Method and inspection system for inspecting bipolar plate of electrochemical cell, in particular fuel cell

Abstract

The invention relates to a method for inspecting a bipolar plate of an electrochemical cell, in particular a fuel cell, in which an image (12) of a surface of the bipolar plate (14) is created, the image (12) is inspected by an automated image-processing-supporting evaluation system for possible defects (28), and the surface of the bipolar plate (14) is inspected by the evaluation system if the inspected bipolar plate (14) is identified as a suspected plate that may be defective. A detailed inspection (22) of the areas of the suspected sheet identified as likely defective is performed. By means of an evaluation method applied in an automated evaluation system during image processing of the image (12), it is possible to increase the inspection quality during inspection of an electrochemical cell, such as a bipolar plate of a fuel cell, without significant additional effort, such that a low-cost and reliable inspection can be achieved.

Description

[0001] This application is a divisional application of the invention patent application with application number 202180013804.3. The filing date of the original application is February 26, 2021, and the name of the invention is "Methods and inspection facilities for inspecting bipolar plates of electrochemical cells, especially fuel cells." Technical Field

[0002] The invention relates to a method for testing bipolar plates, in particular of electrochemical cells, such as fuel cells, and to a testing facility provided therefor, by means of which bipolar plates of electrochemical cells, such as fuel cells, can be tested for defects during production. Background Art

[0003] A fuel cell stack or stack generally includes a plurality of fuel cells in a stacked configuration. The corresponding fuel cell includes an electrolyte and electrodes contacted via a conductive plate. In the case of a polymer electrolyte fuel cell operating in a low temperature range, a polymer membrane-electrode unit is present. In order to separate the individual polymer electrolyte fuel cells in the stack, a conductive bipolar plate generally made of metal is used. The conductive bipolar plate is not only used to electrically contact the electrodes and conduct the current to adjacent cells, but also to assist in conveying and distributing fuel and cooling media, as well as to conduct heat and reaction products. For this purpose, the bipolar plate generally has a gas distribution area, which, due to its structure, results in an optimal distribution of gaseous fuels, mainly hydrogen and oxygen, with respect to the membrane surface. The bipolar plate can be manufactured here by two or more thin-walled metal sheets welded to each other, and the metal sheets constitute the desired flow path, cooling channel and / or opening.

[0004] It is known from US 2019 / 0296379 A1 to check a fuel cell by applying and evaluating different voltages.

[0005] US2019 / 0340747A1 discloses a quality monitoring system and method in the field of fuel cell production lines.

[0006] US 2013 / 0230072 A1 describes a method and a device for fault detection of fuel cell components, in particular bipolar plates.

[0007] DE 103 93 237 B4 discloses a method for detecting electrical defects in thin-film electrode arrangements.

[0008] DE 38 09 221 A1 describes a method and a device for detecting defects, in particular cracks and / or constrictions, on punched parts or other workpieces.

[0009] DE 10 2015 221 697 B3 discloses a device for determining surface properties of a component surface, in particular defects.

[0010] DE 10 2016 211 449 A1 describes an inspection system having a portable inspection unit and a method for inspecting a component.

[0011] DE 10 2009 059 765 A1 discloses a method for producing a bipolar plate. The bipolar plate is formed stepwise from a first plate and a second plate, wherein after each step, an actual pattern consisting of a large number of lines is generated by projection onto an intermediate product representing the respective step and recorded with the aid of a camera. This is then compared with a reference pattern.

[0012] There is a long-standing need to cost-effectively and reliably test components of electrochemical cells, such as fuel cells, before they are installed together with other components in an electrochemical cell, such as a fuel cell, or a fuel cell stack. Summary of the invention

[0013] The object of the present invention is to specify measures which allow for a cost-effective and reliable testing of components of electrochemical cells, in particular fuel cells.

[0014] This object is achieved by a method having the features of claim 1 and by a testing arrangement having the features of claim 7. Preferred embodiments of the invention are given in the dependent claims and in the following description, which can each represent aspects of the invention individually or in combination.

[0015] One embodiment relates to a method for inspecting bipolar plates of electrochemical cells, in particular fuel cells, wherein an image of the surface of the bipolar plate is created, the image is inspected for possible defects by an automated evaluation system that supports image processing, and if the evaluation system identifies the inspected bipolar plate as a suspected plate that may be defective, a detailed inspection of the area of ​​the suspected plate identified as possibly defective is performed, wherein the evaluation system has a neural network for identifying suspected plates that is trained based on a plurality of images of the bipolar plate showing non-defective and defective bipolar plates, and the neural network is further trained with the help of the results of the detailed inspection, wherein the neural network performs reinforcement learning.

[0016] See also the article entitled “So functional reinforcement learning” from April 11, 2019, available at: https: / / www.alexanderthamm.com / de / blog / einfach-erklaert-so-functional-reinforcement-learning /

[0017] In the sense of the present invention, a bipolar plate is understood to be a component of an electrochemical cell that provides the functions of an electrode and a flow-conducting element. This can be, for example, a bipolar plate of a fuel cell or an electrolyzer or also an electrode plate of a redox flow battery.

[0018] In the context of bipolar plate inspection, in particular, other important inspection results or process data related to the manufacture of bipolar plates are used as labels or reward information for reinforcement learning. Therefore, it is preferred to take into account the results of the sealing inspection, online process data of the welding process, detailed inspection by thermal imaging, morphological inspection by determining the quality of the weld, or X-ray / CT information about the corresponding weld. Detailed inspections can be included 100% online or randomly in the case of slower process times. In this case, only the set of welds that are inspected in detail is considered in the training data set of the machine learning method, which grows during the bipolar plate production process and increases in clarity over time.

[0019] The main manufacturing step in manufacturing fuel cells, especially polymer electrolyte fuel cells, is to produce bipolar plates, which are especially made by welding two or more metal plates or plates. The weld seams produced during welding can be well checked by visual inspection. However, the visual inspection is performed automatically in the form of recording the surface of the bipolar plate, especially the two flat sides of the bipolar plate, in an image by means of a camera and conveying them to an image processing device of an automated evaluation system. The evaluation system can process the image and, for example, identify a specific shape and / or color and / or brightness and / or contrast according to optical parameters, and check whether these are within specific expected values. This inspection activity can be well automated and can be implemented by means of the evaluation software of the evaluation system. In addition, it can be checked whether the weld seam is extended in the correct relative position relative to, for example, the stamping or structure of the channel that guides the medium in the gas distribution area. The stamping and structure of the welded bipolar plate can be identified in the image of the surface of the bipolar plate and used, for example, as a reference for the correct positioning of the weld seam.

[0020] In addition to defects related to the welded connections at the bipolar plates, the plates and their structures, in particular the three-dimensional structure in the area of ​​the gas distribution area, the dimensional deviations of the openings formed in the plate ends and in the area of ​​the gas distribution area, etc., can also be checked in order to identify and single out irregularities and intervene in a corrective manner within the scope of the corresponding method steps that may have generated the deviations.

[0021] However, if it is not possible in the image processing evaluation of the images to identify, for example, the correct design of the stampings or structures in the region of the gas distribution area with the required reliability, which may be due to manufacturing errors, the bipolar plates can also be checked for such defects with the aid of the evaluation system.

[0022] Bipolar plates that are considered to be "abnormal" by the evaluation system are subjected to detailed inspection as suspected plates, in which inspection is performed with significantly higher costs if necessary. However, since only suspected plates are subjected to detailed inspection, the inspection costs can be reduced. For example, eddy current inspection is not performed in each bipolar plate, but only in suspected plates that are found to be suspicious in automated visual inspection by the evaluation system. Here, the following knowledge is used, and it is not a problem to visually look almost perfect welds according to experience, and although there may be sufficient, especially tight enough connections in seemingly imperfect welds, this cannot be guaranteed by pure visual inspection in all cases. Thus, bipolar plates that are evaluated as unsuspicious in automated image processing can be further processed without further detailed inspection, and after detailed inspection, only suspected plates that do not actually meet the preset standards in the suspected plates are picked out as waste, and the suspected plates that still meet the established standards can be transported again for further processing. Avoiding the unnecessary removal of suspected plates as waste due to practical reasons avoids unnecessary costs. By means of the evaluation method used during image processing of images in an automated evaluation system, the inspection quality during the inspection of fuel cells can be increased without requiring significant additional effort, so that a cost-effective and reliable inspection of fuel cells is possible.

[0023] In particular, at least one non-destructive inspection, in particular a penetration inspection, a magnetic particle inspection, an ultrasonic inspection, a radiographic inspection and / or an eddy current inspection, is performed on the suspected plate in the area identified as possibly defective during the detailed inspection. Thus, damage to the suspected plate during the detailed inspection can be avoided. In the case where the detailed inspection should show that the suspected plate still meets the required profile, the suspected plate can be delivered to the subsequent manufacturing steps as a conventional bipolar plate again. Thus, unnecessary costs in the manufacture of fuel cells can be avoided. At the same time, compared to a pure visual inspection, the relationship existing inside the suspected plate can be measured, thereby achieving a particularly reliable inspection. The detailed inspection can be performed automatically and / or by inspectors trained for this purpose.

[0024] The requirement profile may relate in particular to the tightness and position of weld seams, the permeability and position of structures in the gas distribution region, the arrangement of openings for the combustion gas supply, the design and position of seals and the like.

[0025] According to the present invention, the evaluation system has a neural network for identifying suspected plates, which is trained based on a plurality of images of bipolar plates showing non-defective and defective bipolar plates. The neural network can be trained by machine learning so that the image evaluation of images in the evaluation system can be improved. As a result, it is particularly feasible to consider a large number of different, possibly very differently weighted parameters when evaluating images. Here, the weights of the parameters are automatically adjusted during the training of the neural network so that in the image evaluation of the images used for training, in which it is already known in advance whether the bipolar plates with such an appearance are normal, the correct images are evaluated as "normal" or "abnormal". This makes it possible to automatically identify complex damage patterns even in the case of complex geometries. As a result, the inspection quality is improved.

[0026] According to the present invention, the neural network is further trained with the help of the results of the detailed inspection. When evaluating bipolar plates, the detailed inspection of the suspected plates leads to additional knowledge, wherein the additional training data is permanently fed back to the evaluation system. This leads to reinforcement learning of the neural network, especially as part of a deep learning structure with multiple intermediate layers of the neural network. "Deep learning" means a machine learning method using an artificial neural network with a large number of intermediate layers, so-called "hidden layers", between the input layer and the output layer. A large-scale internal structure is thus generated. This is a method of information processing. Machine learning is an adaptive algorithm. Deep learning is a subset of machine learning and uses a series of hierarchical layers or conceptual levels to perform the machine learning process. As a result, the inspection quality is always further improved over time. Reinforcement learning can be performed continuously or incrementally.

[0027] In particular, in order to produce a bipolar plate, two or more plates are welded, wherein during welding a plurality of images of the weld seam produced during welding are created and evaluated and / or an image of the entire weld seam is created and evaluated after welding. The image evaluation in the evaluation system can already be carried out during the welding process, so that the melting zone produced when the area to be welded melts can also be taken into account in the evaluation. However, it is also possible to carry out the evaluation only at the end of the welding process, thereby reducing the amount of data to be taken into account in the evaluation, in particular the number of images. It is also possible to illuminate the bipolar plate and create an image using a camera independently of the welding process and / or independently of the light effects and / or reversible thermal expansion effects produced during welding.

[0028] Preferably, the process control of the welding is adjusted with the aid of the result of the evaluation by the evaluation system. In particular, if the bipolar plate is evaluated as "abnormal", the evaluation system can classify possible defects according to the criteria used for this purpose. This makes it possible to associate a specific identified defect with a specific classification and possibly a specific cause.

[0029] If a defect of the suspected plate that is associated with a particular class of defects can be associated with an erroneous process control during welding, it is possible to adjust the process control by means of the welding process control by feeding back the results of the evaluation by the evaluation system so that such defects are avoided in the future. The evaluation system can be part of a closed control loop for the welding process control, for example, in order to make wear-related adjustments to the control variables. Manufacturing errors during welding can thus be detected very early and automatically corrected, so that unnecessary scrap and unnecessary costs are avoided.

[0030] Another embodiment relates to an inspection facility for inspecting bipolar plates of fuel cells by means of the method according to the invention, comprising: a camera for creating at least one image of the surface of the bipolar plate; an automated evaluation system supporting image processing for checking the image for possible defects; and an adapter operable by the evaluation system for exporting bipolar plates identified by the evaluation system as suspected plates with possible defects to a detailed inspection station. The inspection facility is designed to carry out the aforementioned method. The inspection facility is hereby constructed and improved as explained above with respect to the method.

[0031] Bipolar plates that are considered as "abnormal" by the evaluation system are subjected to detailed inspection as suspected plates, in which the inspection is carried out with significantly higher costs if necessary. However, since only the suspected plates are subjected to detailed inspection in the detailed inspection station, the inspection costs can be reduced. Bipolar plates that are evaluated as non-suspicious in the automated image processing in the evaluation system can thus be processed further without further detailed inspection, and after the detailed inspection, only the suspected plates that do not actually meet the preset standards are picked out as waste. Suspected plates that still meet the established standards are conveyed again for further processing. Unnecessary rejection of suspected plates as waste is avoided, so that unnecessary costs are avoided. The evaluation method used when processing images in the automated evaluation system can improve the inspection quality when inspecting bipolar plates without significant additional costs, so that low-cost and reliable inspections can be achieved.

[0032] The evaluation system of the inspection facility according to the present invention has a neural network for identifying suspected plates, which is trained based on a plurality of images showing non-defective and defective bipolar plates, wherein the evaluation system has an interface coupled to the neural network, communicating with a detailed inspection station, and for feeding in the results of the detailed inspection, for the purpose of further training the neural network with the help of the results of the detailed inspection. The detailed inspection of the suspected plates in the detailed inspection station results in additional knowledge when evaluating the bipolar plates, which is fed back to the evaluation system as training data via the interface of the evaluation system. This leads to reinforcement learning of the neural network, especially as part of a deep learning structure with multiple intermediate layers of the neural network. As a result, the inspection quality is always further improved over time.

[0033] In particular, the evaluation unit has an output port that can be coupled to a production unit for manufacturing bipolar plates, and is used to adjust the process control of the production unit according to the results of the evaluation by the evaluation system. Such production units are, for example, laser welding facilities, stamping facilities, forming facilities, etc. In particular, when the bipolar plate is evaluated as "abnormal", the evaluation system can classify possible defects according to the standards used for this purpose. This can be achieved by associating specific identified defects with specific classifications and possibly specific causes. If the defects of the suspected plate associated with a specific category of defects can be associated with erroneous process control in the production unit, it is feasible to adjust the process control with the help of the production unit by feeding back the results of the evaluation via the output port of the evaluation system so that such defects can be avoided in the future. The evaluation system can be part of a closed-loop control loop for process control of the production unit, for example, in order to make wear-related adjustments to the control variables. Manufacturing errors can thus be identified very early and automatically corrected, so as to avoid unnecessary scrap and unnecessary costs.

[0034] Preferably, the detailed inspection station has at least one inspection device for non-destructive inspection of the suspected plate, in particular penetration inspection and / or magnetic particle inspection and / or ultrasonic inspection and / or radiographic inspection and / or eddy current inspection. In this way, damage to the suspected plate during the detailed inspection can be avoided. In the case where the detailed inspection should show that the suspected plate still meets the required profile, the suspected plate can be transported again to the subsequent manufacturing steps as a conventional bipolar plate. In this way, unnecessary costs in the manufacture of components for fuel cells can be avoided. At the same time, compared with a pure visual inspection, the relationships existing inside the suspected plate can be measured, thereby achieving a particularly reliable inspection. The detailed inspection is performed automatically, wherein the inspector can be eliminated due to the continuous improvement of the inspection quality.

[0035] The method according to the invention and the testing device according to the invention are particularly suitable for testing bipolar plates used in fuel cells, in particular polymer electrolyte fuel cells, wherein the bipolar plates are formed from two or more metal sheets. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In the following, the invention is explained by way of example based on preferred embodiments with reference to the accompanying drawings, wherein the features shown below can each represent aspects of the invention individually and in combination. The accompanying drawings show:

[0037] Figure 1 : a schematic flow chart for performing an inspection of a bipolar plate, and

[0038] Figure 2 : Schematic perspective view of a part of an inspection installation for performing inspections in bipolar plates. DETAILED DESCRIPTION

[0039] exist Figure 1 In the process 10 shown in FIG. 1 , a plurality of images 12 are provided in a first step, the images showing in particular the surface of a bipolar plate 14 after a welding process. In a second step, the images 12 are divided, for example, into two categories, wherein one category 16 is provided for images 12 showing "normal" (OK) bipolar plates 14, and another category 18 is provided for images 12 showing "not normal" (NOK) bipolar plates 14. In particular, the category 18 provided for defective bipolar plates 14 may have a plurality of subcategories, which are associated with different causes for different defect modes. In a third step, a neural network 20 is trained with the aid of the images 12 divided into categories 16 and 18. The neural network 20 then independently gives an assessment of "OK" or "NOK" during the manufacturing process of the fuel cell 14 in a fourth step. The bipolar plates 14 considered "normal" are delivered to further production steps. In a fifth step, a detailed inspection 22 is performed on the suspected plates assessed as possibly "not normal" in order to confirm or correct the assessment given by the neural network 20 as "NOK". In the detailed inspection 22, a result is automatically generated, in which the neural network 20 classifies the image 12 as a real "NOK" or still classifies it as "OK" with a very high degree of reliability, at least in doubtful cases. This corresponds to the state after the second step, in which training data for the neural network 20 are generated. The data automatically generated as a result of the detailed inspection 22 are fed back to the neural network 20 as additional training data in order to achieve a reinforcement learning of the neural network 20.

[0040] As in Figure 2 As shown in FIG, the inspection device 24 can have a camera 26, which generates an image 12 from the surface of the bipolar plate 14, which is in particular welded. The camera 26 can also generate a plurality of images 12 or even a film sequence. For this purpose, the bipolar plate 14 can be suitably illuminated, so that all important features can be seen in the image 12, and possible defects 28 can be identified by an evaluation system with a neural network 20 through optical evaluation.

[0041] In addition to defects related to the welded connections at the bipolar plates, the plates and their structures, in particular the three-dimensional structure in the area of ​​the gas distribution area, the dimensional deviations of the openings formed in the plate ends and in the area of ​​the gas distribution area, etc., can also be checked in order to identify and single out irregularities and intervene in a corrective manner within the scope of the corresponding method steps that may have generated the deviations.

[0042] Reference numerals list

[0043] 10 Process

[0044] 12 images

[0045] 14 Bipolar Plate

[0046] 16 A category

[0047] 18 Another category

[0048] 20 Neural Networks

[0049] 22 Detailed inspection

[0050] 24 Inspection facilities

[0051] 26 Camera

[0052] 28 Possible defects

Claims

1. A method for inspecting a bipolar plate of an electrochemical cell, in particular a fuel cell, wherein an image (12) of the surface of the bipolar plate (14) is created, The image (12) is checked for possible defects (28) by an automated image processing-supported evaluation system, and In the event that the evaluation system identifies the inspected bipolar plate (14) as a suspected plate that may be defective, performing a detailed inspection (22) of the area of ​​the suspected plate identified as being defective, wherein the evaluation system has a neural network (20) for identifying suspicious plates, which is trained based on a plurality of images (12) of bipolar plates (14) showing non-defective and defective bipolar plates (14), and further trains the neural network (20) with the aid of the results of the detailed inspection (22), wherein the neural network (20) performs reinforcement learning; wherein during the detailed inspection (22) at least one non-destructive inspection is performed on the suspected plate in the area identified as possibly defective; As the non-destructive inspection, a penetrant inspection, a magnetic particle inspection, an ultrasonic inspection, a radiographic inspection, and / or an eddy current inspection are performed on the suspected plate in the area identified as possibly defective.

2. The method according to claim 1, wherein two or more sheets are welded together to produce the bipolar plate (14), wherein during welding a plurality of images (12) of the weld seam produced during welding are created and evaluated.

3. The method according to claim 2, wherein for the production of the bipolar plate (14), an image (12) of the entire weld seam is created and evaluated after welding. 4 . The method according to claim 2 , wherein the process control of the welding is adjusted with the aid of the result of the evaluation by the evaluation system.

5. An inspection facility for performing the method according to any one of claims 1 to 4, the inspection facility comprising: a camera (26) for creating at least one image (12) of the surface of the bipolar plate (14), an automated image processing-supported evaluation system for checking the image (12) for possible defects (28), and an adapter operable by the evaluation system for exporting the bipolar plates identified by the evaluation system as suspected plates that may be defective to a detailed inspection station having at least one inspection device for performing non-destructive inspection of the suspected plates, in particular penetrant inspection and / or magnetic particle inspection and / or ultrasonic inspection and / or radiographic inspection and / or eddy current inspection; The evaluation system comprises a neural network (20) for identifying suspected plates, which is trained based on a plurality of images (12) of bipolar plates (14) showing non-defective and defective bipolar plates (14), wherein the evaluation system comprises an interface coupled to the neural network (20), communicating with the detailed inspection station, and for feeding in the result of the detailed inspection (22), for the purpose of further training the neural network (20) with the help of the result of the detailed inspection (22).

6. The inspection facility according to claim 5, characterized in that: The evaluation unit has at least one output port for adjusting the process control of the production unit according to the result of the evaluation by the evaluation system, and the output port can be coupled to at least one production unit, in particular a welding facility and / or a stamping facility and / or a forming facility, to produce the bipolar plate (14).

7. The inspection facility according to claim 6, characterized in that: The at least one production unit is a welding facility and / or a punching facility and / or a forming facility.

Images