Manufacturing method and manufacturing system for manufacturing fuel cell stack

A machine learning-based method for fuel cell production accurately classifies cells using unsupervised learning, reducing incorrect identifications and extending the service life of fuel cell stacks by identifying faulty cells through real-time voltage profiling.

CN119096384APending Publication Date: 2024-12-06ROBERT BOSCH GMBH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202380034096.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2022-04-14
Filing Date
2023-04-03
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

Existing fuel cell production methods rely on predefined check limits that often incorrectly identify defective cells, leading to inefficiencies and reduced service life due to premature aging.

Method used

A machine learning-based method that evaluates voltage profiles during discharge to automatically classify fuel cells as non-faulty or faulty, eliminating the need for predefined limits and improving accuracy by using unsupervised learning models trained on real-time data.

Benefits of technology

Minimizes incorrect identification of defective cells, enhances production efficiency by accurately distinguishing between faulty and non-faulty cells, and extends the service life of fuel cell stacks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119096384A_ABST
    Figure CN119096384A_ABST
Patent Text Reader

Abstract

The invention relates to a manufacturing method (100) for manufacturing a fuel cell stack. The manufacturing method (100) comprises: manufacturing (101) a plurality of fuel cells on a pipeline (201) by a plurality of manufacturing steps; ascertaining (103) a voltage value of a voltage of a corresponding fuel cell of the plurality of fuel cells from a start time point, at which the fuel supply to the fuel cells is interrupted, to an end time point; assigning (105) the voltage values by means of a machine learning device to a first group indicating a fault-free state of the fuel cells or to a second group indicating a fault state of the fuel cells, screening (107) the voltage values of which are assigned by the machine learning device to the fuel cells of the second group; and assembling (109) only the fuel cells whose voltage values are assigned to the first group by the machine learner into a fuel cell stack.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a manufacturing method and a manufacturing system for manufacturing a fuel cell stack. Background Art

[0002] According to the prior art, during production, fuel cells are checked by means of individual checks specific to the respective production point, in that the determined measured values ​​are respectively compared with a predefined check limit. Fuel cells whose measured values ​​deviate too greatly from the check limit are screened out. Summary of the invention

[0003] Within the framework of the present invention, a production method and a production system for producing a fuel cell stack are proposed. Further features and details of the invention are derived from the corresponding dependent claims, the description and the drawings. The features and details described here in conjunction with the production method according to the invention are of course also applicable to the production aspect according to the invention, and vice versa, so that the disclosures for the individual inventive aspects are always mutually referenced or can be mutually referenced.

[0004] The present invention is particularly useful for providing a robust fuel cell system.

[0005] Therefore, according to a first aspect of the present invention, a production method for producing a fuel cell stack is proposed.

[0006] The manufacturing method includes manufacturing a plurality of fuel cells through a plurality of manufacturing steps on an assembly line, obtaining voltage values ​​of corresponding fuel cells among the plurality of fuel cells from a starting time point to an ending time point, interrupting the fuel supply to the fuel cell at the starting time point, assigning the voltage value to a first group indicating a non-faulty state of the fuel cell or to a second group indicating a faulty state of the fuel cell with the aid of a machine learning machine, screening out the fuel cells whose voltage values ​​are assigned to the second group by the machine learning machine, and assembling only the fuel cells whose voltage values ​​are assigned to the first group by the machine learning machine into a fuel cell stack.

[0007] In the context of the present invention, a faulty state is understood to be a state of the fuel cell which is abnormal and leads to premature aging or a reduced service life compared to a normal state.

[0008] The proposed invention is based on the principle that the behavior of the fuel cell is determined based on the voltage profile at the end of the operating cycle of the fuel cell, i.e., during the so-called "discharge", and is used to evaluate the state of the fuel cell, and the fuel supply to the fuel cell is interrupted during the discharge. For this purpose, the voltage profile of the corresponding fuel cell is evaluated from a starting time point to an end time point, and the fuel supply to the fuel cell is interrupted at the starting time point, the end time point being, for example, a predetermined operating state or a predetermined time interval to the starting time point, which interval can last, for example, one second.

[0009] In order to assess the state of the fuel cell, a machine learning machine is used according to the present invention, which assigns the voltage values ​​determined during "discharging", i.e., a plurality of voltage values ​​or a course of voltage values, to a first group indicating a non-faulty state or to a second group indicating a faulty state. Accordingly, the machine learning machine automatically identifies a faulty fuel cell and a non-faulty fuel cell.

[0010] By using a machine learning engine, it is possible to dispense with predetermined check limits, minimizing the number of fuel cells that are incorrectly identified as defective. Instead, defective or abnormal fuel cells are identified based on an assignment logic learned and continuously adapted by the machine learning engine.

[0011] In particular, it is provided that, for training the machine learning engine, only voltage values ​​are transmitted to the machine learning engine, and correspondingly no so-called “label data”, ie, basic facts predefined by a user.

[0012] In particular, it can be provided that if the machine learning engine assigns the voltage value to the second group or not to the first group, a fault message is output on an output unit such as a display and / or a loudspeaker. This means that if a faulty fuel cell is identified, a fault message is output.

[0013] Assigning a voltage value to the second group implies an abnormality in the behavior or construction of the corresponding fuel cell, such as a gas leak or a change in electrical conductivity (e.g. a short circuit), such that a defect of the fuel cell is likely to occur in the near future and a corresponding indication for the expected defect can be output.

[0014] It can be provided that the end time point is predetermined or the end time point is dynamically determined according to the start time point. The dynamically determined end time point can be, for example, a time point selected according to the operating state of the fuel cell system, so that the end time point changes according to the operating state currently set in the fuel cell system. For this purpose, for example, an allocation chart can be predetermined, in which each end time point is allocated to a corresponding operating state.

[0015] It may be provided that the machine learning engine comprises an unsupervised learning model which continuously updates itself as a function of the voltage values ​​of the different fuel cells supplied to the machine learning engine, resulting in an assignment of the voltage values ​​to the parameters of the first group.

[0016] The unsupervised learning model is based on a core program that divides detailed data, in the present case the course of voltage changes during "discharge", into groups. The core program itself defines the individual distinguishing features that lead to the division into the individual groups. Accordingly, it is not necessary to provide the core program with physical relationships or models. Instead, the core program itself determines these relationships and thus defines itself conditionally: what is to be understood as the first group and what is to be understood as the second group. Here, the first group is defined as the fault-free state of the fuel cell, which is illustratively done by: the core program obtains more data from the fuel cell in the fault-free state than the data obtained from the fuel cell in the faulty state. Accordingly, the first group is formed with the help of voltage values ​​that have a controlling effect within the normal range, while the second group relates to those voltage values ​​that are outside the first group or the corresponding "point cloud".

[0017] The unsupervised learning model is not provided with pre-screened voltage values ​​for training. Rather, the unsupervised learning model is trained only on the basis of voltage data that are also provided to the unsupervised learning model as input signals during real operation.

[0018] For example, the so-called “k-means algorithm”, “expectation-maximization algorithm” or “density-based spatial clustering of applications with noise algorithm” can be used as unsupervised learning models, which form groups with small differences and similar sizes.

[0019] It can be arranged that a machine learning machine is trained on faulty fuel cells.

[0020] In order to enable the machine learning machine provided according to the present invention to accurately distinguish between a non-faulty fuel cell and a faulty fuel cell or to classify the voltage values ​​into the first group or the second group, in the configuration of the present invention, the machine learning machine can be provided with input values ​​obtained based on the faulty fuel cell so that the machine learning machine can adapt its underlying mathematical model accordingly.

[0021] It can be provided that the warning indication “faulty manufacturing” is assigned to the fuel cell if the machine learning machine assigns the voltage value ascertained at the corresponding fuel cell to the second group.

[0022] An alarm indication, such as a fault report, output on a display or a loudspeaker can inform the user that the machine learning machine has assigned the corresponding fuel cell to the second group, so that the user can, for example, remove the fuel cell from the production process or supply it to a post-processing process.

[0023] Furthermore, it can be provided that the voltage values ​​are preprocessed by means of mathematical methods.

[0024] During preprocessing, for example in a preparation phase, dimensions are defined according to which the machine learning machine can make an allocation. This preprocessing can maximize the accuracy and speed of the allocation made by the machine learning machine because the number of dimensions potentially available for use is reduced.

[0025] Furthermore, it can be provided that at least one mathematical method includes a correlation of the voltage value with a number of predefined geometrical parameters.

[0026] By pre-processing geometric parameters, in particular outliers and / or examples, it is possible to dispense with a check limit for identifying defective fuel cells, so that the number of fuel cells that are incorrectly identified as defective is minimized.

[0027] Furthermore, it can be provided that an intervention limit is determined by means of a least squares difference method, starting from which the fault message is output.

[0028] By means of the intervention limit, it is possible to avoid the output of an error message in the case of small deviations of the voltage values ​​(for example, if the deviation lies within the permissible tolerance).

[0029] Furthermore, it can be provided that the intervention limit is increasingly lowered as the number of production cycles carried out increases.

[0030] By lowering the intervention margin more and more, as the machine learning machine is continuously trained, the machine learning machine is gradually used as a decision maker for outputting fault reports. Accordingly, the untrained machine learning machine is protected from making erroneous negative decisions by the high intervention margin, and the trained machine learning machine is burdened by the low intervention margin. In this case, a new untrained machine learning machine can be preconfigured by a calculation model predefined by a central database, so that the machine learning machine further refines the provided calculation model if necessary and adapts it to specific given conditions in the corresponding fuel cell system if necessary.

[0031] According to a second aspect of the present invention, a manufacturing system for manufacturing a fuel cell stack is provided.

[0032] The production system comprises a production line and a computing unit, wherein the computing unit is configured to carry out a possible configuration of the proposed production method.

[0033] A computing unit in the context of the present invention is understood to be a computer, in particular a processor or a subprocessor or any other programmable switching circuit.

[0034] Further advantages, features and details of the invention are apparent from the following description, in which exemplary embodiments of the invention are described in detail with reference to the accompanying drawings. The features mentioned in the claims and in the description are essential to the invention individually or in any combination. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Here it is shown:

[0036] Figure 1 : Schematic illustration of possible configurations of the proposed fabrication method.

[0037] Figure 2 : Schematic diagram of possible configurations of the proposed manufacturing system. DETAILED DESCRIPTION

[0038] exist Figure 1 100 is described in the specification. The manufacturing method 100 includes: a manufacturing step 101, in which a plurality of fuel cells are manufactured on a production line through a plurality of manufacturing steps; an obtaining step 103, in which the voltage value of the voltage of the corresponding fuel cell among the plurality of fuel cells is obtained from a starting time point to an ending time point, at which the fuel supply to the fuel cell is interrupted; an allocation step 105, in which the voltage value is allocated to a first group indicating a non-faulty state of the fuel cell or to a second group indicating a faulty state of the fuel cell with the aid of a machine learning machine; a screening step 107, in which the fuel cells whose voltage values ​​are allocated to the second group by the machine learning machine are screened out; and an assembling step 109, in which only the fuel cells whose voltage values ​​are allocated to the first group by the machine learning machine are assembled into a fuel cell stack.

[0039] The machine learning engine can be configured in particular to detect a defective fuel cell or to distinguish it from a non-defective fuel cell using so-called "end-of-line" electrochemical measurements, such as "discharge time" or a corresponding point cloud of measured values ​​as a basis for assigning the measured values ​​to one of the at least two groups. For this purpose, the machine learning engine can, for example, determine geometrical dimensions (such as distances, tolerance compliance, etc.) or receive them as input parameters by means of transmission.

[0040] Accordingly, the machine learning engine detects characteristics which lead to outstanding discharge values ​​or particularly low current / voltage values, so that the corresponding fuel cells can be screened out and rejected before being assembled into a fuel cell stack.

[0041] exist Figure 2, a manufacturing system 200 is described. The manufacturing system 200 comprises an assembly line 201, for example several machines 205, in particular several manufacturing robots, and a computing unit 203. The computing unit is configured to execute the manufacturing method 100. For this purpose, the computing unit 100 is connected, for example, to sensors 207 on the assembly line and controls the individual machines 205 of the assembly line in such a way that these machines 205 discard the currently processed fuel cell if the fuel cell is assigned to the second group by the machine learning machine or is identified as faulty.

Claims

1. A method (100) for manufacturing a fuel cell stack, in, The manufacturing method (100) comprises: - manufacturing (101) a plurality of fuel cells on a production line (201) by a plurality of manufacturing steps; - obtaining (103) a voltage value of a corresponding fuel cell among the plurality of fuel cells from a starting time point to an ending time point, interrupting the fuel supply to the fuel cell at the starting time point, - assigning (105) the voltage value to a first group describing a non-faulty state of the fuel cell or to a second group describing a faulty state of the fuel cell by means of a machine learning machine, - filtering out the fuel cells whose voltage values ​​are assigned to the second group by the machine learning machine (107), - Only the fuel cells whose voltage values ​​are assigned by the machine learning machine to the first group are assembled (109) into a fuel cell stack.

2. The manufacturing method (100) according to claim 1, It is characterized in that The end time point is predetermined or dynamically determined according to the start time point.

3. The manufacturing method (100) according to claim 1 or 2, It is characterized in that The machine learning machine includes an unsupervised learning model that continuously updates itself based on the voltage values ​​of different fuel cells supplied to the machine learning machine, resulting in the assignment of voltage values ​​to a first set of parameters.

4. The manufacturing method (100) according to any one of the preceding claims, It is characterized in that The machine learning engine is trained on faulty fuel cells.

5. The manufacturing method (100) according to any one of the preceding claims, It is characterized in that If the machine learning engine assigns the voltage value ascertained at the corresponding fuel cell to the second group, the warning indication "defective manufacturing" is assigned to this fuel cell.

6. The manufacturing method (100) according to claim 5, It is characterized in that An intervention limit is determined by means of a least squares error method, starting from which a warning indication is assigned to the corresponding fuel cell.

7. The manufacturing method (100) according to claim 6, It is characterized in that As the number of executed manufacturing cycles increases, the intervention margin decreases.

8. The manufacturing method (100) according to any one of the preceding claims, It is characterized in that The voltage value is preprocessed by means of at least one mathematical method.

9. The manufacturing method (100) according to claim 8, It is characterized in that The at least one mathematical method comprises a correlation of the voltage value with a number of predefined geometrical parameters.

10. A manufacturing system (200) for manufacturing a fuel cell stack, in, The manufacturing system comprises: - assembly line (201), - a calculation unit (203), The computing unit (203) is configured to implement the manufacturing method (100) according to any one of claims 1 to 9.