Manufacture of quality inspection system

Abstract

Manufacturing of a quality inspection system. This invention relates to a method for manufacturing a quality inspection system, wherein the quality inspection system is configured to execute a quality inspection model, wherein the quality inspection model has a filter mask and a quality model, wherein the quality inspection system is configured to determine at least one quality characteristic of at least one battery pack and has at least one electrochemical impedance spectroscopy unit for detecting spectral inspection data of the at least one battery pack within a frequency range, wherein a first machine learning method and a second machine learning method are executed in a coupled and coordinated manner based on the spectral learning data, such that the first machine learning method creates a filter mask, the filter mask determining the analytical data, and the second machine learning method creates a quality model, the quality model being optimized for maximizing the quality of the at least one quality characteristic.

Description

Technical Field

[0001] This invention relates to a method for manufacturing a quality inspection system, wherein the quality inspection system is configured to perform a quality inspection model, which includes a filter mask and a quality model. Furthermore, this invention relates to related quality inspection systems. Background Technology

[0002] Given the rapidly growing demand, the efficient and scalable production of battery packs, especially lithium-ion battery packs, is becoming increasingly important. However, industrial manufacturing processes also involve high defect and scrap rates, as well as significant uncertainties in estimating the actual capacity or operational life of individual battery pack cells. Furthermore, predicting these and other quality parameters, particularly from dedicated self-discharge measurements, is time-consuming, sometimes many times the duration of production, or is associated with high resource requirements for energy and / or space.

[0003] A much more precise and faster method for checking the quality of battery pack cells lies in what is known as electrochemical impedance spectroscopy (EIS). EIS measures the response of battery pack cells to signal strength or impulse responses to specific current or voltage excitations over a wide frequency range (typically from mHz to several kHz).

[0004] The known problem in this context is the evaluation of complex frequency response measurements, which involve up to thousands of measurement points, each with a real / imaginary part of its frequency-dependent impedance value. Furthermore, deriving or modeling the mass characteristics of interest in the battery from high-dimensional raw data is particularly challenging.

[0005] To date, statistical data analysis and spectral analysis methods have been particularly employed to derive battery characteristics from the obtained impedance spectra. A common approach is to fit existing analytical models—i.e., models based on physical equations—to the spectral data. However, the effectiveness of such models is limited and typically does not reflect all observed characteristics in the measurement data. Furthermore, these methods are not suitable for identifying new relationships or correlations, but rather only allow for discerning how well previous understandings or models fit the data.

[0006] Recently, purely data-based methods, particularly regression methods or "supervised machine learning," have also been mentioned or applied in the literature. Here, most involve subjecting a series of battery pack cells to impedance spectroscopy (EIS) methods based on laboratory tests, followed by physical measurements to determine actual capacity or stress tests to determine remaining life. This generates a dataset from which the impedance spectrum can be correlated with subsequently measured quality characteristics. A regression model can then be trained to predict expected characteristics for unknown battery pack cells. However, what remains unresolved in this case is the understanding of which characteristics of the EIS measurements decisively influence quality. To determine the quality of unknown battery pack cells, especially during manufacturing, EIS measurements must be performed across the entire frequency range to generate the input data needed for the regression model to predict quality. The disadvantage of this approach is its very high time consumption. Summary of the Invention

[0007] The objective of this invention is to provide a solution for improving the quality analysis of battery pack cells.

[0008] The invention is derived from the features of the independent claims. Advantageous improvements and designs are the subject of the dependent claims. The designs, applicability, and advantages of the invention are illustrated in the following description and drawings.

[0009] This invention relates to a method for manufacturing a quality inspection system. The quality inspection system is configured to execute a quality inspection model, wherein the quality inspection model includes a filter mask and a quality model.

[0010] The quality inspection system further comprises at least one quality characteristic for determining at least one battery pack cell, and has at least one electrochemical impedance spectroscopy (EIS) unit for detecting spectral inspection data of at least one battery pack cell within a frequency range.

[0011] The method includes the following steps:

[0012] - Create a quality inspection model, and

[0013] - A quality inspection system is manufactured by implementing a quality inspection model onto the computational unit of the quality inspection system.

[0014] The steps to create a quality inspection model include the following:

[0015] - Spectral learning data of at least one battery pack cell were detected using at least one electrochemical impedance spectroscopy unit.

[0016] - A filter mask is created using a first machine learning method, wherein the filter mask constitutes analytical data for determining analytical data from spectral test data, wherein determining the analytical data includes selecting at least one sub-range of the frequency range of the spectral test data by querying the filter mask, wherein the at least one sub-range covers at most 50%, and particularly at most 30%, of the entire frequency range of the spectral test data, and

[0017] - A quality model is created using a second machine learning method, wherein the quality model constitutes at least one quality feature for calculating at least one battery pack cell based on the analysis data.

[0018] The first and second machine learning methods are executed in a coupled and coordinated manner based on the spectral learning data. The first machine learning method creates a filter mask that determines the analysis data, while the second machine learning method creates a quality model that optimizes the maximization of the quality of at least one quality feature. This quality feature may also be referred to as a quality characteristic or a fitness feature.

[0019] One aspect of the invention involves arranging a first machine learning method and a second machine learning method and coordinating them to create a filter mask that enables the creation of a quality model, the quality model being optimized for maximizing the quality of at least one quality feature. The filter mask constitutes a significant limitation and / or reduction of a frequency range. Filtering through the filter mask results in a reduced number of frequency windows. A subrange of the frequency range of the spectral test data queried by the filter mask particularly includes 50% of the entire frequency range of the spectral test data. In another embodiment, a subrange of the frequency range of the spectral test data queried by the filter mask particularly includes 30%, 25%, 20%, 15%, 10%, 5%, 3%, or 1%. Higher reduction and therefore smaller subranges have the advantages of less complex analysis and thus achieving meaningful quality criteria with less expense and resource usage.

[0020] To achieve this, the first and second machine learning methods are coordinated, coupled, and performed simultaneously, at the same time, and / or based on their coordination. In other words, a filter mask is created to enable the creation of a model with the best possible quality. Technically, this is achieved in particular by executing the first and second machine learning methods using data generated from the same first spectral data, respectively.

[0021] In other words, this means that: training a filter mask is performed by a first machine learning method, such that the first machine learning method provides a filter mask that selects at least one subrange of a frequency range, such that a second machine learning method can train a quality model that computes quality features with the highest accuracy, i.e., maximizing the quality of the quality model, based on at least one subrange.

[0022] The spectral data may optionally be normalized and / or consist of N discrete frequency data points.

[0023] Therefore, the invention proposed herein aims to perform complex electrochemical impedance spectroscopy (EIS) transformations using machine learning methods, which allow for the identification of information from the EIS spectra that is important for a given quality characteristic. The resulting quality characteristics can then be used to accelerate and refine quality inspection systems, particularly quality methods, maintenance methods, and testing methods.

[0024] Therefore, the results of the proposed method can be used to further reduce the frequency range by only needing to check the number of cells in each battery pack. This makes the method more efficient and scalable for application in manufacturing and quality assurance processes.

[0025] The processing method proposed here is not limited to EIS measurements, but can also be applied to similar measurement methods, especially for measuring frequency spectra, for battery characterization, but also for material testing methods performed by means of infrared spectroscopy or spectra of the contents.

[0026] In an improved embodiment of the present invention, the at least one quality feature includes at least:

[0027] - Charge storage capacity,

[0028] - Operating life,

[0029] - Remaining lifespan, and / or

[0030] - Self-discharge rate.

[0031] The quality model is designed to calculate at least one quality characteristic of at least one battery pack cell based on analytical data. Different configurations of at least one quality characteristic have the advantage of allowing for different statements regarding the quality of the battery pack cells.

[0032] In another improvement of the invention, the quality inspection system is further configured to use at least one quality characteristic in the manufacture of at least one battery pack cell.

[0033] - control,

[0034] - Test methods, and / or

[0035] - Maintenance.

[0036] Using at least one quality characteristic to control the manufacturing of at least one battery pack particularly includes modifying manufacturing parameters, adapting machine parameters, and / or adapting the manufacturing process.

[0037] Inspection methods that use at least one quality characteristic in the manufacture of at least one battery pack cell particularly include switching additional inspection cycles to subject the inspection results to further analysis and / or testing of additional inspection parameters.

[0038] Maintenance that uses at least one quality characteristic in the manufacture of at least one battery pack includes, in particular, modifying manufacturing parameters, temporarily suspending manufacturing, performing maintenance work, adapting machine parameters, and / or adapting to future manufacturing processes.

[0039] In another improvement of the invention, determining the analytical data further includes applying an autoencoder, particularly a variational autoencoder. An autoencoder can also be referred to as a feature extractor, and feature extraction can be performed, particularly from spectral testing and learning data. The autoencoder is applied to the spectral testing and / or learning data, in addition to a filter mask, and thus creates the analytical data.

[0040] In another improved embodiment of the invention, the creation of the quality inspection model further includes the following steps:

[0041] - Create an autoencoder using a third machine learning method.

[0042] In another improvement of the invention, the first, second, and third machine learning methods utilize data in a coupled and coordinated manner according to the spectral learning data, such that the first machine learning method creates a filter mask and the third machine learning method creates an autoencoder, wherein the filter mask and the autoencoder determine the analysis data, such that the second machine learning method creates a quality model that is optimized for maximizing the quality of at least one quality feature.

[0043] Similar to the coordinated creation of filter masks and quality models, the additional coordination in the creation of autoencoders has the advantage of creating a quality model that calculates quality features with the highest accuracy, i.e., maximizing the quality of the quality model, based on at least one subrange and analysis data.

[0044] In another improvement of the invention, the analysis data is a latent representation of the spectral test data, particularly modified by an autoencoder. The latent representation can also be referred to as a compressed representation or a representation that includes a reduced number of dimensions. Therefore, the autoencoder has the advantage of reducing the complexity of the spectral test and / or learning data.

[0045] In another improvement of the invention, the analytical data has reduced complexity compared to the spectral test data. This has the advantages of providing the quality model with data of reduced complexity and, in particular, feature extraction has already been performed.

[0046] In summary, a (variational) autoencoder is trained on a masked input spectrum to generate a latent representation of the masked input spectrum with reduced complexity (= the encoder's output). The autoencoder can perform clustering, classification, reconstruction, feature extraction, and / or filtering of the masked input spectrum. This can also be referred to as regularization.

[0047] In another improvement of the invention, the quality model further constitutes a means for calculating at least one second quality characteristic of at least one battery pack cell based on analytical data. This has the advantage that, particularly based on the magnitude of the uncertainty of the determined quality characteristic, various other processing steps can be introduced, particularly enabling a complete remeasurement or direct sorting of the EIS spectrum of the battery pack cells under consideration.

[0048] In another improvement of the invention, the quality inspection model has at least one second quality model, which constitutes a second quality characteristic for calculating at least one battery pack cell based on analytical data. In this design, multiple quality models can be trained, particularly by increasing the frequency of the number of models, especially by reducing L1 regularization during training. In an implementation, the quality models will be implemented in a chain: a subsequent quality model is evaluated only if the uncertainty predicted by the preceding quality model is too large. The advantage thus achieved is time savings by avoiding the need for non-mandatory frequency measurements.

[0049] In another design, it is conceivable to train different frequency masks and quality models for different types of battery packs.

[0050] In another improvement of the invention, determining the analytical data further includes:

[0051] - Apply dimensionality reduction methods, especially t-SNE.

[0052] - Apply transformation methods, and / or

[0053] - Apply feature extraction to spectral test data.

[0054] This has the following advantages: different methods can be used to construct autoencoders or as alternatives to autoencoders.

[0055] In another improvement of the invention, it is determined that the analytical data further includes applied spectral test data:

[0056] - Grouping,

[0057] - Classification,

[0058] - Regularization,

[0059] - Reproduction, and / or

[0060] - Filtering.

[0061] This has the following advantages: it reduces the complexity of spectral testing and / or learning data, which may be reduced upon availability, and thus reduces the cost of using quality models.

[0062] In another improvement of the invention, the spectral learning data includes both test data and training data. This has the advantage that the machine learning method can not only be trained but also tested.

[0063] In another improvement of the invention, the quality model utilizes:

[0064] - Convolutional Neural Network (CNN).

[0065] - Deep neural networks,

[0066] - Linear models, especially partial least squares (PLS) models.

[0067] - Bayesian Neural Network (BNN), or

[0068] - Bayesian Networks (BN)

[0069] To execute.

[0070] Bayesian neural networks (BNNs) can generate statements about the predictable uncertainty of predictions, in addition to the average value of quality features.

[0071] Furthermore, the present invention includes a quality inspection system manufactured according to the method of the present invention.

[0072] Furthermore, the present invention includes a method for using the quality inspection system according to the present invention.

[0073] It has the following steps:

[0074] - Determine analytical data from spectral test data using a filter mask, wherein determining the analytical data comprises selecting at least one sub-range of the frequency range of the spectral test data by querying the filter mask, wherein the at least one sub-range covers at most 50%, and particularly at most 30%, of the entire frequency range of the spectral test data.

[0075] - Calculate at least one quality characteristic of at least one battery pack cell based on the analysis data using a quality model.

[0076] This invention offers the following advantages: According to current technology, performing EIS measurements across the entire frequency range results in high time consumption, especially when measuring low frequencies simultaneously. This means that the proposed method achieves time savings if only the frequencies characterizing battery quality must also be measured. In the case of a large number of batteries to be characterized, lengthy measurements would be impractical.

[0077] Furthermore, the method according to the invention enables not only the assessment of the end-of-line quality of the battery, but also the determination of statements regarding the expected lifespan during battery pack operation. This requires, in part, recording impedance spectra at regular time intervals during laboratory stress tests on a small number of batteries during aging tests.

[0078] Furthermore, EIS measurements can replace the long-term, continuous determination of battery self-discharge during manufacturing. Therefore, during the training cycle, a complete EIS spectrum is recorded in addition to the self-discharge determination, allowing this EIS spectrum to be used to train the method described above.

[0079] In subsequent production, self-discharge measurements can then be omitted, and instead, time-optimized EIS measurements can be performed only up to that point. This results in significant time savings during battery manufacturing. Attached Figure Description

[0080] The features and advantages of the present invention will become clear from the following explanation of several embodiments based on the illustrative drawings.

[0081] Figure 1 A flowchart of the method according to the present invention is shown.

[0082] Figure 2 The diagram illustrates the process of creating a quality inspection model, and...

[0083] Figure 3 The diagram and process of the application quality inspection system are shown. Detailed Implementation

[0084] Figure 1 A manufacturing quality inspection system 1 according to the present invention is shown (in...) Figure 2 and / or Figure 3 The flowchart of the method shown is illustrated, wherein the quality inspection system 1 is configured to execute the quality inspection model 2 (in... Figure 2 and / or Figure 3 As shown in the figure), where the quality inspection model 2 has a filter mask 2a (in Figure 2 and / or Figure 3 (shown in) and quality model 2c (in) Figure 2 and / or Figure 3(as shown in the diagram), wherein the quality inspection system 1 is configured to determine at least one battery pack cell 4 (in Figure 2 and / or Figure 3 At least one quality feature 3c (shown in) Figure 2 and / or Figure 3 (as shown in the diagram), and has at least one electrochemical impedance spectroscopy unit for detecting spectral test data 3 of at least one battery pack cell 4 within a frequency range (in Figure 2 and / or Figure 3 (as shown in the image).

[0085] The method includes the following steps:

[0086] - Step S1: Create quality inspection model 2, and

[0087] - Step S2: Manufacturing the quality inspection system 1 by implementing the quality inspection model 2 onto the computing unit of the quality inspection system 1.

[0088] The steps for creating quality inspection model 2 are as follows:

[0089] - Step S21: Detect the spectral learning data 3 of at least one battery pack cell 4 using at least one electrochemical impedance spectroscopy unit (in... Figure 2 and / or Figure 3 (as shown in the image)

[0090] - Step S22: Create a filter mask 2a using a first machine learning method, wherein the filter mask 2a constitutes the analytical data 3a to be determined based on the spectral test data 3 (see step S24). Figure 2 and / or Figure 3 (shown in), 3b (in) Figure 2 and / or Figure 3 (as shown in the image)

[0091] - Step S24: Wherein the analysis data 3a, 3b are determined to contain at least one subrange of the frequency range of the spectral test data 3 selected by querying the filter mask 2a, wherein the at least one subrange covers at most 50%, and particularly at most 30%, of the entire frequency range of the spectral test data 2a, and

[0092] - Step S25: Create a quality model 2c using a second machine learning method, wherein the quality model 2c constitutes a quality feature 3c for calculating at least one quality feature 3c of at least one battery pack cell 4 based on the analysis data 3a, 3b.

[0093] The first machine learning method and the second machine learning method are executed in a coupled and coordinated manner based on the spectral learning data 3, such that the first machine learning method creates a filter mask 2a (see step S22), the filter mask determines the analysis data 3a, 3b (see step S24), and the second machine learning method creates a quality model 2c (see step S25), the quality model being optimized for maximizing the quality of at least one quality feature 3c.

[0094] Optionally, the analysis data 3a and 3b are determined (see step S24) and further include the application (S24b, not shown) of an autoencoder 2b, especially a variational autoencoder 2b (in Figure 2 and / or Figure 3 (As shown in the image).

[0095] Therefore, creating (see step S1) the quality inspection model 2 optimally includes the following steps:

[0096] - Step S23: Create autoencoder 2b using a third machine learning method.

[0097] The first, second, and third machine learning methods thus utilize data in a coupled and coordinated manner based on the spectrum learning data, such that the first machine learning method creates a filter mask 2a (see step S22) and the third machine learning method creates an autoencoder 2b (see step S23), wherein the filter mask 2a and the autoencoder 2b determine the analysis data 3a, 3b (see step S24), such that the second machine learning method creates a quality model 2c (see step S25), which is optimized for maximizing the quality of at least one quality feature 3c.

[0098] Figure 2 The creation process is shown (see step S1). Figure 1 ) Schematic diagram and process of quality inspection model 2.

[0099] Input value 3 is the unprocessed (possibly normalized) EIS spectrum 3 (consisting of N discrete frequency data points) from the training dataset of different battery packs 4.

[0100] The concept described herein stipulates that, through a combination of "supervised learning" and "unsupervised learning" methods, such as using a "variational auto-encoder" (VAE) 2b or a dimensionality reduction method 2b (such as t-SNE 2b), information about the target feature 3c (lifetime, capacity, ...) is transformed in the original EIS data 3, such that a constraint 2a on the relevant frequency range 3a (or a combination of ranges) can be derived from it, which is required for optimally evaluating unknown battery packs 4 with respect to their quality or suitability in the future.

[0101] Therefore, measurements of frequency ranges that are therefore irrelevant can be eliminated, saving time and resources during (mass) manufacturing.

[0102] Similarly, a trainable vector layer 2a (also called an "attention layer" 2a or frequency mask 2a) with N weights in the value range [0, 1] first transforms the input spectrum 3 into a reduced number of frequencies 3a (the masked input spectrum 3a). The training of these weights 2a is coupled with the subsequent optimization of the quality model 2c and the autoencoder 2b. That is, the training of the attention layer 2a is optimized and coordinated to filter out ranges that the quality model 2c utilizes to provide optimal results. Therefore, the training of the attention layer 2a and the quality model 2c is performed simultaneously. Therefore, it is preferable to select those frequency ranges that are particularly relevant to the subsequent adaptation of the learning objective 3c (quality model 2c and autoencoder 2b).

[0103] An autoencoder 2b (variable) is trained on the masked input spectrum 3a to generate a latent representation of the masked input spectrum 3a with reduced complexity (= the output 3b of encoder 2b). The autoencoder 2b can perform clustering, classification, reconstruction, feature extraction, and / or filtering of the masked input spectrum 2a. This can also be referred to as regularization 2b.

[0104] The output 3b of the autoencoder 2b, i.e. the latent representation 3b of the original data spectrum 3, is also used as input for training a supervised deep quality model 2c (e.g., composed of a convolutional neural network (CNN) 2c or a Bayesian network (BNN) 2c with uncertainty), in which predictions of the quality features 3c of the battery 4 are learned (the target variable 3c for this learning step comes from the measurements performed on the battery pack battery 4 after the EIS spectrum 3).

[0105] If training is successfully completed, the reduced representation 3b of the original spectrum 3 in the masked frequency space 3a (the output of the attention layer 2a) contains only those frequency ranges that are necessary and relevant for the prediction (by the quality model 2c) or the reproduction (by the autoencoder 2b) of the quality standard 3c.

[0106] Furthermore, implicit correction or suppression of sensor noise is performed in the impedance spectroscopy measurement 3. This can be done in particular by appropriate encoding in the latent representation 3b, i.e., learning to map the noisy sensor value 3 to the same latent representation 3b as the noisy value 3b, and thus reconstructing the latter from the noisy measurement 3.

[0107] Figure 3 The diagram and flow of the application quality inspection system 1 are shown.

[0108] When manufacturing battery pack cell 4, the trained quality model 2c can be used to improve efficiency as follows:

[0109] - A batch of battery pack cells 4 undergoes EIS scanning only within the selected frequency range; the required frequency mask 2a is obtained from the model's one-time training (see...). Figure 2 Extract the description.

[0110] - The results of the restricted EIS spectrum of battery 4 were used as input to a (CNN / deep) quality model 2c pre-trained for these frequency ranges in order to predict the expected quality 3c respectively.

[0111] – Based on the predictions of quality model 2c, individual batteries 4 can be sorted, or undergo a complete re-inspection if quality model 2c is not clear.

[0112] Although the invention has been further illustrated and described in detail by way of embodiments, the invention is not limited to the disclosed examples and other variations can be derived by those skilled in the art without departing from the scope of protection of the invention.

Claims

1. A method for manufacturing a quality inspection system (1), wherein the quality inspection system (1) is configured to perform a quality inspection model (2), wherein the quality inspection model (2) has a filter mask (2a) and a quality model (2c). The quality inspection system (1) comprises at least one quality characteristic (3c) for determining at least one battery pack cell (4) and has at least one electrochemical impedance spectroscopy unit for detecting spectral inspection data of the at least one battery pack cell (4) within a frequency range. It has the following steps: - Create the quality inspection model (2) described in (S1), and - The quality inspection system (1) is manufactured (S2) by implementing the quality inspection model (2) onto the computing unit of the quality inspection system (1). The creation of the quality inspection model (2) involves the following steps: - Spectral learning data of at least one battery pack cell (4) are detected by the at least one electrochemical impedance spectroscopy unit (S21). - A filter mask (2a) is created (S22) using a first machine learning method, wherein the filter mask (2a) constitutes analytical data (3a, 3b) for determining from the spectral test data, wherein the determined analytical data (3a, 3b) comprises at least one subrange of the frequency range of the spectral test data selected (S24a) by querying the filter mask (2a), wherein the at least one subrange covers up to 50% of the entire frequency range of the spectral test data. - The quality model (2c) is created (S25) by a second machine learning method, wherein the quality model (2c) constitutes at least one quality feature (3c) of the at least one battery pack cell (4) for calculating based on the analysis data (3a, 3b). The first machine learning method and the second machine learning method are executed in a coupled and coordinated manner based on the spectral learning data, such that the first machine learning method creates (S22) a filter mask (2a), the filter mask determines (S24) the analysis data (3a, 3b), and the second machine learning method creates (S25) a quality model (2c), the quality model being optimized for maximizing the quality of the at least one quality feature (3c).

2. The method of claim 1, wherein the at least one subrange covers up to 30% of the entire frequency range of the spectral test data.

3. The method according to claim 1 or 2, The at least one quality feature (3c) mentioned above includes at least: - Charge storage capacity, - Operating life, - Remaining lifespan, and / or - Self-discharge rate.

4. The method according to any one of claims 1 to 2 above, The quality inspection system (1) is further configured to use the at least one quality feature (3c) in the manufacture of the at least one battery pack cell (4). - control, - Test methods, and / or - Maintenance.

5. The method according to any one of claims 1 to 2 above, The analysis data (3a, 3b) determined in (S24) further includes the application of (S24b) an autoencoder (2b).

6. The method according to claim 5, wherein the autoencoder is a variational autoencoder.

7. The method according to claim 5, The creation of the quality inspection model (2) described in (S1) further includes the following steps: - The autoencoder (2b) is created by a third machine learning method (S23).

8. The method according to claim 7, The first, second, and third machine learning methods utilize data to perform in a coupled and coordinated manner based on the spectrum learning data, such that the first machine learning method creates (S22) a filter mask (2a) and the third machine learning method creates (S23) an autoencoder (2b), wherein the filter mask (2a) and the autoencoder (2b) determine (S24) the analysis data (3a, 3b), such that the second machine learning method creates (S25) a quality model (2c), wherein the quality model is optimized for maximizing the quality of at least one quality feature (3c).

9. The method according to any one of claims 1 to 2, The analytical data (3a, 3b) therein are potential representations of the spectral test data.

10. The method according to any one of claims 1 to 2 above, The analytical data (3a, 3b) described therein have reduced complexity compared to the spectral test data.

11. The method according to any one of claims 1 to 2 above, The quality model (2c) further constitutes a second quality characteristic for calculating at least one battery pack cell (4) based on the analytical data (3a, 3b), or The quality inspection model (2) therein has at least one second quality model, which constitutes at least one second quality characteristic of the at least one battery pack cell (4) based on the analysis data (3a, 3b).

12. The method according to any one of claims 1 to 2, The analytical data (3a, 3b) determined in (S24) further include: - Apply dimensionality reduction method, - Apply transformation methods, and / or - Apply feature extraction to the spectral test data.

13. The method of claim 12, wherein the dimensionality reduction method is t-SNE.

14. The method according to any one of claims 1 to 2, The analytical data determined in (S24) further includes data from the application of the spectral test data: - Grouping, - Classification, - Regularization, - Reproduction, and / or - Filtering.

15. The method according to any one of claims 1 to 2 above, The spectral learning data mentioned therein includes test data and training data.

16. The method according to any one of claims 1 to 2, The quality model (2c) mentioned above is achieved by: - Convolutional Neural Network (CNN). - Deep neural networks, - Linear model, - Bayesian Neural Network (BNN), or - Bayesian Networks (BN) To execute.

17. The method of claim 16, wherein the linear model is a partial least squares (PLS) model.

18. A quality inspection system (1), said quality inspection system being manufactured according to the method of any one of claims 1 to 17.

19. A method for using the quality inspection system (1) according to claim 18, It has the following steps: - Analytical data (3a, 3b) are determined from the spectral test data using the filter mask (2a), wherein the determined analytical data (3a, 3b) comprises at least one subrange of the frequency range of the spectral test data selected by querying the filter mask (2a), wherein the at least one subrange covers up to 50% of the entire frequency range of the spectral test data, and - Calculate at least one quality characteristic (3c) of the at least one battery pack cell (4) based on the analysis data (3a, 3b) using the quality model (2c).

20. The method of claim 19, wherein the at least one subrange covers up to 30% of the entire frequency range of the spectral test data.

Images