). The average of the eight marginal contributions is then calculated. This average marginal contribution is a SHAP value, which is characteristic of the influence of variable y1 on the value of combustion parameter v0. From this, the influence factor 28 to be determined can be derived.

[0156] The described procedure is repeated analogously for the remaining variables y2, y3, and y4, so that a SHAP value, and thus an influencing factor, can also be determined for each of these variables. Fig. Figure 4 shows a so-called waterfall plot (50), which serves to visualize the determined SHAP values. The combustion quantity, normalized to 1, is plotted as a function of the variables on the x-axis (52).

[0157] In other words, the point f(x)=1 (reference symbol 56) corresponds to the initial combustion parameter v0, as contained in the original, unmodified dataset. The starting point is the expected value E[f(x)] (reference symbol 58), which in this example has been arbitrarily set to 0.2. This expected value is the average of all determined first and second combustion parameters. On the y-axis 54, the variables y1, y2, y3, and y4 (reference symbols 60, 62, 64, 66) are arranged according to their SHAP values ​​(variable y4 has the greatest influence here). Reference symbols 68, 70, 72, and 74 denote the SHAP values ​​for the four variables. In the example shown, variables y1 and y2 have a positive influence of 0.2, while variable y3 has a negative influence of -0.1. Variable y4 has the greatest influence with a positive value of 0.5.

[0158] In Fig.Figure 5 is a schematic representation of a preferred embodiment of a method according to the invention for determining a new fuel composition or for optimizing the influence of a component using the trained combustion property determination model. In this context, the combustion parameter can be understood as a function of a multitude of variables.

[0159] In step S4, the determined influencing factors (as described above) are evaluated, and those variables that significantly affect the combustion parameter are selected. The number of selected variables and / or the corresponding values ​​of the associated influencing factors depend, firstly, on what exactly is to be determined, i.e., whether a new fuel composition, component optimization, or something similar is to be determined. For example, if a new fuel composition is to be determined, all variables relating to engine components or engine settings (component parameters) can be kept constant in the subsequent process. On the other hand, if specific component influences or engine settings are to be optimized, all variables relating to the fuel used (analysis parameter, proportion parameter, fuel parameter) could be kept constant.

[0160] In a further step S5, a local extreme point of the combustion parameter is determined as a function of one of the selected variables. It is possible to determine such an extreme point for each of the selected variables individually and then combine the results. Alternatively, such an extreme point could also be determined as a function of several variables in a single operation. While the latter method requires more computational power, it offers the advantage of taking into account that the individual variables are not completely independent of each other, but rather influence one another. For example, the sum of the proportions always equals 1 (100 vol%). Conversely, the proportion (proportion) or type of a component (analysis parameter) also influences a property of the fuel mixture, such as the vapor pressure (fuel parameter).

[0161] Determining the extreme point involves either identifying a minimum or a maximum combustion parameter. This naturally depends on the nature of the combustion parameter. If the combustion parameter is an exhaust gas composition, such as the proportion of soot particles, then the combustion parameter should be minimized accordingly. Conversely, if the combustion parameter is an efficiency parameter, then it should be maximized accordingly.

[0162] In step S6, a new fuel composition, component influence, engine setting, or similar parameter is determined based on the identified extreme point. In a further step S7, the results obtained can be validated. This can be done either using computer-aided methods or by comparison with real-world measurements. For example, the determined new fuel composition could be measured on a test bench to verify whether the measured combustion parameter corresponds to the determined minimized or maximized combustion parameter.

[0163] The applicant reserves the right to claim all features disclosed in the application documents as essential to the invention, provided they are novel individually or in combination compared to the prior art. It is further noted that the individual figures also describe features which may be advantageous on their own. A person skilled in the art will immediately recognize that a particular feature described in a figure may be advantageous even without incorporating other features from that figure. Furthermore, a person skilled in the art will recognize that advantages may also arise from a combination of several features shown in individual or different figures. Reference symbol list 2 Input size 4 Trained machine learning models 6 Initial size 10 Analysis parameter 12 Proportion size 14 Fuel size 16 Component size 18 Combustion size 20 training data set 22 Untrained combustion property determination model 24 Trained combustion property determination model 26 Trained explainable combustion property determination model 28 influencing factors 30 Graph on LIME approach 32 x-axis 34 y-axis 36 Measured value Variable x0 38 Permuted variables x1 to x6 40 Measured value combustion quantity v0 42 Permuted combustion parameters v1 to v6 44 Compensation function 50 Waterfall Plot SHAP Approach 52 x-axis 54 y-axis 56 Functional value / measured value combustion quantity 58 Expected value 60 Variable y1 62 Variable y2 64 Variable y3 66 Variable y4 68 SHAP value for variable y1 70 SHAP value for variable y2 72 SHAP value for variable y3 74 SHAP value for variable y4 S1 to S7 process steps QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited patent literature

[0000] DE 10 2022 207 017 A1

[0005] US 2023 / 0260342 A1

[0008] KR 2023 0165 495 A

[0009] Cited non-patent literature

[0000] DIN EN ISO 22854 [0037, 0132] ISO / DIS 22854:2020 [0037, 0132] ISO 22854

[0038] Ribeiro et al. (2016

[0078]

Claims

Computer-implemented method for generating a trained combustion property determination model (24) of machine learning for determining at least one combustion parameter (18) of a given fuel using analysis data sets, each of which is characteristic of a composition of a given fuel, wherein each analysis data set for the respective given fuel comprises at least one analysis parameter (10) and a proportion parameter (12) associated with the analysis parameter (10), wherein the analysis parameter (12) is characteristic of a group of compounds from the given fuel which have at least one given functionality, and wherein the proportion parameter (12) is characteristic of a quantity fraction of the compounds present in the respective fuel.which are grouped in the group characterized by the analysis variable (10) comprising the steps: - Providing a trainable combustion property determination machine learning model (22) which includes a set of trainable parameters and which, based on an analysis data set acquired for a given fuel or data derived therefrom as input variable (2), generates at least one combustion variable (18) as output variable (6), wherein the combustion variable (18) is characteristic of at least one combustion property of the fuel; - Generating a training data set (20) comprising a plurality of acquired analysis data sets for training fuels, which are used to generate the training data set, and a plurality of combustion variables (18), wherein the combustion variables are each characteristic of at least one combustion property of the respective training fuel.- Training the combustion property determination machine learning model (22) based on the training dataset; - Determining at least one influencing variable (28) by applying a method to explain the trained combustion property determination machine learning model (24), wherein the at least one influencing variable (28) is characteristic of an influence of the at least one proportion variable (12) and / or the at least one analysis variable (10) on the combustion variable (18). The method according to claim 1, characterized in that at least one analysis data set comprises at least one fuel parameter (14) which is characteristic of at least one fuel property, wherein the at least one fuel property is preferably selected from a group of fuel properties which includes ignition quality, energy content, density, vapor pressure, boiling point, viscosity, knock resistance, cetane number (CN), characteristic values ​​such as points of a distillation curve, yield sooting index (YSI), research octane number (RON), motor octane number (MON), front octane number (FOZ), road octane number (SOZ) and the like, as well as combinations thereof. Method according to at least one of the preceding claims, characterized in that at least one analysis data set comprises at least one component size (16), wherein the component size (16) is characteristic of at least one property of an engine which was used to detect at least one combustion size (18) associated with the at least one analysis data set. Method according to claim 2 or 3, characterized in that an influencing factor (28) is determined for the at least one fuel parameter (14), wherein the influencing factor (28) is characteristic of an influence of the at least one fuel parameter (14) on the combustion parameter (18) and / or an influencing factor (28) is determined for the at least one component parameter (16), wherein the influencing factor (28) is characteristic of an influence of the at least one component parameter (16) on the combustion parameter (18). Method according to at least one of the preceding claims, characterized in that the at least one determined influencing variable (28) is characteristic for a global and / or a local interpretability of the trained combustion property determination model of machine learning (24). Method according to at least one of the preceding claims, characterized in that the training data set comprises a plurality of input variables (2) and a plurality of output variables (6) assigned to the input variables (2), wherein the plurality of input variables (2) each comprises a plurality of variables, wherein the plurality of variables is selected from a group of variables comprising an analysis variable (10), a proportion variable (12), a fuel variable (14) and a component variable (16), and wherein the output variable (6) is a combustion variable (18). The method according to claim 6, characterized in that the determination of the at least one influencing variable (28) by applying a method for explaining the trained combustion property determination model of machine learning (24) comprises at least the steps of: - generating a plurality of permuted input variables based on an input variable (2) by permuting at least one variable of this input variable (2), wherein the permuted input variables differ from each other pairwise by a value of this variable; - determining a plurality of permuted combustion variables (42) based on the permuted input variables by applying the trained combustion property determination model of machine learning (24);- Fitting a simple model to the permuted input variables and to the permuted combustion variables (42), weighting the permuted combustion variables (42) based on their similarity to the original combustion variable (18); - Determining an influence variable (28) for this variable based on the fitted simple model.; The method according to claim 6, characterized in that the determination of the at least one influencing variable (28) by applying a method for explaining the trained combustion property determination model of machine learning (24) comprises at least the steps: - selecting a variable of the input variable (2); - determining a plurality of coalitions, wherein the remaining, in particular unselected, variables of the input variable are taken into account when determining the plurality of coalitions;- Determining a first combustion parameter and a second combustion parameter for each of the plurality of coalitions by applying the trained combustion property determination machine learning model, wherein, in determining the first combustion parameter, the plurality of coalitions are used in combination with the selected variable with its original value as input, and wherein, in determining the second combustion parameter, the plurality of coalitions are used in combination with the selected variable with a modified value as input; - Determining a marginal contribution of the selected variable for each of the plurality of coalitions by determining a difference between the first combustion parameter and the second combustion parameter; - Determining an influence parameter (28) for the selected variable by forming a mean of the plurality of marginal contributions.; Method for the particularly predictive determination of a particularly new fuel composition using the trained combustion property determination model of machine learning (24) according to at least one of the preceding claims comprising the steps of determining a particularly local extremum of a combustion parameter (18) as a function of at least one variable and preferably as a function of a plurality of variables, wherein the at least one variable and preferably the plurality of variables are selected from a group of variables comprising an analysis parameter (10), a proportion parameter (12), a fuel parameter (14) and a component parameter (16);- Determining a new fuel composition based on the determined, in particular local, extreme point, wherein the determination of the new fuel composition includes the determination of a plurality of analytical parameters (10) and the determination of a plurality of proportion parameters (12) assigned to the analytical parameters (10). Device for generating a trained combustion property determination model of machine learning (24) for determining at least one combustion parameter (18) of a given fuel using analysis data sets, each of which is characteristic of a composition of a given fuel, wherein each analysis data set for the respective given fuel comprises at least one analysis parameter (10) and a proportion parameter (12) associated with the analysis parameter (10), wherein the analysis parameter (10) is characteristic of a group of compounds from the given fuel which have at least one given functionality, and wherein the proportion parameter (12) is characteristic of a quantity fraction of the compounds in the respective fuel which are grouped in the group characterized by the analysis parameter, wherein the device is configured toto provide a trainable combustion property determination model of machine learning (22) which comprises a set of trainable parameters and which, based on an analysis data set acquired for a given fuel or data derived therefrom as input (2), generates at least one combustion property (18) as output (6), wherein the combustion property (18) is characteristic of at least one combustion property of the fuel, wherein the device is configured to generate a training data set comprising a plurality of acquired analysis data sets of training fuels, which are used to generate the training data set, and a plurality of combustion properties (18), wherein the combustion properties (18) are each characteristic of at least one combustion property of the respective training fuel, wherein the device is configured toto train the combustion property determination model of machine learning (22) on the basis of the training data set, characterized in that the device is configured to determine at least one influencing variable (28) by applying a method for explaining the trained combustion property determination model of machine learning (26), wherein the at least one influencing variable (28) is characteristic of an influence of the at least one proportion variable (12) and / or the at least one analysis variable (10) on the combustion variable (18). Device for the particularly predictive determination of a new fuel composition using a trained combustion property determination model of machine learning (24), which was trained according to one of the preceding claims 1 to 8, wherein the device is configured to determine a particularly local extreme point of a combustion parameter (18) as a function of at least one variable and preferably as a function of a plurality of variables, wherein the at least one variable and preferably the plurality of variables are selected from a group of variables comprising an analysis parameter (10), a proportion parameter (12), a fuel parameter (14) and a component parameter (16), wherein the device is configured to take into account the determined influencing parameters (28) when determining the particularly local extreme point, wherein the device is configured toto determine a new fuel composition based on the determined, in particular local, extreme point, wherein the determination of the new fuel composition comprises the determination of at least one and preferably a plurality of analytical parameters (10) and the determination of at least one and preferably a plurality of proportion parameters (12) assigned to the analytical parameters (10).

Images