Diagnosis method for determining a state of a fuel cell system
The diagnostic method employs machine learning to assess nitrogen permeability in fuel cells by retraining models over time, optimizing purge processes and extending fuel cell stack service life by adapting to changing conditions.
Patent Information
- Application Number
- PCT/EP2024/062524
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-05-19
- Filing Date
- 2024-05-07
- Publication Date
- 2024-11-28
AI Technical Summary
Fuel cells exhibit increasing nitrogen permeability over time, leading to higher hydrogen consumption and necessitating more frequent purging processes, which is not effectively managed by existing technologies, impacting maintenance and service life.
A diagnostic method using machine learning to determine nitrogen permeability by training a model initially and retraining it over time to adjust purge processes based on changing gas concentrations, allowing for adaptive control of purge intervals and membrane protection.
This method enables accurate assessment of fuel cell state and nitrogen permeability, optimizing purge processes and extending fuel cell stack service life by adapting to changing conditions.
Smart Images

Figure EP2024062524_28112024_PF_FP_ABST
Abstract
Description
[0001] Description
[0002] title
[0003] Diagnostic method for determining a condition of a
[0004] The presented invention relates to a diagnostic method for determining a state of a fuel cell system, a fuel cell system and a program product according to the appended claims.
[0005] State of the art
[0006] Nitrogen permeability, i.e., the ability of fuel cells and their membranes to pass through nitrogen, changes over time. In particular, older fuel cells exhibit higher nitrogen permeability than newer or newer fuel cells.
[0007] If the nitrogen permeability of a fuel cell is too high, more thorough purging processes—for example, longer or more frequent—are required. This leads to significantly higher hydrogen consumption compared to a new condition.
[0008] Monitoring the nitrogen permeability of respective fuel cells is therefore advantageous in order to plan a possible necessary replacement of a fuel cell stack or its maintenance.
[0009] Furthermore, if the nitrogen permeability is known, condition-dependent measures can be taken to protect the membranes and maximize the lifetime of a particular fuel cell stack.
[0010] Disclosure of the invention Features and details which are described in connection with the diagnostic method according to the invention naturally also apply in connection with the fuel cell system according to the invention or the program product according to the invention and vice versa, so that with regard to the disclosure of the individual aspects of the invention, reference is or can always be made to each other.
[0011] The invention presented serves in particular to determine a change in nitrogen permeability of fuel cells of a fuel cell system.
[0012] Thus, according to a first aspect of the invention presented, a diagnostic method for determining a state of a fuel cell system is presented.The diagnostic method comprises training a machine learner at a first point in time such that it determines at least one output gas concentration in an anode subsystem of the fuel cell system based on at least one input gas concentration measured in a measuring line of a fuel cell system, retraining the machine learner at a second point in time such that it determines at least one output gas concentration in the anode subsystem of the fuel cell system based on at least one input gas concentration measured in the measuring line of the fuel cell system, and determining a nitrogen permeability of fuel cells of the fuel cell system based on a change in the machine learner between the first point in time and the second point in time.
[0013] In the context of the invention presented, a machine learner is understood to be a mathematical model that assigns an output concentration to a measured input concentration, i.e., an input concentration determined by a sensor. A machine learner can, for example, comprise an artificial neural network and / or a Gaussian process model, in particular with a NARX structure. In particular, the machine learner can be used to adjust a purge process for purging the fuel cell system, i.e., to adapt purge times or purge intervals to a respective specific output concentration. Accordingly, the machine learner can provide a "virtual mass spectrometer" that determines a gas concentration, in particular a nitrogen concentration and / or a hydrogen concentration in the anode subsystem of the fuel cell system.
[0014] The presented diagnostic method is based on multiple or repeated training of the machine learner provided according to the invention. For this purpose, the machine learner is trained at a first point in time, such as the time of delivery, and later, at a second point in time, such as during maintenance or inspection, it is trained or retrained.
[0015] Since the physical properties, in particular the nitrogen permeability of fuel cells of the fuel cell system, change over time, retraining at the second time leads to a change in the machine learner or a mathematical model underlying the machine learner. Accordingly, the machine learner trained at the second time or second machine learner assigns different values or different output concentrations to the respective input gas concentrations than the machine learner trained at the first time or first machine learner. Accordingly, the first machine learner assigns first output gas concentrations to the respective input gas concentrations, and the second machine learner assigns second output gas concentrations to the respective input gas concentrations.
[0016] The second machine learner is a modified version of the first machine learner. For this purpose, the first machine learner can be modified, or a new machine learner based on the first machine learner or a copy of the first machine learner can be used and adapted accordingly through training at the second time. Accordingly, it can be provided that the first machine learner and the second machine learner are operated in parallel, for example, to determine a difference between the outputs of the respective machine learners.
[0017] The presented invention is based on the principle that the change of the machine learner from the first time to the second time or the change between the first machine learner and the second machine learner is assigned a nitrogen permeability, in particular a characteristic value that quantifies a nitrogen permeability.
[0018] It can be provided that the change in the machine learner is determined based on a deviation between a first initial concentration determined by the machine learner trained at the first time and a second initial concentration determined by the machine learner trained at the second time.
[0019] A difference in the state of the fuel cell system between the first point in time and the second point in time results in a difference between the first output gas concentrations determined by the first machine learner and the second output gas concentrations determined by the second machine learner. Accordingly, the state of the fuel cell system, in particular a nitrogen permeability, can be deduced from the difference between the first output gas concentrations and the second output gas concentrations. For this purpose, for example, an assignment scheme can be used that assigns a nitrogen permeability value to each difference between the first output gas concentrations and the second output gas concentrations. This assignment scheme can, for example, be another machine learner that was trained using various training fuel cell systems that exhibit different nitrogen permeabilities.The training fuel cell systems can of course be real or simulated.
[0020] It can further be provided that a control parameter for a purge valve of the fuel cell system is selected as a function of a respective initial concentration determined by the machine learner, and the change in the machine learner is determined on the basis of a deviation between a first control parameter, which was determined as a function of the machine learner trained at the first time, and a second control parameter, which was determined as a function of the machine learner trained at the second time.
[0021] By indirectly determining a change in the machine learner and, consequently, the state of the fuel cell system via a control parameter selected as a function of the machine learner's initial gas concentration, a control loop implemented in the fuel cell system can be used to infer the nitrogen permeability of the fuel cells of the fuel cell system. For this purpose, for example, a time between two purge processes or a purge interval can be evaluated as a control parameter. Accordingly, a characteristic value for quantifying the nitrogen permeability can be directly assigned to a change in the control parameter after the second time point. For this purpose, for example,An assignment scheme can be used that assigns a nitrogen permeability value to a respective difference between a first control parameter determined as a function of the first machine learner and a second control parameter determined as a function of the second machine learner. This assignment scheme can, for example, be another machine learner that was trained using various training fuel cell systems exhibiting different nitrogen permeabilities. The training fuel cell systems can, of course, be real or simulated.
[0022] It can further be provided that the change in the machine learner is determined based on a deviation between a first model parameter of a mathematical model of the machine learner trained at the first point in time and a second model parameter of a mathematical model of the machine learner trained at the second point in time.
[0023] A change in the machine learner from the first point in time to the second point in time requires, in particular, a change in model parameters, such as a weighting of regression parameters of a Gaussian process, of the machine learner. Accordingly, a characteristic value for quantifying the nitrogen permeability can be assigned based on a difference between the respective model parameters of the first machine learner and the respective model parameters of the second machine learner. For this purpose, for example, an assignment scheme can be used that assigns a nitrogen permeability value to a respective difference between a first model parameter of the first machine learner and a second model parameter of the second machine learner. This assignment scheme can, for example, be another machine learner that was trained using various training fuel cell systems that exhibit different nitrogen permeabilities.The training fuel cell systems can of course be real or simulated.
[0024] It can further be provided that the machine learner is trained at least during the second time point based on a ground truth which is determined by means of a gas concentration sensor arranged on the anode subsystem of the fuel cell system.
[0025] A gas concentration sensor arranged on the anode subsystem of the fuel cell system, such as a hydrogen concentration sensor and / or a nitrogen concentration sensor, determines real values that represent actual physical conditions in a respective fuel cell system, ie, a respective nitrogen permeability.
[0026] Accordingly, training the machine learner with a ground truth that represents or encompasses the actual physical conditions requires the machine learner to adapt to the actual physical conditions, so that it assigns input gas parameters to output gas parameters, taking the actual physical conditions into account, and based on this assignment, the actual physical conditions can be inferred. In particular, it can be provided that the output concentration includes a concentration of hydrogen and / or nitrogen.
[0027] It may further be provided that the fuel cell system is adjusted depending on the determined nitrogen permeability.
[0028] By adjusting the fuel cell system depending on the specific nitrogen permeability, the fuel cell system can be adapted to the current nitrogen permeability.
[0029] It may further be provided that the determined nitrogen permeability is displayed on a display and / or stored in a memory.
[0030] Based on the nitrogen permeability displayed on a display, which can be presented, for example, as a quantitative value or as a traffic light system, the status of the fuel cell system can be presented to the user. For example, a recommended maintenance interval can be determined dynamically based on the nitrogen permeability.
[0031] Using a nitrogen permeability stored in a memory, subsequent processes, such as a control of the fuel cell system, can access the nitrogen permeability or its value and adapt dynamically to the nitrogen permeability accordingly.
[0032] It may further be provided that the machine learner is trained at the first point in time using a training fuel cell system.
[0033] Using a training fuel cell system, such as an identical reference system or a simulated fuel cell system, a delivery state of a fuel cell system can be validly mapped or simulated, thus eliminating the need to operate a respective fuel cell system in its delivery state for training purposes. According to a second aspect, the presented invention relates to a fuel cell system for converting energy.
[0034] The proposed fuel cell system comprises a fuel cell stack comprising an anode subsystem and a cathode subsystem, an exhaust line, a purge line, a measuring line arranged downstream of a connection between the purge line (409) and the exhaust line (407) in the flow direction, and a computing unit. The computing unit is configured to execute a possible embodiment of the proposed diagnostic method.
[0035] According to a third aspect, the presented invention relates to a program product comprising instructions which, when executed by a computing unit, configure the computing unit to carry out a possible embodiment of the presented diagnostic method.
[0036] Advantages that have been described in detail for the diagnostic method for determining a state of a fuel cell system according to the first aspect of the invention apply equally to the fuel cell system for converting energy according to the second aspect of the invention and to the program product according to the third aspect of the invention.
[0037] Further advantages, features, and details of the invention will become apparent from the following description, which describes exemplary embodiments of the invention in detail with reference to the drawings. The features mentioned in the claims and in the description may be essential to the invention individually or in any combination.
[0038] They show:
[0039] Figure 1 shows a possible design of the presented diagnostic procedure,
[0040] Figure 2 shows a change in output gas concentration values caused by a change in a machine learner through retraining, Figure 3 shows a change in control parameters of a purge process caused by a change in a machine learner through retraining,
[0041] Figure 4 shows a possible design of the presented fuel cell system,
[0042] Figure 1 shows a diagnostic method 100 for determining a state of a fuel cell system.
[0043] The diagnostic method 100 comprises a first training step 101 in which a machine learner is trained at a first point in time such that it determines at least one output gas concentration in an anode subsystem of the fuel cell system based on at least one input gas concentration measured in a measuring line of a fuel cell system, such as a training fuel cell system.
[0044] Furthermore, the diagnostic method 100 comprises a second training step 103, in which machine learners previously trained in the first training step 101 are trained at a second time such that they determine at least one output gas concentration in the anode subsystem of the fuel cell system based on at least one input gas concentration measured in the measuring line of a respective fuel cell system.
[0045] Furthermore, the diagnostic method 100 comprises a determination step 105 in which a nitrogen permeability of fuel cells of the fuel cell system is determined based on a change in the machine learner between the first time point and the second time point.
[0046] Figure 2 shows a diagram 200 that extends over time on its abscissa and over an initial gas concentration of, for example, nitrogen or hydrogen on its ordinate. A first curve 201 corresponds to initial gas concentration values determined by a machine learner that was trained at a first point in time.
[0047] A second curve 203 corresponds to initial gas concentration values determined by the machine learner, which was trained at a second time point.
[0048] Based on a deviation between the first curve 201 and the second curve 203, as indicated by arrow 205, a change in the fuel cell system, in particular a nitrogen permeability of the respective fuel cells of the fuel cell system, can be concluded.
[0049] Figure 3 shows a diagram 300 which extends on its abscissa over time and on its ordinate over an activity of a purge valve of a fuel cell system.
[0050] A first curve 301 corresponds to an activity of the purge valve, which was set or regulated as a function of a machine learner trained at a first point in time.
[0051] A second curve 303 corresponds to an activity of the purge valve, which was set or regulated as a function of the machine learner trained at a second point in time.
[0052] Based on a deviation between the first curve 301 and the second curve 303, as indicated by arrow 305, a change in the fuel cell system, in particular a nitrogen permeability of the respective fuel cells of the fuel cell system, can be concluded.
[0053] Figure 4 shows a fuel cell system 400 for converting energy. The fuel cell system 400 comprises a fuel cell stack 401, which includes an anode subsystem 403 and a cathode subsystem 405, an exhaust line 407, a purge line 409, a measurement line 411 arranged downstream of a connection between the purge line (409) and the exhaust line (407), and a computing unit 413. The computing unit 413 is configured to execute the diagnostic method 100 according to Figure 1.
Claims
Claims 1 . Diagnostic method (100) for determining a state of a fuel cell system (400), the diagnostic method (100) comprising: Training (101) a machine learner at a first time such that it determines at least one output gas concentration in an anode subsystem of the fuel cell system based on at least one input gas concentration measured in a measuring line of a fuel cell system, Retraining (103) the machine learner at a second time such that it determines at least one output gas concentration in the anode subsystem (403) of the fuel cell system (400) based on at least one input gas concentration measured in the measuring line (411) of the fuel cell system (400), Determining a nitrogen permeability (105) of fuel cells of the fuel cell system (400) based on a change in the machine learner between the first time point and the second time point.
2. Diagnostic method (100) according to claim 1, characterized in that the change in the machine learner is determined based on a deviation between a first initial concentration determined by the machine learner trained at the first time and a second initial concentration determined by the machine learner trained at the second time.
3. Diagnostic method (100) according to claim 1, characterized in that that a control parameter for a purge valve of the fuel cell system (400) is selected as a function of a respective initial concentration determined by the machine learner, and the change in the machine learner is determined based on a deviation between a first control parameter, which was determined as a function of the machine learner trained at the first time, and a second control parameter, which was determined as a function of the machine learner trained at the second time.
4. Diagnostic method (100) according to claim 1, characterized in that the change in the machine learner is determined based on a deviation between a first model parameter of a mathematical model of the machine learner trained at the first time and a second model parameter of a mathematical model of the machine learner trained at the second time.
5. Diagnostic method (100) according to one of the preceding claims, characterized in that the machine learner is trained at least during the second time point on the basis of a ground truth which is determined by means of a gas concentration sensor arranged on the anode subsystem (403) of the fuel cell system (400).
6. Diagnostic method (100) according to one of the preceding claims, characterized in that the initial concentration comprises a concentration of hydrogen and / or nitrogen.
7. Diagnostic method (100) according to one of the preceding claims, characterized in that the fuel cell system (400) is adjusted depending on the determined nitrogen permeability.
8. Diagnostic method (100) according to one of the preceding claims, characterized in that the determined nitrogen permeability is output on a display and / or stored in a memory.
9. Diagnostic method (100) according to one of the preceding claims, characterized in that the machine learner is trained at the first time using a training fuel cell system.
10. A fuel cell system (400) for converting energy, the fuel cell system (400) comprising: a fuel cell stack (401) comprising an anode subsystem (403) and a cathode subsystem (405), an exhaust gas line (407), a purge line (409), a measuring line (411) arranged downstream of a connection between the purge line (409) and the exhaust gas line (407), a computing unit (413), the computing unit (413) being configured to carry out a diagnostic method (100) according to any one of claims 1 to 9.
11. A program product, wherein the program product comprises instructions which, when executed by a computing unit (413), configure the computing unit (413) to execute a diagnostic method (100) according to one of claims 1 to 9.
Citation Information
Patent Citations
Fuel cell controlling and regulating method for electrical vehicles, involves determining one or multi dimensional connections between dynamic input variable and dynamic output variable in training algorithm
DE102007026314A1
Method for operational monitoring of at least one fuel cell device and fuel cell device
DE102020202881A1
Method and apparatus for operating a technical system
DE102021210393A1
