Method for monitoring an electrochemical system, and an electrochemical system

Abstract

The invention relates to a method for monitoring an electrochemical system, in particular a high-temperature electrochemical system, preferably a fuel cell system, wherein in at least one method step a functionality of the electrochemical system is determined on the basis of at least one machine learning process (14, 16). According to the invention, the functionality is determined by a degradation model which is at least partly separated from a reference model of the electrochemical system.

Description

[0001] Description

[0002] Method for monitoring an electrochemical system and an electrochemical system

[0003] State of the art

[0004] A method for monitoring a fuel cell system, in particular a high-temperature fuel cell system, has already been proposed, in which in at least one method step a functionality of the fuel cell system is determined based on at least one machine learning process.

[0005] Disclosure of the invention

[0006] The invention is based on a method for monitoring an electrochemical system, in particular a high-temperature electrochemical system, preferably a fuel cell system, wherein in at least one method step a functionality of the electrochemical system is determined based on at least one machine learning process.

[0007] In particular, an electrochemical system is provided for converting chemical energy into electrical energy or vice versa. An electrochemical system can, for example, be a fuel cell system, which in particular enables the conversion of a fuel, e.g. hydrogen, into electrical energy, or an electrolysis cell system, which in particular enables the conversion of electrical energy into chemical energy, preferably for storage. Accordingly, a high-temperature electrochemical system can, for example, be a high-temperature fuel cell system or a high-temperature electrolysis cell system. In the following, the invention is described for a fuel cell system by way of example, without being limited to this. Thus, the invention can also be provided, for example, for an electrolysis cell system instead of a fuel cell system.

[0008] It is proposed that the functionality be determined using a degradation model that is at least partially separated from a reference model of the fuel cell system. The method is preferably intended to determine, in particular predict, degradation of the fuel cell system. Degradation preferably refers to a, in particular gradual, decline in the functionality of the fuel cell system, for example due to wear, aging processes, deposits, or the like. The functionality is preferably evaluated based on at least one target parameter determined using the reference model and / or the degradation model.The at least one target parameter can be an operating parameter of the fuel cell system, for example, an electrical voltage provided by the fuel cell system, or an abstract key figure of the fuel cell system, which in particular incorporates several operating parameters of the fuel cell system. The at least one target parameter is preferably output to an operator of the fuel cell system and / or to a higher-level control system. Before output, the at least one target parameter can optionally be further processed into an actual state of degradation, a degradation rate, an expected remaining service life of the fuel cell system, a time at which the next maintenance should be performed, or the like.Additionally or alternatively, the at least one target parameter is used by a control or regulation unit of the fuel cell system to control or regulate the fuel cell system, for example to initiate a safety operating mode of the fuel cell system that restricts the scope of performance if functional losses of the fuel cell system exceed a predetermined threshold value.

[0009] The reference model preferably maps at least one reference input parameter to a reference value of the at least one target parameter. The degradation model preferably maps at least one degradation input parameter to a degradation value of the at least one target parameter. The word part "reference" or "degradation" in reference input parameter and degradation input parameter merely represents a naming convention for assigning the respective input parameter to the corresponding model. In principle, any parameter available during operation of the fuel cell system can be used as a reference input parameter and / or as a degradation input parameter. In particular, an available parameter can be used only as a reference input parameter, only as a degradation input parameter, or both as a reference input parameter and as a degradation input parameter.The at least one reference input parameter and / or the at least one degradation input parameter can be detected by at least one sensor of the fuel cell system or determined by a computing unit of the fuel cell system, in particular as a function of a variable detected by the at least one sensor. The at least one reference input parameter and / or the at least one degradation input parameter can, in particular, be physical variables or abstract key figures. The computing unit preferably evaluates the reference model and / or the degradation model to determine the target parameter.

[0010] The reference model and the degradation model preferably complement each other in the description of the fuel cell system, in particular in the determination of the at least one target parameter. The degradation model preferably describes a change, preferably all significant changes, in the at least one target parameter due to a degradation of the fuel cell system. Examples that are used as degradation input parameters or that flow into the at least one degradation input parameter include a cumulative operating time of the fuel cell system, a cumulative current throughput of the fuel cell system, a cumulative number of thermal cycles of the fuel cell system, or another variable that describes or characterizes an aging of the fuel cell system.For example, a combination of a cumulative aging value and an actual value of an operating parameter of the fuel cell system, such as a product of an actual value of a current throughput of the fuel cell system and a cumulative operating time of the fuel cell system, is used as a degradation input parameter and / or is incorporated into the at least one degradation input parameter. For example, a characterization variable of the fuel cell system, which is recorded, for example, after production and before commissioning of the fuel cell system, is used as a degradation input parameter and / or is incorporated into the at least one degradation input parameter.

[0011] The reference model preferably describes a dependency of the target parameter on an actual value of an operating parameter and / or an operating condition of the fuel cell system. Examples of operating parameters that are used as reference input parameters or that flow into the reference input parameters include, for example, an electrical current generated by the fuel cell system, a fuel gas utilization of the fuel cell system, a fuel gas utilization of a fuel cell stack of the fuel cell system, a volume flow of a fuel gas, a volume flow of an oxygen-containing fluid, a characterization variable of the fuel cell system, which is recorded, for example, after production and before commissioning of the fuel cell system, or the like.Examples of operating conditions that are used as reference input parameters or that are included in the reference input parameters are, for example, an ambient temperature, an ambient air pressure or the like.

[0012] The reference model is preferably at least substantially independent of degradation of the fuel cell system. The reference model should preferably be considered substantially independent of degradation if a change in the reference value due to this degradation under otherwise constant operating conditions is at least less than 75%, preferably less than 50%, particularly preferably less than 25%, extremely preferably less than 10%, of a change in the degradation value due to this degradation over the same period. Preferably, the at least one reference input parameter, in particular all reference input parameters, is / are independent of degradation, for example because it / they is / are kept constant by means of a control system.Alternatively, a dependence of the at least one reference input parameter on a degradation within the reference model is at least partially compensated, in particular by a dependence of another reference input parameter on this degradation. Preferably, a dependence of the at least one reference input parameter on a degradation of the fuel cell system, in particular due to compensation and / or for a reason inherent in the reference input parameter, is so small that the reference model is considered essentially independent of this degradation. The reference model can describe the fuel cell system, for example, in a fully functional state, wherein the reference value represents an ideal value and is increasingly corrected by the degradation value as degradation progresses.Alternatively, the reference model describes a permissible minimum functional state of the fuel cell system, with the reference value forming a base value that is supplemented by the degradation value as degradation progresses. Alternatively, the reference model describes the fuel cell system in a state with average functionality, particularly with respect to the expected service life of the fuel cell system.

[0013] In at least one embodiment, the degradation model and the reference model are separated in the sense that a determination of the degradation value is / is carried out independently of a determination of the reference value and / or a determination of the reference value is / is carried out independently of a determination of the degradation value. The reference model and / or the degradation model are / is preferably created by a learning unit using the at least one machine learning process. The learning unit can be the aforementioned computing unit or an external further computing unit, wherein the reference model and / or the degradation model is / are transferred to a memory of the computing unit after the creation of the reference model and / or the degradation model, in particular automatically or manually.Depending on the choice of the target parameter, the specific design of the fuel cell system, and / or the choice of the machine learning process, the learning unit creates the reference model and the degradation model independently of one another, in a unilateral, or mutually dependent relationship. Preferably, the learning unit is designed to separate, in particular as far as possible, a dependency of the at least one target parameter on the at least one reference input parameter and the at least one degradation input parameter. It is conceivable for the reference model and the degradation model to overlap, for example, when using a neural network as the machine learning process, to have one or more common input layers.

[0014] The learning unit preferably creates the reference model and the degradation model by evaluating training data. The training data include a dependency of the target parameter on at least one, in particular a plurality of, input parameters. The training data can be created by simulating the fuel cell system and / or acquired from other fuel cell systems. The learning unit preferably analyzes the training data to identify the at least one input parameter as a reference input parameter or as a degradation input parameter. For example, the learning unit analyzes the training data with regard to a correlation between the at least one input parameter and the at least one target parameter. If the learning unit finds a correlation between the at least one input parameter and a long-term trend of the at least one target parameter, it preferably identifies this input parameter as a degradation input parameter.If the learning unit finds no correlation between the at least one input parameter and a long-term trend of the at least one target parameter, it preferably identifies this input parameter as a reference input parameter. Alternatively, a programmer of the learning unit divides the input parameters into degradation input parameters and reference input parameters. Preferably, the degradation model is trained on the training data from which the at least one reference input parameter has preferably been removed or is not included, hereinafter referred to as degradation training data for short. Preferably, the reference model is trained on the training data from which the at least one degradation parameter has preferably been removed or in which the at least one degradation parameter is not included, hereinafter referred to as reference training data for short.

[0015] The inventive design advantageously allows a degradation and / or a degradation rate of the fuel cell system to be reliably determined. The risk that a trend in the target parameter that is not attributable to degradation and is described in particular by the reference model is incorrectly identified as degradation can be advantageously kept low. In particular, when determining the degradation and / or the degradation rate, an advantageously high level of plausibility can be achieved by interpreting the target parameter. Furthermore, the degradation model can be accessed directly, so that a modeled course of the target parameter can advantageously be obtained directly without short-term fluctuations.Furthermore, an advantageously easy-to-interpret machine learning process can be used for the degradation model, for example linear regression, in particular while at the same time an advantageously complex and / or flexible machine learning process can be used for the reference model.

[0016] It is further proposed that the degradation model be created using the machine learning process, and the reference model be created using a further machine learning process. The machine learning process and the further machine learning process can be of the same type or different types of machine learning processes. The machine learning process and the further machine learning process can be executed by the same learning unit or by different learning units. For example, the degradation model is determined using linear regression or a regularized variant of linear regression, such as lasso regression or ridge regression. It is also conceivable that a higher-dimensional model, in particular a quadratic or cubic model, is used for the degradation model.For example, the reference model is determined using a Gaussian process, a support vector machine, a decision tree, an ensemble variant of the aforementioned methods, in particular using random forest methods or gradient boost methods, a neural network, or the like. The inventive design allows the machine learning processes to be advantageously specifically adapted to the determination of the reference value and the determination of the degradation value of the target parameter.

[0017] Furthermore, it is proposed that the degradation model and the reference model have independent dependencies. The degradation model is preferably, in particular only, dependent on the at least one degradation parameter and in particular independent of the at least one reference parameter. The reference model is preferably, in particular only, dependent on the at least one reference parameter and in particular independent of the at least one degradation parameter. The degradation model and the reference model are preferably determined using a machine learning process that is inherently separable with regard to the input parameters. For example, the degradation model and the reference model are determined jointly from the training data using linear regression or using a regularized variant of linear regression, such as using lasso regression or ridge regression.The degradation model preferably has one, in particular one, coefficient assigned to the at least one degradation input parameter. To determine or predict the degradation value, the computing unit evaluates, for example, the at least one coefficient of the degradation model, in particular while neglecting coefficients of the reference model. Alternatively, the degradation model and the reference model are evaluated jointly by the computing unit to determine the degradation value, wherein the at least one reference input parameter is set to a constant value. The inventive design advantageously allows for simple creation and evaluation of the degradation model and / or the reference model.

[0018] Furthermore, it is proposed that the degradation model and the reference model be determined through alternating iteration steps. The method preferably comprises a reference iteration step. In the reference iteration step, the learning unit preferably creates a reference submodel using the further machine learning process and the reference training data, in which the target parameter is replaced by a current transfer value. In a first pass of the reference iteration step, the transfer value is set equal to the value of the target parameter contained in the training data. In the reference iteration step, the learning unit preferably determines an iteration value of the target parameter using the reference submodel determined in this reference iteration step.In the reference iteration step, the computing unit preferably determines a new transfer value depending on a deviation of the iteration value of the target parameter determined in the reference iteration step from the previous transfer value. The method preferably comprises a degradation iteration step. The new transfer value is transferred to the degradation iteration step as the current transfer value. In the degradation iteration step, the learning unit preferably creates a degradation submodel using the machine learning process and the degradation training data, in which the target parameter is replaced by the current transfer value. In the degradation iteration step, the learning unit preferably determines an iteration value of the target parameter using the degradation submodel determined in this degradation iteration step.In the degradation iteration step, the computing unit preferably determines a new transfer value depending on a deviation of the iteration value of the target parameter determined in the degradation iteration step from the previous transfer value. The new transfer value is preferably transferred to the reference iteration step as the current transfer value. The reference iteration step and the degradation iteration step are preferably repeated until a deviation measure, for example a mean square deviation, of the iteration value from the current transfer parameter is smaller than a predetermined threshold value. To determine or predict the degradation value using the degradation model, or the reference value using the reference model, the computing unit preferably evaluates all degradation submodels or reference submodels created in the corresponding iteration steps, and their results are summed.The inventive design advantageously allows any machine learning processes to be used to create the reference model and the degradation model.

[0019] It is further proposed that the degradation model be determined based on the reference model. Preferably, in a model determination step of the method, the reference model is determined by the learning unit using the training data set. In particular, the reference model is dependent on the at least one reference input parameter and the at least one degradation parameter. Preferably, in a further model determination step of the method, the degradation model is determined by the learning unit. In the further model determination step, the learning unit preferably uses the degradation training data, in which the target parameter is replaced by a pseudo-value of the target parameter, to determine the degradation model.The learning unit preferably determines the pseudo value using the reference model determined in the model determination step, wherein the at least one degradation input parameter is used from the training data and a substitute value is used for the at least one reference input parameter. The substitute value is, for example, a constant value stored in a memory of the learning unit or an average of the values ​​of the at least one reference input parameter contained in the training data. The computing unit preferably determines an actual value of the degradation using the degradation model. The computing unit determines a forecast of the target parameter, for example, using only the reference model. Due to the embodiment according to the invention, the reference model and the degradation model can advantageously be created quickly and advantageously evaluated quickly.

[0020] It is further proposed that the degradation model and the reference model for determining at least one operating parameter of the fuel cell system have a common base value, in particular one different from zero. The degradation model and the reference model are determined, for example, as discussed in the last paragraph. Preferably, the learning unit determines the base value, in particular additionally. Preferably, the computing unit determines the base value as the target parameter taken from the training data, less the degradation value determined from the training data using the degradation model and less the reference value determined from the training data using the reference model. When determining the reference value, the learning unit preferably inserts a constant value, in particular zero, into the reference model instead of the degradation input parameter of the training data.Preferably, the learning unit determines the base value for several, in particular all, data sets of the training data and calculates an average from the obtained base values, which is stored as a base value in a memory of the computing unit. Preferably, the computing unit adds the stored base value, a forecast of the degradation value created using the degradation model, and a forecast of the reference value created using the reference model to a forecast of the target parameter, wherein the at least one degradation input parameter is kept constant in the reference model.

[0021] Due to the design according to the invention, the reference model and the degradation model can advantageously be created quickly and advantageously evaluated quickly, whereby the degradation model can additionally be validated.

[0022] It is further proposed that the degradation model be extracted from a higher-level model in order to determine the reference model. The reference model is preferably determined as the remaining part of the higher-level model after the degradation model has been extracted from the higher-level model. In the model determination step of the method, the higher-level model is preferably determined by the learning unit using the training data set. In particular, the higher-level model is dependent on the at least one reference input parameter and the at least one degradation input parameter. In the further model determination step, the learning unit preferably uses the degradation training data, in which the target parameter is replaced by the pseudo-value of the target parameter, to determine the degradation model.The learning unit preferably determines the pseudo value using the higher-level model determined in the model determination step, wherein the at least one degradation input parameter from the training data is used, and the substitute value is used for the at least one reference input parameter. The reference model is preferably determined using the reference training data, in which the target parameter is determined by a difference between the target parameter contained in the reference training data and the degradation value that the learning unit determines using the degradation model and the degradation training data. The inventive design allows the reference model and the degradation model to be advantageously quickly created and advantageously quickly evaluated, wherein the degradation model can additionally be validated.

[0023] Furthermore, an electrochemical system, in particular a high-temperature electrochemical system, preferably a fuel cell system, with at least one computing unit for carrying out a method according to the invention is proposed.

[0024] Here, too, the invention is described below using a fuel cell system as an example, but is not limited to this. Thus, the invention may also relate, for example, to an electrolysis cell system instead of a fuel cell system.

[0025] A "computing unit" or a "learning unit" should be understood in particular as a unit with an information input, an information processing unit, and an information output. The computing unit or learning unit advantageously has at least one processor, a memory, input and output means, further electrical components, an operating program, control routines, control routines, and / or calculation routines. The components of the computing unit or the learning unit are preferably arranged on a common circuit board and / or advantageously arranged in a common housing. The computing unit or learning unit is preferably specifically configured and / or specifically programmed to carry out the method according to the invention. The fuel cell system preferably comprises at least one fuel cell device. The fuel cell device preferably comprises at least one fuel cell, in particular at least one fuel cell stack.The at least one fuel cell is preferably designed as a high-temperature fuel cell, for example as a solid oxide fuel cell (SOFC) or as a molten carbonate fuel cell (MCFC). The at least one fuel cell is preferably intended to convert a fuel gas, in particular hydrogen and / or at least one hydrocarbon, with the addition of oxygen in order to generate an electrical current. The fuel cell device preferably comprises components for handling and / or processing the fuel gas and / or an oxygen-containing fluid, in particular ambient air. Examples of these components are at least one fan, at least one reformer, at least one afterburner, at least one heat exchanger, at least one exhaust gas recirculation system, and / or other components deemed appropriate by a person skilled in the art.The fuel cell system preferably comprises at least one sensor arranged in the fuel cell device to detect at least one operating parameter of the fuel cell system, in particular of the fuel cell device. The fuel cell system preferably comprises at least one output unit, for example a display, indicator lights, a loudspeaker, or the like, to output the target parameter and / or a variable derived from the target parameter. The output unit, the computing unit, and / or the learning unit can be arranged on the fuel cell device, in particular within a housing or on a housing of the fuel cell device, or can be formed separately from the fuel cell device and communicate, for example, with the fuel cell device, in particular the sensor, and / or with each other via a data network, in particular via the Internet or via a private data network.The design according to the invention makes it possible to provide a fuel cell system whose degradation can advantageously be reliably monitored.

[0026] The method according to the invention and / or the fuel cell system according to the invention should not be limited to the application and embodiment described above. In particular, the method according to the invention and / or the fuel cell system according to the invention can have a number of individual elements, components, units, and method steps that differs from the number stated herein to fulfill a functionality described herein. Furthermore, in the value ranges specified in this disclosure, values ​​within the stated limits are also to be considered disclosed and can be used arbitrarily.

[0027] Drawings

[0028] Further advantages will become apparent from the following description of the drawings. The drawings illustrate an exemplary embodiment of the invention. The drawings, the description, and the claims contain numerous features in combination. Those skilled in the art will also expediently consider the features individually and combine them into useful further combinations.

[0029] They show:

[0030] Fig. 1 is a schematic representation of a fuel cell system according to the invention and

[0031] Fig. 2 a schematic flow diagram of a method according to the invention, description of the embodiment

[0032] Figure 1 shows a fuel cell system 12. The fuel cell system 12 preferably comprises a fuel cell device 20 for generating an electric current by electrochemically converting a fuel gas with oxygen. The fuel cell system 12 preferably comprises at least one sensor 22 for detecting at least one operating parameter of the fuel cell device 20. The fuel cell system 12 comprises a computing unit 18 for carrying out a method 10, which is explained in more detail in Figure 2. The computing unit 18 is provided for determining a target parameter, in particular as a function of the at least one operating parameter detected by the at least one sensor 22. A degradation model and a reference model, by means of which the computing unit 18 determines the target parameter, are preferably stored in a memory of the computing unit 18.The degradation model and / or the reference model can be created by the computing unit 18 or by an external learning unit 26. In the following, "learning unit" can refer to the computing unit 18 or the external learning unit 26. The computing unit 18 is preferably provided to interpret the target parameter with respect to a degradation of the fuel cell device 20. The fuel cell system 12 preferably comprises an output unit 24 for outputting the target parameter and / or interpreting the target parameter with respect to a degradation of the fuel cell device 20.

[0033] Figure 2 shows the method 10 for monitoring the fuel cell system 12, in particular the fuel cell device 20. The method 10 preferably comprises a machine learning process 14. The method 10 preferably comprises a further machine learning process 16. The method 10 preferably comprises an evaluation step 34. In the evaluation step 34 of the method 10, a functionality of the fuel cell system 12 is determined by the computing unit 18 based on at least one of the machine learning processes 14, 16. The functionality is determined by the computing unit 18 using the degradation model that is at least partially separated from the reference model of the fuel cell system 12.The degradation model is created by the learning unit using the machine learning process 14, and the reference model is created by the further machine learning process 16. Preferably, the computing unit 18 determines a degradation value of a target parameter using the degradation model. Preferably, the computing unit 18 determines a reference value of the target parameter using the reference model.

[0034] The method 10 preferably comprises a data acquisition step 28. In the data acquisition step 28, training data is acquired, by means of which the machine learning processes 14, 16 are created. The training data preferably comprise data sets consisting of input parameters and values ​​of the target parameter. The input parameters are preferably equal to the operating parameters detectable with the sensor 22 or a variable or characteristic derived therefrom. The input parameters can be selected, in particular, as any desired combination of operating parameters during feature engineering. Particularly preferably, the target parameter is equal to an electrical voltage linked to the electrical current generated by the fuel cell device 20. The training data can be created by simulating the fuel cell device 20 or by recording data from other fuel cell devices.

[0035] The method 10 preferably comprises a preprocessing step 30 for preprocessing the training data. The training data is preferably filtered, in particular with regard to an operating time of the fuel cell device 20. For example, only data sets of the training data that were acquired or created within a predefined nominal load range of the fuel cell device 20 are further used. For example, only data sets of the training data are further used in which the input parameters fluctuate by less than a predefined tolerance value and which, in particular, describe a steady state of the fuel cell device 20. Optionally, a preselection is made of the types of fuel cell devices 20 to be considered, which are to be described by the degradation model and the reference model.

[0036] The learning unit or a programmer of the learning unit assigns the input parameters, depending on the influence of the input parameters on the target parameter, preferably to the reference model as a reference input parameter, particularly in the case of a relatively low influence, or to the degradation model as a degradation input parameter, particularly in the case of a relatively high influence. Preferably, the computing unit 18 creates reference training data from the training data, which includes the reference input parameters and the target parameter, and degradation training data, which includes the degradation input parameters and the target parameter. The learning unit executes the machine learning process 14 to create the degradation model, preferably using the degradation training data set. The learning unit executes the further machine learning process 16 to determine the reference model, preferably using the reference training data set.

[0037] The method 10 optionally includes a validation step 32. In the validation step 32, a value of the target parameter is predicted using the determined degradation model and / or the determined reference model and compared with test data. If the test data and the predicted value sufficiently match, the degradation model and the reference model are stored in a memory of the computing unit 18. If the test data and the predicted value do not sufficiently match, the machine learning process 14 and the further machine learning process 16 are repeated, preferably based on a different machine learning model. Alternatively, several machine learning models known to the learning unit are executed one after the other, and in the validation step 32, it is determined which machine learning model was used to create the most precise degradation model and the most precise reference model.When the machine learning process 14 and the further machine learning process 16 are executed using linear regression or another machine learning model that is inherently separable with respect to the input parameters, the degradation model and the reference model exhibit mutually independent dependencies on these input parameters. When the machine learning process 14 and the further machine learning process 16 are executed, the degradation model and the reference model are determined through alternating iteration steps. When the machine learning process 14 and the further machine learning process 16 are executed, the degradation model is determined based on the reference model.When the machine learning process 14 and the further machine learning process 16 are executed, the learning unit determines a base value common to the degradation model and the reference model, in particular a base value different from zero, which the computing unit 18 uses to determine the target parameter. When the machine learning process 14 and the further machine learning process 16 are executed, the learning unit extracts the degradation model from a higher-level model to determine the reference model.

[0038] In the evaluation step 34, the computing unit 18 evaluates the reference model and / or the degradation model based on the operating parameters determined with the at least one sensor 22 in order to obtain an actual value of the target parameter or a forecast of the target parameter. Preferably, the computing unit 18 interprets the target parameter with respect to a degradation and / or a degradation rate of the fuel cell device 20. The method 10 preferably comprises an output step 36 in which the output unit 24 outputs the target parameter and / or its interpretation with respect to a degradation and / or a degradation rate.

[0039] In an alternative exemplary embodiment, which is not shown in detail, an electrolysis cell system could be designed analogously to the fuel cell system 12 shown in Fig. 1. The electrolysis cell system 12 could accordingly preferably comprise an electrolysis cell device for generating hydrogen by electrochemical conversion of water with the assistance of electrical energy. The electrolysis cell system could also accordingly comprise at least one sensor for detecting at least one operating parameter of the electrolysis cell device. The electrolysis cell system could also comprise a computing unit for carrying out a method analogous to Fig. 2. The computing unit could accordingly be provided to determine a target parameter, in particular as a function of the at least one operating parameter detected by the at least one sensor.A degradation model and a reference model, by means of which the computing unit determines the target parameter, could preferably be stored in a memory of the computing unit 18. The degradation model and / or the reference model can be created by the computing unit or by an external learning unit. In the following, "learning unit" could refer to the computing unit or the external learning unit. The computing unit could preferably be provided to interpret the target parameter with respect to a degradation of the electrolysis cell device. Accordingly, the electrolysis cell system could also preferably comprise an output unit for outputting the target parameter and / or interpreting the target parameter with respect to a degradation of the fuel cell device.

[0040] Accordingly, according to the alternative exemplary embodiment, a method for monitoring the electrolysis cell system, in particular the electrolysis cell device, could be carried out analogously to Fig. 2. The method could preferably comprise a machine learning process. The method could also preferably comprise a further machine learning process 16. The method could preferably comprise an evaluation step. In the evaluation step of the method, a functionality of the electrolysis cell system could be determined by the computing unit based on at least one of the machine learning processes. The functionality could be determined by the computing unit using the degradation model that is at least partially separated from the reference model of the electrolysis cell system. The degradation model could be created by the learning unit using the machine learning process 14 and the reference model using the further machine learning process.Preferably, the computing unit determines a degradation value of a target parameter using the degradation model. Preferably, the computing unit determines a reference value of the target parameter using the reference model.

[0041] For the alternative embodiment, the method could preferably also include a data acquisition step. In the data acquisition step, training data could be acquired, which is used to create the machine learning processes. The training data could preferably comprise data sets consisting of input parameters and values ​​of the target parameter. The input parameters could preferably be the same as the operating parameters detectable with the sensor or a variable or characteristic derived therefrom. The input parameters can be selected, in particular, as any desired combination of operating parameters during feature engineering. Particularly preferably, the target parameter would be equal to the amount of hydrogen generated by the electrolysis cell device. The training data can be created by simulating the electrolysis cell device or by recording data from other electrolysis cell devices.

[0042] Within the scope of the alternative embodiment, the method could also preferably comprise a preprocessing step for preprocessing the training data. The training data is preferably filtered, in particular with regard to an operating time of the electrolysis cell device. For example, only data sets of the training data that were acquired or created within a predefined nominal load range of the electrolysis cell device are further used. For example, only data sets of the training data are further used in which the input parameters fluctuate by less than a predefined tolerance value and which, in particular, describe a steady state of the electrolysis cell device. Optionally, a preselection is made of the types of electrolysis cell devices to be considered, which are to be described by the degradation model and the reference model.

[0043] Within the scope of the alternative embodiment, the learning unit or a programmer of the learning unit could assign the input parameters, depending on the influence of the input parameters on the target parameter, preferably to the reference model as a reference input parameter, particularly in the case of a relatively low influence, or to the degradation model as a degradation input parameter, particularly in the case of a relatively high influence. Preferably, the computing unit could create reference training data from the training data, which includes the reference input parameters and the target parameter, and degradation training data, which includes the degradation input parameters and the target parameter. The learning unit could preferably execute the machine learning process for creating the degradation model using the degradation training data set.The learning unit could preferably perform the further machine learning process to determine the reference model using the reference training dataset.

[0044] Within the scope of the alternative embodiment, the method could also optionally include a validation step. In the validation step, a value of the target parameter could be predicted using the determined degradation model and / or the determined reference model and compared with test data. If the test data and the predicted value sufficiently match, the degradation model and the reference model could be stored in a memory of the computing unit 18. If the test data and the predicted value do not sufficiently match, the machine learning process and the further machine learning process could be repeated, preferably based on a different machine learning model.Alternatively, several machine learning models known to the learning unit could be executed sequentially, and the validation step could determine which machine learning model generated the most precise degradation model and the most precise reference model. If the machine learning process and the subsequent machine learning process are executed using linear regression or another machine learning model that is inherently separable with respect to the input parameters, the degradation model and the reference model could exhibit independent dependencies on these input parameters. If the machine learning process and the subsequent machine learning process are executed, the degradation model and the reference model could be determined through alternating iteration steps.When the machine learning process and the further machine learning process are executed, the degradation model could be determined based on the reference model. When the machine learning process and the further machine learning process are executed, the learning unit could determine a base value common to the degradation model and the reference model, in particular a base value different from zero, which the computing unit could use to determine the target parameter. When the machine learning process and the further machine learning process are executed, the learning unit could extract the degradation model from a higher-level model to determine the reference model.

[0045] Within the scope of the alternative embodiment, the computing unit could also evaluate the reference model and / or the degradation model in the evaluation step 34 based on the operating parameters determined with the at least one sensor 22 in order to obtain an actual value of the target parameter or a forecast of the target parameter. Preferably, the computing unit could interpret the target parameter with respect to a degradation and / or a degradation rate of the electrolysis cell device. The method could also preferably include an output step in which the output unit outputs the target parameter and / or its interpretation with respect to a degradation and / or a degradation rate.

Claims

Claims 1. Method for monitoring an electrochemical system, in particular a high-temperature electrochemical system, preferably a fuel cell system, wherein in at least one method step a functionality of the electrochemical system is determined based on at least one machine learning process (14, 16), characterized in that the functionality is determined by means of a degradation model that is at least partially separated from a reference model of the electrochemical system.

2. Method according to claim 1, characterized in that the degradation model is created using the machine learning process (14) and the reference model is created using a further machine learning process (16).

3. Method according to one of the preceding claims, characterized in that the degradation model and the reference model have independent dependencies on each other.

4. Method according to one of the preceding claims, characterized in that the degradation model and the reference model are determined by alternating iteration steps.

5. Method according to one of the preceding claims, characterized in that the degradation model is determined based on the reference model.

6. Method according to one of the preceding claims, characterized in that the degradation model and the reference model are used to determine at least one operating parameter of the electrochemical Systems have a common, in particular non-zero, base value. A method according to one of the preceding claims, characterized in that the degradation model is extracted from a higher-level model in order to determine the reference model. An electrochemical system, in particular a high-temperature electrochemical system, preferably a fuel cell system, with at least one computing unit (18) for carrying out a method according to one of the preceding claims.

Images