Generating training data to train machine learning model applied to control production process

Abstract

A system (10) for generating training data to train a machine learning model (14) applied to a classification task in a production process (13. N) for manufacturing at least one workpiece and / or product, the system comprising a central server (11) and a number of local servers (12.1, 12.2, 12.k) configured to perform: a. Generating a central data set (CD), the method comprises the steps of: a. Providing a central data set (CD) comprising synthetic data points encoded for feature sets and values of tags assigned to each feature set, b. Transmitting copies of the central data set (CD) to a number of local servers (12.1, 12.2, 12.k), c. At each local server (12.1, 12.2, 12.k), transmitting the copies of the central data set (CD) to a plurality of local servers (12.1, 12.2, 12.k) by means of processing the process-specific data sets (PD1, PD2, PD3, PD4, PD5, PD6, PD6, PD6, PD7, PD6, PD7, PD6, PD7, PD6, PD7, at the central server (11), encoding knowledge in the local servers (12.1, 12.2, 12.4) into the features and / or tags of each data point resulting in a process-specific current purified data set to optimize the features and / or tags of each data point of a copy of the central data set (DSk), d. At the central server (11), from the local servers (12.1, 12.2, 12.4, 12.4, 12.4, 12.4, 12.4, 12.4, 12.4, 12.4, 12.4, 12.4, 12.4, 12.4, 12.4, 12.4, 12.4, 12.k) receives a copy of the process-specific current purified data set and synthesizes all of the current purified data sets into a process-agnostic and purified aggregated central data set, e. Repeating steps b-d by applying the aggregated central data set as the central data set until a termination criterion is met, e. Repeating steps b-d by applying the aggregated central data set as the central data set until the termination criterion is met. F. Providing the resulting aggregated central data set (TD) as an output and using it as an input for training a machine learning model (14) for the production process (13. N).

Description

[0001] The present disclosure relates to a computer-implemented method and a corresponding system for generating training data for training a machine learning model for a classification task in a production process for manufacturing at least one workpiece and / or product. In addition, the present disclosure relates to a corresponding computer-implemented method and a computer program product for controlling a production process.

[0002] Artificial intelligence ("AI") systems, such as machine learning models, are known from the prior art. Machine learning models are software programs whose behavior is learned from data rather than being explicitly programmed. The learning process is called "training", which requires large amounts of data and considerable computing resources. Thus, the trained machine learning model solves the specific task for which it was trained, such as the prediction of a machine or process property. Unlike conventional software, the learning behavior of a machine learning model is highly dependent on the data, i.e., the training data, including the parameters used during training.

[0003] In industrial manufacturing, there are highly diverse manufacturing scenarios. For example, for the same production process, there may be different machine types used from many machine manufacturers. The same or similar production process can be performed in different production sites with different machine settings. Similarly, wear of tooling of different machines and other production-related factors, such as changes in the manufacturing process, affect the distribution of collected data. In the result, the distribution of data recorded from different machines may be different. Such a shift in data distribution usually leads to a degradation in the performance of a machine learning model developed for a certain production process.

[0004] Furthermore, while machine learning models perform well on in-domain data, i.e., data similar to the datasets they were trained on, e.g., data from the same production process, operated with the same machine but from another manufacturer, they fail to generalize to out-of-domain data incurred by strong distribution shifts, e.g., data from another production process operated with the same machine. The distribution shift problem is also addressed by collecting data from a new distribution, e.g., from a new manufacturing machine, and labeling it before training a new ML model or fine-tuning an existing ML model thereon. This approach has two drawbacks. First, in most cases, only very little labeled data is available, which is randomly sampled. Few data samples for new, unseen manufacturing scenarios for customizing an adaptive learning model or even more for customizing an untrained learning model result in mediocre performance of the model.

[0005] WO 2021 / 080577 A1 discloses a system for implementing federated learning, which performs the following actions in each round of multiple rounds of model optimization: select a set of one or more clients; each client in the set obtains information describing the global version of the model from a server; each client in the set updates the model based on their local data; the updated model or model update is sent by each client to the server; the server aggregates updates and improves the global model.

[0006] US2021 / 0272014 A1 discloses that data samples are transmitted from a central server to at least one local server device. The central server receives a set of predictions from the at least one local server device, the set of predictions being based on the set of transmitted data samples.

[0007] One way to generalize machine learning models to unseen out-of-domain data is to train the ML model on data from different manufacturing conditions, such as different machine types. This would allow the ML model to be used directly on unseen data collected by unknown machine types without requiring any labeling or retraining. The problem here is that data owners, such as manufacturing companies, are reluctant to share their data in order to protect data privacy and know-how and possibly prevent reverse engineering. Such an approach accelerates the development of ML models, but typically requires high data transfers to different data sources.

[0008] Therefore, the purpose of this application is to train an ML model on data from different manufacturing conditions in such a way that the model can be used directly on unseen data collected from unknown production processes, for example using unknown machine types, without requiring any labeling or retraining in a fast and low memory consumption manner while protecting data privacy.

[0009] This object is solved by the features of the independent claims. The dependent claims contain further developments of the invention.

[0010] A first aspect relates to a computer-implemented method for generating training data for training a machine learning model for a classification task in a production process for manufacturing at least one workpiece and / or product, the method comprising:

[0011] a. generating, on a central server, a central data set comprising synthetic data points encoding values ​​for feature sets and labels assigned to each feature set,

[0012] b. transmitting a copy of the central data set to several local servers, wherein each local server stores a process-specific data set collected at a local production process among several local production processes, and the collected data sets of the local production processes have different data distributions,

[0013] c. at each local server, optimizing the features and / or labels of each data point of its copy of the central dataset by encoding the knowledge contained in its process-specific dataset into the features and / or labels of each data point resulting in a process-specific current distilled dataset,

[0014] d. at the central server, receiving copies of process-specific current cleansed data sets from at least a subset of the local servers and aggregating all current cleansed data sets into a process-agnostic cleansed aggregated central data set,

[0015] e. Repeat steps bd by applying the aggregated central dataset as the central dataset until the termination criteria is met,

[0016] f. Provide the resulting aggregated central dataset as output and use it as input for training a machine learning model for the production process.

[0017] The proposed method applies a federated learning approach at the dataset level rather than at the model level. This is achieved by optimizing a collection of synthetic data points rather than optimizing a machine learning model. By using a federated learning setting, the proposed method does not need direct access to the data of different production processes, i.e., a data domain, where each owner of a production process stores their private data on local servers. These local servers communicate with a central server that aggregates them in a secure manner, for example by aggregating information received from different local servers. Therefore, the proposed method protects data privacy.

[0018] The disadvantage of federated learning, i.e., sending large datasets between a central server and each local would be computationally expensive, is overcome by encoding the knowledge contained in its process-specific dataset into the features and / or labels of each data point resulting in a process-specific current purified dataset. The purified dataset has a reduced size compared to the process-specific dataset. Therefore, the volume of data transmitted for each repetition is significantly reduced. The purified dataset is synthesized into a small number of data points that do not need to be from the correct data distribution, but when given as training data to a machine learning model, will approximate the model trained on the original data. Therefore, the proposed method generates a purified dataset, i.e., a small dataset that encodes the task knowledge, by combining a dataset purification method. The method reduces cost and time because the newly proposed method is faster and has a lower memory footprint.

[0019] In an embodiment, the production process is a new production process, which is similar to a local production process.

[0020] Similarity here means that the data collected at the new production process and the data collected at the local production process have different distributions. For example, the new process is performed by the same machine as in the local production process but manufactures another product, and thus data with different data distribution is collected at the new process. Another example, the new production process manufactures the same product, but uses a machine from another machine manufacturer, so that the data collected at the new production process has a difference (a different) from the data of the local production process.

[0021] In an embodiment, the production process is one of a local production process.

[0022] Using the resulting aggregated central dataset as input for training a machine learning model for the local production process results in an updated ML model for the local production process that provides better results even if the wear of the tooling has changed during operation time or the tool has been replaced by another tool.

[0023] In an embodiment, training is performed on a local server using the resulting aggregated central dataset and the process-specific dataset of the local process.

[0024] Such combination of datasets used to train the local production process emphasizes information about specific features of the local production process related to process-agnostic features in the resulting aggregated central dataset, which optimizes the resulting trained ML model of the local production process.

[0025] In an embodiment, the initial central data set includes data points of random values, or includes a data set of the local production process or any of the production processes, or a cleansed data set of a subset of the local production processes.

[0026] This allows for a variety of different starting datasets. A random dataset provides a uniform and unbiased data distribution about a local or new production process. Applying any local production process or production process dataset or cleaned dataset accelerates reaching the termination criteria.

[0027] According to the invention, the optimization is performed by inputting the process-specific data set and a copy of the central data set into a data set cleansing algorithm.

[0028] Transferring large data sets between the central server and each local server would be computationally expensive. Optimizing the data set through a data set cleansing algorithm reduces the amount of data that must be sent and thus reduces data transfer costs, and reduces processing capacity in the local and central servers.

[0029] In an embodiment, aggregation is performed by averaging the features and / or labels of all current cleansed datasets resulting in an aggregated central dataset.

[0030] The aggregated dataset alone does not reveal any information about the process-specific datasets of different local production processes. The collection of client-specific cleansed datasets, each of which is biased towards a specific data domain / production process, produces a dataset that is domain-agnostic, i.e., not biased towards a specific domain / production process.

[0031] In further embodiments, the data points are characterized by parameters measured on at least one machine performing the production process, or on at least one workpiece and / or product manufactured by the production process.

[0032] In an embodiment, the features of the data points include information of a feature map of an image of the production process.

[0033] The feature map of an image provides the most discriminative information about the objects on the image. Compared with a dataset that encodes the complete image, the data points that only include the information of the feature map have a reduced volume.

[0034] In an embodiment, the process-specific data set of each local production process is private data, and this private data is not transmitted to the central server.

[0035] In an embodiment, the production process is one of an additive manufacturing process or a milling process, and the classification task is one of anomaly detection, fault classification, condition monitoring of a machine in the production process, or quality control, product sorting of workpieces and / or products manufactured in the production process.

[0036] A second aspect relates to a computer-implemented method for controlling a production process for manufacturing at least one workpiece and / or product, comprising:

[0037] - training a machine learning model with an aggregated central dataset generated according to claims 1-10 to perform classification tasks in a production process,

[0038] - Input the data points collected during the production process into the trained machine learning model,

[0039] - outputting classification results of the production process from the trained machine learning model depending on the input data points, and

[0040] -Control the production process depending on the classification results.

[0041] A machine learning model trained by an aggregated central dataset provides high performance with respect to the quality of the classification results in the case of having no or only a small amount of training data for the production process to which it is applied. The high performance of the trained ML model and the accuracy of the output classification results, such as the fault class of a product, are used, for example, to adapt the production process and / or the parameter settings of the machine performing the production process.

[0042] The trained ML model not only has high domain generalization capabilities, but also uses an aggregated central dataset, which is a cleansed cross-domain dataset that encodes the knowledge contained in process-specific datasets from different local production processes that have been used for training. The aggregated central dataset can be useful for reducing the engineering burden of developing different data-driven applications beyond the task at hand, such as applications requiring continuous learning properties or neural architecture search. The advantage is cost reduction and / or performance increase through the reuse of the cleansed cross-domain dataset, i.e., the aggregated central dataset.

[0043] In an embodiment, the classification task is one of anomaly detection, fault classification, condition monitoring of a machine in a production process, or quality control, product sequencing of workpieces and / or products manufactured in the production process.

[0044] A third aspect relates to a system for generating training data for training a machine learning model, the machine learning model being applied to a classification task in a production process for manufacturing at least one workpiece and / or product, the system comprising a central server and a plurality of local servers, configured to execute:

[0045] a. Generate a central dataset, including synthetic data points encoding the feature sets and the values ​​of the labels assigned to each feature set,

[0046] b. transmitting a copy of the central data set to several local servers, wherein each local server stores a process-specific data set collected (measured) at one local production process among several local production processes, and the collected data sets of the local production processes have different data distributions,

[0047] c. at each local server, optimizing the features and / or labels of each data point of its copy of the central dataset by encoding the knowledge contained in its process-specific dataset into the features and / or labels of each data point resulting in a process-specific current cleansed dataset,

[0048] d. at the central server, receiving copies of process-specific current cleansed data sets from at least a subset of the local servers and aggregating all current cleansed data sets into a process-agnostic and cleansed aggregated central data set,

[0049] e. Repeat steps bd by applying the aggregated central dataset as the central dataset until the termination criteria is met,

[0050] f. The resulting aggregated central dataset is provided as output and used as input for training a machine learning model for the production process, where optimization is performed by inputting the process-specific dataset and a copy of the central dataset into a dataset cleansing algorithm.

[0051] The system provides an aggregated central dataset for training ML methods in a fast, processing power and data bandwidth efficient manner.

[0052] A fourth aspect relates to a computer program product directly loadable into the internal memory of a digital computer, comprising software code portions for performing the steps as previously described when said product is run on said digital computer.

[0053] The computer program product may be provided by a computer readable memory medium having computer program instructions stored thereon which, when executed by a processor, perform the steps of the method according to the preceding claim.

[0054] The present invention will be explained in more detail with reference to the accompanying drawings. Similar objects will be marked with the same reference numerals.

[0055] Figure 1 An embodiment of the system of the present invention is schematically illustrated.

[0056] Figure 2 An embodiment of the inventive method for generating training data is shown, which is illustrated by a schematic information flow diagram.

[0057] Figure 3 An embodiment of the method of the invention for controlling a production process is shown which is illustrated by a schematic information flow diagram.

[0058] Note that in the following detailed description of the embodiments, the drawings are schematic only, and the elements shown are not necessarily shown to scale. On the contrary, the drawings are intended to illustrate the functions and collaboration of the components. Here, it will be understood that any connection or coupling of a functional block, device, component or other physical or functional element may also be achieved by indirect connection or coupling, for example, via one or more intermediate elements. The connection or coupling of an element or component or node may be achieved, for example, by a combination based on a wired, wireless connection and / or based on a wired and wireless connection. The functional unit may be implemented by dedicated hardware, such as a processor or firmware, and / or by a combination of dedicated hardware and firmware and software. Further note that each functional step of the method may be performed at a functional unit on a related system.

[0059] Figure 1A system 10 is shown which is configured to carry out the method according to the invention, namely to generate training data TD for training a machine learning model 14 for classification tasks in a production process 13.n for manufacturing at least one workpiece and / or product.

[0060] The system 10 comprises a central server 11 and a number of local servers 12.1, 12.2, 12.k. Each of the local servers 12.1, 12.2, 12.k comprises a data interface for communication with the central server 10 and a storage unit which stores process-specific data sets PD1, PD2, PDk collected at the local production processes 13.1, 13.2, 13.k. The process-specific data sets PD1, PD2, PDk of each local production process 13.1, 13.2, 13.k are private data, i.e. owned by different customers, and throughout the method steps performed by the system, such private data are not transmitted to the central server 11.

[0061] In industrial manufacturing, there are highly diverse manufacturing scenarios. For example, in a production process, there may be different machine types used from many machine manufacturers. A local production process is such a variant of a production process. The data collected from several local production processes have different data distributions. The data or collected data from the local production processes 13.1, 13.2, 13.k are sensor data recorded over time by sensors or image data from a workpiece captured by a camera, wherein the sensors measure physical parameters such as torque, rotation speed, temperature, pressure, voltage, and the like during the production process.

[0062] A data point of a data set comprises a feature and a label assigned to the feature. A feature comprises a collection of different sensor data measured at a certain point in time or a period of time. The label indicates a class characterizing a production process or a machine performing the production process or an artifact manufactured by the production process when the data point is recorded. Examples of classes are the described normal / failure modes, modes of associated maintenance requirements or Failure A, Failure B, Failure C. In this specification, labeled or labeled data is used as a synonym for data for annotation or annotation.

[0063] The central server 11 is configured to generate a central data set CD including synthetic data points in an initial step. In a federated learning method performed in collaboration between the central server and several local servers, the synthetic data points are repeatedly optimized, and finally the central server 11 outputs an aggregated central data set TD, which is used as training data to train a machine learning model 14, which is applied to a classification task in a production process 13.n of manufacturing at least one workpiece and / or product.

[0064] The production process 13.n and the local production processes 13.1, 13.2, 13.k for which the ML model will be trained are processes that collect data sets having the same structure but showing different data distributions. The local production processes 13.1, 13.2, 13.k and the production process 13.n are preferably one of an additive manufacturing process or a milling process. The classification task is one of anomaly detection, fault classification, condition monitoring of a machine in the production process, or quality control of workpieces and / or products manufactured in the production process, product sorting.

[0065] An embodiment of the method of the present invention performed by the central server 11 and each local server 12.1, 12.2, 12.k is described in Figure 2 The following steps are illustrated in FIG.

[0066] In a first step S1, a central data set CD is generated by the central server 11. The central data set CD comprises synthetic data points encoding the values ​​of feature sets and labels assigned to each feature set. The central data set CD comprises data points of random values ​​or a data set of a local production process or any production process in the production process or a cleaned data set of a subset of the local production processes PD1, PD2, PDk.

[0067] In a next step S2, a copy of the central dataset DS is transmitted to each of the local servers 12.1, 12.2, 12.k. At each of the local servers 12.1, 12.2, 12.k, the features and / or labels of each data point of the copy of the central dataset CD are optimized by encoding the knowledge contained in its process-specific dataset PD1, PD2, PDk into the features and / or labels of each data point, resulting in a process-specific current cleansed dataset, see step S3. The optimization is performed by inputting the process-specific datasets and the copy of the central dataset CD into a dataset cleansing algorithm, which outputs a process-specific current cleansed dataset. Examples of possible dataset cleansing algorithms and techniques are published under httpS: / / arxiv.org / abs / 1811.10959, httpS: / / arxiv.org / abs / 2006.05929, https: / / arxiv.org / abs / / 2110.04181 or httpS: / / arxiv.org / abs / 2107.13034.

[0068] In a variant, only the features of the process-specific data sets PD1, PD2, PDk are optimized and the labels remain unchanged. The features of the data points are parameters measured on at least one machine performing the production process or on at least one workpiece and / or product manufactured by the production process. The features of the data points can encode a feature map of the image, which provides the most discriminating information for the objects on the image. The current cleansed data set is sent from each of the local servers 12.1, 12.2, 12.k back to the central server 11.

[0069] The central server 11 receives a copy of the process-specific current cleansed dataset from each of the local servers 12.1, 12.2, 12.k and aggregates all cleansed datasets into a process-agnostic and cleansed aggregated central dataset, see step S4. Preferably, the aggregation is performed by averaging the features and / or labels of all current cleansed datasets resulting in an aggregated central dataset. The aggregation of process-specific cleansed datasets, each of which is biased towards a specific data domain, results in a dataset that is domain-agnostic, i.e. not biased towards a specific production process.

[0070] In step S5, the current aggregated central data set is compared to a termination criterion and further communication rounds are performed until the termination criterion is met. The termination criterion may be one of a predefined convergence criterion, a predefined number of repetitions, a predefined time limit, and the like. If the termination criterion does not match, the aggregated central data set is applied as the central data set and steps S2 to S4 are performed again. That is, in each round, a copy of the previous central data set is replaced by the current cleansed aggregated central data set and sent from the central server 11 to each of the local servers 12.1, 12.2, 12.k. The data cleansing algorithm optimizes the cleansed aggregated central data set using the process-specific data sets PD1, PD2, PDk, resulting in a "new" current cleansed aggregated central data set. The "new" current cleansed aggregated central data set is sent from the local servers 12.1, 12.2, 12.k back to the central server 11, which aggregates all the "new" current cleansed data sets into a process-agnostic and "new" cleansed aggregated central data set. If the “new” cleansed aggregated central dataset meets the termination criteria at step S5 , the “new” current cleansed aggregated central dataset is output and used as input for training the machine learning model 14 of the production process 13 . n .

[0071] The algorithm of the method is shown in pseudo code as follows. Here, consider a scenario where each different manufacturing condition is a different type of manufacturing machine that generates data from a specific domain and saves it on client k, which is a local server k. In the pseudo code, N represents the number of training repetitions and Ncom Indicates the number of communication rounds.

[0072] Algorithm 1: Training data generation for cross-domain joint data purification

[0073]

[0074] In step 10, a dataset purification algorithm is applied. In the following, we provide the pseudo code of a simple and high-performance dataset purification algorithm recently published (https: / / arxiv.org / pdf / 2110.04181.pdf).

[0075]

[0076] After the method in step S5 converges, ie after the optimization / learning of the obtained aggregated central dataset TD as the cleansed cross-domain dataset, the obtained aggregated central dataset TD is used to train the ML model of the production process 14, see step S6.

[0077] The production process may be a new production process, which is similar to the local production process. Alternatively, the production process is one of the local production processes 13.1, 13.2, 13.k, wherein training is performed on the local server 12.1, 12.2, 12.k using the process-specific datasets PD1, PD2, PDk of the local processes 13.1, 13.2, 13.k and the resulting cleaned aggregated central dataset TD.

[0078] The output of the described method is not only the resulting aggregated central dataset TD, but also a trained ML model using the resulting aggregated central dataset as training data. The machine learning model can be a deep learning model, such as a deep autoencoder. But the generated cleansed aggregated central dataset can be used to train any machine learning model, including SVM, decision trees, boosting methods, and more models.

[0079] In fact, the resulting cleansed aggregated central dataset TD contains task-related knowledge from different data domains and can therefore be used during the development of other data-driven applications beyond the task at hand. Other application examples of related tasks include pre-training of deep learning models, continuous learning, neural architecture search, and hyperparameter tuning.

[0080] Another aspect of the invention relates to a method for controlling a production process for manufacturing at least one workpiece and / or product. The ML model and the production process are as described above and in Figure 1 and Figure 2 The method is illustrated in Figure 3 Illustration of the information flow diagram in .

[0081] In a first step S11 of the method, the machine learning model 14 is trained to utilize the Figure 2 The aggregated central data set generated by the method is used to perform classification tasks in the production process 13.n. In a second step S12, the data points collected during the production process 13.n are input into a trained machine learning model 14, which outputs a classification result for the production process 13.n depending on the input data points, see step S13. Subsequently, the production process is controlled depending on the classification result.

[0082] In order to apply our method to a new production process 13.n, it is assumed that the data collected by it contains the same features as the process-specific dataset. In industrial manufacturing, these features are usually recorded using a set of reasonable sensors (including cameras), which measure, for example, the torque of each axis in a milling machine and the control deviation. The collected dataset must be annotated and preprocessed using the same preprocessing steps applied to the process-specific dataset presented on the local server, which was used for dataset purification. See step S6, the application of the model in the new manufacturing scenario consists in deploying the trained ML model in production mode, that is, it makes predictions directly on live data. The live data comes from, for example, a new machine type Mi. The classifier model is initialized using the domain-agnostic parameters θ^* found during model training, see step S6. The class probabilities of live data are predicted using the model θ^*.

[0083] It will be understood that the above description of the examples is intended to be illustrative and that the illustrated components are susceptible to various modifications. For example, the concepts described can be applied to different technical systems and in particular to different subtypes of the corresponding technical systems with only minor adaptations.

Claims

1. A computer-implemented method for generating training data for training a machine learning model (14) for a classification task in a production process (13.n) for manufacturing at least one workpiece and / or product, the method comprising: a. generating (S1) on a central server (11) a central data set (CD) comprising synthetic data points encoding values ​​for feature sets and for labels assigned to each feature set, b. transmitting (S2) a copy of the central data set (CD) to a number of local servers (12.1, 12.2, 12.k), wherein each local server (12.1, 12.2, 12.k) stores a process-specific data set (PD1, PD2, PDk) collected at a local production process (13.1, 13.2, 13.k) among the number of local production processes, and the collected data sets (PD1, PD2, PDk) of the local production processes have different data distributions, c. at each of said local servers (12.1, 12.2, 12.k), optimizing (S3) said features and / or labels of each data point of said copy of said central data set (CD) by encoding the knowledge contained in its process-specific data set (PD1, PD2, PDk) into the features and / or labels of each data point resulting in a process-specific currently cleansed data set, d. at the central server (11), receiving copies of the process-specific current cleansed data sets from at least a subset of the local servers (12.1, 12.2, 12.k) (12.1, 12.212.k) and aggregating all current cleansed data sets into a process-agnostic and cleansed aggregated central data set, e. Repeating (S5) steps bd by applying the cleansed aggregated central dataset as the central dataset until the termination criteria are met, f. Providing (S6) the obtained aggregated central dataset (TD) as output and using it as input for training the machine learning model (14) for the production process (13.n), wherein optimization is performed by inputting the process-specific datasets (PD1, PD2, PDk) and the copies of the central dataset (CD) into a dataset purification algorithm.

2. The method according to claim 1, wherein: The production process (13.n) is a new production process, which is similar to the local production process (13.1, 13.2, 13.k).

3. The method according to claim 1 or 2, wherein: The production process (13.n) is one of the local production processes (13.1, 13.2, 13.k).

4. The method according to claim 3, wherein: The training is performed on the local server (12.1, 12.2, 12.k) using the obtained aggregated central dataset (TD) and the process-specific datasets (PD1, PD2, PDk) of the local processes (13.1, 13.2, 13.k).

5. A method according to any one of the preceding claims, wherein: The initial central data set (CD) includes data points of random values, or includes a data set (PD1, PD2, PDk) of any process in the local production process (13.1, 13.2, 13.k) or a cleansed data set of a subset of the local production process (13.1, 13.2, 13.k).

6. A method according to any one of the preceding claims, wherein: The aggregation is performed by averaging the features and / or labels of all current cleansed datasets resulting in the aggregated central dataset.

7. A method according to any one of the preceding claims, wherein: The data points are characterized by parameters measured on at least one machine performing the production process (13.1, 13.2, 13.k, 13.n) or on at least one workpiece and / or product manufactured by the production process.

8. A method according to any one of the preceding claims, wherein: The characteristics of the data points include information of a feature map of the image of the production process (13.1, 13.2, 13.k, 13.n).

9. A method according to any one of the preceding claims, wherein: The process-specific data sets (PD1, PD2, PDk) of each local production process (13.1, 13.2, 13.k) are private data and are not transmitted to the central server (11).

10. A method according to any one of the preceding claims, wherein: The production process (13.1, 13.2, 13.k, 13.n) is one of an additive manufacturing process or a milling process, and the classification task is anomaly detection, fault classification, condition monitoring of a machine in the production process (13.1, 13.2, 13.k, 13.n), or quality control, product sorting of the workpiece and / or product manufactured in the production process (13.1, 13.2, 13.k, 13.n).

11. A computer-implemented method for controlling a production process (13.n) for manufacturing at least one workpiece and / or product, comprising: - training (S11) a machine learning model (14) with an aggregated central data set (TD) generated according to claims 1-10 to perform classification tasks in said production process (13.n), - inputting (S12) the data points collected during the production process (13.n) into a trained machine learning model (14), - outputting (S13) a classification result of the production process (13.n) from the trained machine learning model (14) depending on the input data points, and - controlling (S14) the production process (13.n) depending on the classification result.

12. The method according to claim 11, wherein: The classification task is one of abnormality detection, fault classification, and condition monitoring of a machine in the production process, or quality control and product classification of workpieces and / or products manufactured in the production process.

13. A system (10) for generating training data for training a machine learning model (14), the machine learning model (14) being applied to a classification task in a production process (13.n) for manufacturing at least one workpiece and / or product, the system comprising a central server (11) and a number of local servers (12.1, 12.2, 12.k), configured to perform: a. generating a central data set (CD) comprising synthetic data points encoding values ​​for feature sets and for labels assigned to each feature set, b. transmitting a copy of the central data set (CD) to the plurality of local servers (12.1, 12.2, 12.k), wherein each local server (12.1, 12.2, 12.k) stores a process-specific data set (PD1, PD2, PDk) collected at a local production process (13.1, 13.2, 13.k) among the plurality of local production processes, and the collected data sets (PD1, PD2, PDk) of the local production processes have different data distributions, c. at each of said local servers (12.1, 12.2, 12.k), optimizing the features and / or labels of each data point of the copy of the central dataset (DSk) by encoding the knowledge contained in its process-specific dataset (PD1, PD2, PDk) into the features and / or labels of each data point resulting in a process-specific currently cleansed dataset, d. at the central server (11), receiving copies of process-specific current cleansed data sets from at least a subset of the local servers (12.1, 12.2, 12.k) and aggregating all current cleansed data sets into a process-agnostic and cleansed aggregated central data set, e. Repeat steps bd by applying the aggregated central dataset as the central dataset until the termination criteria is met, f. providing the resulting aggregated central data set (TD) as output and using it as input to the machine learning model (14) for training the production process (13.n), The optimization is performed by inputting the process-specific data sets (PD1, PD2, PDk) and a copy of the central data set (CD) into a data set cleansing algorithm.

14. A computer program product directly loadable into the internal memory of a digital computer, comprising software code portions for executing the steps of claims 1 to 10 when said product is run on said digital computer.

Images