Method and apparatus for generating training data, for training a machine learning model and for determining a tool state

Abstract

The invention relates to a method for generating training data, a method for training a machine learning model, and a method for determining a tool state and conditioning a grinding tool using a trained machine learning model.

Description

[0001] The invention relates to a method and a device for generating training data, a method and a device for training a machine learning model, and a method and a device for determining a tool state and for conditioning a grinding tool. State of the art

[0002] Grinding processes are an important part of manufacturing technology, ensuring precise and high-quality machining of workpieces. To achieve optimum manufacturing quality and cost, these processes should be specifically tailored to the respective machining task. Numerous factors play a role here, including the selection of the grinding tool, the adjustment of the process parameters, and / or the continuous monitoring of the tool and the process.

[0003] In ongoing manufacturing operations, grinding tools are subject to natural wear, which is dependent on both load and time. This wear manifests itself in various undesirable phenomena such as increased process forces, elevated process temperatures, and increased vibrations. Such changes ultimately lead to a reduction in component quality. To counteract these negative effects and maintain the optimal performance of the grinding tools, they undergo a conditioning process at regular intervals. This process serves to restore the grinding tool to its original, optimal operating condition.

[0004] The practical challenge, however, lies in the fact that the wear behavior of a grinding tool is subject to considerable fluctuations, at least in part. These fluctuations can be caused by different component properties, such as variations in allowances or varying material properties (e.g., hardness). Furthermore, manufacturing-related differences in the grinding tool properties, as well as external factors such as machine vibrations, shocks, and temperature fluctuations, influence the wear behavior.

[0005] To ensure consistently high component quality within the required tolerances despite these fluctuations, grinding tools are often conditioned preventively and at an early stage. The goal is to prevent tool wear from ever reaching a level that could impair the quality of the manufactured components. However, this leads to shorter conditioning intervals than technologically necessary, which negatively impacts the efficiency and productivity of grinding processes in terms of tool consumption and downtime for conditioning operations.

[0006] Overall, the precise tuning and regular conditioning of the grinding tools presents a complex challenge, but it remains essential to achieve the desired manufacturing goals while keeping costs under control.

[0007] It is an object of the invention to specify improved methods and / or improved devices in this respect.

[0008] The problem is solved by a method according to the features of claim 1. The problem is solved by a method according to the features of claim 2. The problem is solved by a method according to the features of claim 4. The problem is solved by a device according to the features of claim 11. The problem is solved by a device according to the features of claim 12. The problem is solved by a device according to the features of claim 13. Disclosure of the invention

[0009] According to a first aspect, a method for generating training data is proposed. Furthermore, a corresponding device is proposed, wherein the device includes an evaluation and computing unit configured to perform the steps of the method. The method comprises the following steps: providing process data for a first grinding operation with a grinding tool, in particular one conditioned immediately before the first grinding operation; providing process data for an Nth grinding operation with the grinding tool, with N > 1 and N ∈ ℕ (element of the set of natural numbers); preprocessing the process data for the first and the Nth grinding operations; extracting features from the preprocessed process data for the first and the Nth grinding operations.Reducing the number of extracted features by applying principal component analysis (PCA) to obtain a PCA feature cluster for the first and Nth grinding operations; comparing the PCA feature cluster obtained for the first grinding operation with the PCA feature cluster for the Nth grinding operation to determine wear of the grinding tool from the process data for the Nth grinding operation, particularly depending on a comparison and / or wear criterion; and if, based on the comparison, it is determined that the grinding tool is worn, providing the PCA feature cluster for the process data of the Nth grinding operation as training data for training a machine learning model to determine a tool condition.

[0010] Using PCA, a large number of individual features can be reduced to a few (e.g., 2 to 3), so-called Principal Components (PC values), without a significant loss of information from the original data.

[0011] To identify the wear behavior of the grinding tool, for example a grinding wheel, during a specific grinding process and to provide a sufficiently large training data set, a large number of grinding processes are preferably recorded metrologically and the resulting process data are processed.

[0012] The grinding processes and their measurement data are preferably ordered according to their chronological sequence, i.e., all first grinding processes directly after dressing, all second grinding processes after dressing, ... all Nth grinding processes after dressing, etc., and the data pre-processed using PCA (PCA feature clusters) are compared with each other. For example, the PCA feature clusters of the first grinding processes are compared with PCA feature clusters of the second grinding processes after dressing. For example, PCA feature clusters of the first grinding processes are compared with PCA feature clusters of the third grinding processes. For example, PCA feature clusters of the first grinding process are generally compared with PCA feature clusters of the Nth grinding process after dressing. PCA feature clusters of the Nth grinding processes are preferably those PCA feature clusters belonging to grinding processes that take place immediately before the next conditioning.Conditioning the grinding tool can, for example, involve dressing.

[0013] With increasing differences between the respective (wear) condition of the grinding tool and the (reference) condition of the grinding tool immediately after dressing (represented by the first grinding process), a clear separation of the PCA feature clusters can be seen.

[0014] Once the grinding tool has reached a stable worn state, the spacing of the PCA feature clusters remains almost constant. The PC values ​​from the PCAs of the grinding processes thus determined with similar wear states (for example, those with a consistently dull grinding tool) are then used to train a machine learning model, in particular an autoencoder.

[0015] According to a second aspect, a method for training a machine learning model is proposed. Furthermore, a corresponding device is proposed, the device comprising an evaluation and computing unit configured to perform the steps of the method. The method comprises the following steps: providing training data containing a multitude of PCA feature clusters, each generated by Principal Component Analysis (PCA) from process data relating to a grinding operation with a grinding tool determined to be worn, or each generated by Principal Component Analysis (PCA) from process data relating to a grinding operation with a grinding tool determined to be unworn; training the machine learning model, in particular an autoencoder, using the training data; and providing the trained machine learning model.Preferably, the multitude of PCA feature clusters is provided as training data by the method according to the first aspect.

[0016] Based on the training data, the machine learning model, especially the autoencoder, can learn a stable wear state of the grinding tool and, in particular during subsequent application, recognize the difference to a different wear state.

[0017] If a grinding tool exhibits a nearly stable wear state (low / minimal wear, i.e., no significant change in condition) for several grinding cycles (e.g., 1 to 4) after the conditioning or dressing process, the machine learning model or autoencoder can alternatively be trained on these "good" states of the grinding tool. As soon as a maximum permissible recursion error is exceeded when applying such a trained model, a conditioning or dressing process is required.

[0018] According to a third aspect, a method for determining a tool condition and conditioning a grinding tool using a machine learning model trained as described herein is proposed. Furthermore, a corresponding device is proposed, wherein the device includes an evaluation and computing unit configured to execute the steps of the method.The procedure comprises the following steps: providing preprocessed process data for a grinding operation with the grinding tool; extracting features from the preprocessed process data; reducing the number of extracted features by applying Principal Component Analysis (PCA) to obtain a PCA feature cluster for the grinding operation; determining a reconstruction error in the reconstruction of the obtained feature cluster using the trained machine learning model to determine the tool condition of the grinding tool; and conditioning the grinding tool if the determined reconstruction error meets a predetermined conditioning criterion.

[0019] The determination of the current wear state of a grinding tool is preferably carried out as follows. A grinding process is measured and reduced to PC values ​​or principal component values ​​using feature extraction (PCA). The PC values ​​from the PCA are preferably fed to the trained machine learning model. This model calculates a recursion error between the new data or PCA feature clusters and the learned wear state.

[0020] The larger the calculated recursion error, the further away the considered wear state is from a dull or worn state of the grinding tool, i.e., the more "unworn" the grinding tool is.

[0021] By defining a limit value, such as a maximum permissible recursion error, the optimal time for a technically necessary conditioning or dressing process can always be determined. Regardless of the number of grinding processes, the machine learning model now recognizes the degree of wear on the grinding tool due to the individual or varying load conditions and can, for example, request a new conditioning or dressing process via a command to the grinding machine and / or an output and / or a warning to a user.

[0022] It is understood that the steps according to the invention, as well as further optional steps, do not necessarily have to be carried out in the sequence shown, but can also be carried out in a different sequence. Furthermore, additional intermediate steps may be provided. The individual steps may also comprise one or more sub-steps without thereby departing from the scope of the method according to the invention.

[0023] The statements made regarding the procedure apply accordingly to the device(s) and vice versa. It is understood that linguistic modifications of procedurally formulated features can be reformulated for the device according to standard linguistic practice, without such formulations needing to be explicitly listed here.

[0024] The methods and / or devices claimed herein enable the detection of individual, time-dependent, and / or situational changes in the operating behavior, particularly the wear state, of grinding tools, and thus the determination of an optimal time for a conditioning process. For this purpose, a target-actual comparison of data acquired from monitoring systems of the grinding tool, e.g., sensors, is performed, particularly after each workpiece has been machined, using machine learning methods. For example, a machine learning model can be trained to recognize the state of a "worn" grinding tool, so that a conditioning process can subsequently be initiated, but always only according to the respective situational, time-dependent, and / or individual need.This reduces unproductive downtime in the grinding process, particularly that caused by conditioning. Ultimately, this is achieved because the intervals between two conditioning cycles can be increased.

[0025] The need to perform a conditioning process is determined by comparing process data, which in particular represent the temporal states of a grinding tool, such as a grinding wheel. Such states can be determined, for example, as "good" or "bad," or even in binary terms as "1" or "0."

[0026] The process data at the time of a “bad” condition of a grinding tool can be used to train a machine learning model, for example an autoencoder or a neural network.

[0027] In the subsequent application or inference phase of the machine learning model, a size of a reconstruction error (in the case of an autoencoder) between actual process data of the grinding tool and process data at the time of a "bad" condition of the grinding tool can preferably be used to determine how far the wear state of the grinding tool is from a need to trigger a conditioning process.

[0028] The monitoring concept thus detects whether a grinding tool is still within a defined target state or range, or whether conditioning is required. This assessment is preferably performed after each grinding operation (i.e., after each machining of a component). This significantly increases effectiveness and / or productivity compared to known non-dynamic conditioning concepts.

[0029] The invention can, in principle, be used for all grinding processes that involve conditioning processes.

[0030] In a further aspect, it is proposed that the process data include sensor data and / or drive data of the grinding tool and / or a grinding machine comprising the grinding tool. It is particularly preferred that the sensor data be acquired by a structure-borne sound and / or vibration sensor and / or a force sensor and / or a sensor for measuring electrical current and / or voltage, especially on drive components of the grinding tool and / or the grinding machine.

[0031] Data sources for the process data are preferably those from sensors and / or the drives of a grinding tool or grinding machine. During a grinding process, signals are preferably acquired and stored at these data sources. The sensors can be structure-borne sound and / or vibration sensors and / or sensors for measuring the electrical current or voltage at the drive components (e.g., grinding spindle) of the grinding machine. Other sensors (e.g., force sensors) can also be used. The sampling rate of the structure-borne sound signal can preferably be up to 2 MHz. The sampling rate of the vibration signal and / or the electrical current can preferably be up to 100 kHz.A structure-borne sound sensor and / or a vibration sensor is / are preferably mounted as close as possible to a contact zone between the grinding tool and a workpiece to be ground, for example in or on a tool spindle, or in or on a workpiece spindle, or on a tailstock. The electrical current can preferably be measured on at least one, preferably all three conductors of the grinding spindle and / or on at least one, preferably all three conductors of the workpiece spindle.

[0032] In another aspect, it is proposed that the preprocessing of the respective process data involves trimming to a predetermined time window length and / or transforming at least part of the process data between a time and a frequency domain.

[0033] The processing of the acquired process data, particularly the sensor data, preferably takes place after each grinding operation or after the raw sensor signals have been saved. First, the respective raw sensor signals or process data are preferably trimmed to one or more time windows, each with a length of, for example, 0.5 to 1.0 seconds, in order to focus on time periods that are of particular interest or highly informative. Since the subsequent feature extraction preferably takes place in both the time domain of the respective (sensor) signal and in the frequency and / or time-frequency domain, a signal transformation can preferably be performed to determine a Fourier spectrum and / or a wavelet spectrum and / or a Hilbert spectrum.

[0034] In another aspect, it is proposed that feature extraction involves extracting statistical values ​​from the process data, where the statistical values ​​have a maximum and / or a minimum and / or a median and / or a standard deviation and / or a kurtosis.

[0035] From the preprocessed process data, such as time windows, statistical values ​​(so-called features), such as maximum, minimum, median, standard deviation, kurtosis, etc., are preferably calculated or extracted. To make the derived values ​​(features) more usable for machine learning algorithms, such as the machine learning model, the extracted features are preferably standardized (e.g., between 0 and 1) and / or subjected to feature cleaning. Feature cleaning can, for example, involve filtering out features without significant variance (very low information content) and / or filtering out features that are highly correlated with each other. The remaining features are then further processed using PCA (Principal Component Analysis).

[0036] In another aspect, a control unit of an industrial machine tool is also claimed, wherein the present method for determining a tool state using a machine learning model trained herein can be executed on the control unit in one of its aspects.

[0037] In another aspect, a computer program is claimed to contain program code capable of executing at least parts of the present method in one of its aspects when the computer program is executed on a computer. In other words, a computer program (product) is claimed to comprise instructions that, when executed by a computer, cause it to execute the method(s) in one of its aspects.

[0038] In a further aspect, a computer-readable data carrier containing the program code of a computer program is proposed to execute at least parts of the present method in one of its aspects when the computer program is executed on a computer. In other words, the invention relates to a computer-readable (storage) medium comprising instructions which, when executed by a computer, cause it to execute the method / steps of the method in one of its aspects.

[0039] The described configurations and training programs can be combined in any way desired.

[0040] Further possible embodiments, developments and implementations of the invention also include combinations of features of the invention described previously or subsequently with regard to the exemplary embodiments that are not explicitly mentioned. Brief description of the drawings

[0041] The accompanying drawings are intended to provide a further understanding of the embodiments of the invention. They illustrate embodiments and, in conjunction with the description, serve to explain the principles and concepts of the invention.

[0042] Other embodiments and many of the aforementioned advantages become apparent with reference to the drawings. The elements depicted in the drawings are not necessarily shown to scale. Fig. Figure 1 shows a schematic flowchart of an exemplary embodiment of the present method. Fig. Figure 2 shows a schematic flowchart of an exemplary embodiment of the present method. Fig. Figure 3 shows a schematic flowchart of an exemplary embodiment of the present method. Fig. Figures 4(a)-(c) show exemplary PCA feature cluster comparisons for determining wear of a grinding tool. Fig. Figure 5 shows a schematic progression of a reconstruction error with an increasing number of grinding processes after a final conditioning process of a grinding tool.

[0043] In the figures of the drawings, identical reference symbols denote identical or functionally equivalent elements, parts or components, unless otherwise stated.

[0044] The respective method can be carried out in any embodiment, at least partially, by a device 100, which may comprise several components not shown in detail, for example, one or more provisioning units and / or at least one evaluation and computing unit. It is understood that the provisioning unit may be designed together with the evaluation and computing unit, or it may be different from it. Furthermore, the device 100, which may be part of a system, may comprise a storage unit and / or an output unit and / or a display unit and / or an input unit.

[0045] Fig. Figure 1 shows a schematic flowchart of a computer-implemented procedure for generating training data.

[0046] In step S1, process data for a first grinding operation is provided using a grinding tool, in particular one that has been conditioned immediately before the first grinding operation.

[0047] In step S2, process data for an Nth grinding operation with the grinding tool is provided, with N > 1 and N ∈ ℕ.

[0048] In step S3, the process data for the first and the Nth grinding process is preprocessed.

[0049] In step S4, features are extracted from the pre-processed process data for the first and the Nth grinding process.

[0050] In step S5, a reduction of the extracted features is carried out by applying a Principal Component Analysis (PCA) to obtain a respective PCA feature cluster for the first and the Nth grinding process.

[0051] In step S6, the PCA feature cluster obtained for the first grinding process is compared with the PCA feature cluster for the Nth grinding process to determine wear of the grinding tool from the process data for the Nth grinding process, particularly depending on a comparison and / or wear criterion.

[0052] If, based on the comparison, it is determined that the grinding tool is considered worn, in step S7 the PCA feature cluster is provided to the process data for the Nth grinding process as training data for training a machine learning model to determine a tool condition.

[0053] Fig. Figure 2 shows a schematic flowchart of a computer-implemented procedure for training a machine learning model.

[0054] In step S10, training data is provided that exhibits a large number of PCA feature clusters, each generated by Principal Component Analysis (PCA) from process data of a grinding operation with a grinding tool that was determined to be worn, or each generated by Principal Component Analysis (PCA) from process data of a grinding operation with a grinding tool that was determined to be not worn.

[0055] In step S20, the machine learning model, in particular an autoencoder, is trained using the training data.

[0056] In step S30, the trained machine learning model is deployed.

[0057] Fig. Figure 3 shows a schematic flowchart of a computer-implemented procedure for determining a tool state and conditioning a grinding tool using a trained machine learning model.

[0058] In step S100, pre-processed process data for a grinding operation with the grinding tool is provided.

[0059] In step S200, features are extracted from the pre-processed process data.

[0060] In step S300, a reduction in the number of extracted features is performed by applying a Principal Component Analysis (PCA) to obtain a PCA feature cluster for the grinding process.

[0061] In step S400, a reconstruction error is identified during the reconstruction of the obtained feature cluster using the trained machine learning model to determine the tool condition of the grinding tool.

[0062] In step S500, the grinding tool is conditioned if the determined reconstruction error meets a predetermined conditioning criterion.

[0063] Fig. Figures 4(a)-(c) show exemplary PCA feature cluster comparisons for determining wear of a grinding tool using a trained machine learning model, in particular an autoencoder.

[0064] Fig. Figure 4(a) shows a PCA feature cluster 400, which was determined based on several measurement data series for a first grinding process after a previous conditioning or dressing process, compared to a PCA feature cluster 402, which was determined based on several measurement data series for a second grinding process following the last conditioning or dressing process and occurring after the first grinding process. The two PCA feature clusters 400 and 402 are almost contiguous.

[0065] Fig. Figure 4(b) shows a PCA feature cluster 400, which was determined based on several measurement data series for a first grinding process after a previous conditioning or dressing process, compared to a PCA feature cluster 404, which was determined based on several measurement data series for an i-th grinding process occurring after the last conditioning or dressing process, with i > 2. The two PCA feature clusters 400 and 404 are already further apart.

[0066] Fig. Figure 4(c) shows a PCA feature cluster 400, which was determined based on several measurement data series for a first grinding process after a previous conditioning or dressing process, in comparison to a PCA feature cluster 406, which was determined based on several measurement data series for an Nth grinding process following the last conditioning or dressing process, with N >> 2. The two PCA feature clusters 400 and 406 are now significantly different.

[0067] Fig.Figure 5 shows an example of an increase in a recursion error 500 as the considered grinding process is closer to the last conditioning or dressing process. An example recursion error curve 502 of a trained machine learning model based on PCA feature cluster results from grinding process data is shown. The recursion error 500 decreases steadily with an increasing number N of grinding processes, starting from the last conditioning or dressing process.

Claims

[1] Method for generating training data, the method comprising the steps: - Providing (S1) process data for a first grinding operation with a grinding tool, in particular one conditioned immediately before the first grinding operation; - Providing (S2) process data for an Nth grinding operation with the grinding tool, with N > 1 and N ∈ ℕ; - Preprocessing (S3) of the process data for the first and the Nth grinding process; - Extracting (S4) features from the preprocessed process data for the first and the Nth grinding process; - Reducing (S5) a respective number of extracted features by applying a Principal Component Analysis (PCA) to obtain a respective PCA feature cluster (400, 402, 404, 406) for the first and the Nth grinding operation; - Comparing (S6) the obtained PCA feature cluster (400) for the first grinding operation with the PCA feature cluster (402, 404, 406) for the Nth grinding operation to determine wear of the grinding tool from the process data for the Nth grinding operation, in particular depending on a comparison and / or wear criterion; and - if, based on the comparison (S6), it is determined that the grinding tool is considered worn, provide (S7) the PCA feature cluster (406) to the process data for the Nth grinding operation as training data for training a machine learning model to determine a tool condition. [2] Method for training a machine learning model, the method comprising the steps: - Providing (S10) training data that includes a variety of PCA feature clusters (406) that were each generated by Principal Component Analysis (PCA) to process data for a grinding operation with a grinding tool that was determined to be worn, or that includes a variety of PCA feature clusters (400) that were each generated by PCA to process data for a grinding operation with a grinding tool that was determined to be not worn; - Training (S20) the machine learning model, in particular an autoencoder, using the training data; and - Deployment (S30) of the trained machine learning model. [3] Method according to claim 2, wherein the plurality of PCA feature clusters (406) are each provided as training data by the method according to claim 1. [4] Method for determining a tool condition and for conditioning a grinding tool using a machine learning model trained, in particular according to claim 2 or 3, the method comprising the steps: - Providing (S100) pre-processed process data for a grinding operation with the grinding tool; - Extracting (S200) features from the preprocessed process data; - Reducing (S300) a number of the extracted features by applying a Principal Component Analysis (PCA) to obtain a PCA feature cluster for the grinding process; - Determining (S400) a reconstruction error (500) in the reconstruction of the obtained feature cluster using the trained machine learning model to determine the tool condition of the grinding tool; and - Conditioning (S500) of the grinding tool if the determined reconstruction error (500) meets a predetermined conditioning criterion. [5] Method according to one of the preceding claims, wherein the process data comprise sensor data and / or drive data of the grinding tool and / or a grinding machine comprising the grinding tool. [6] Method according to one of the preceding claims, wherein the sensor data are acquired by a structure-borne sound and / or vibration sensor and / or a force measuring sensor and / or a sensor for measuring an electric current and / or voltage, in particular on drive components of the grinding tool and / or the grinding machine. [7] Method according to one of the preceding claims, wherein the preprocessing of the respective process data comprises cutting to a predetermined time window length and / or transforming between a time and a frequency domain of at least part of the process data. [8] Method according to any of the preceding claims, wherein the feature extraction comprises extracting statistical values ​​from the process data, wherein the statistical values ​​have a maximum and / or a minimum and / or a median and / or a standard deviation and / or a kurtosis. [9] Computer program with program code to execute at least parts of a method according to any one of claims 1 to 8 when the computer program is executed on a computer. [10] Computer-readable data carrier containing program code of a computer program for executing at least parts of a method according to any one of claims 1 to 8 when the computer program is executed on a computer. [11] Device (100) for generating training data, wherein the device (100) has an evaluation and computing unit configured to perform the following steps: - Providing process data for a first grinding operation with a grinding tool, in particular one conditioned immediately before the first grinding operation; - Providing process data for an Nth grinding operation with the grinding tool, with N > 1 and N ∈ ℕ; - Pre-processing of the process data for the first and the Nth grinding process; - Extracting features from the pre-processed process data for the first and the Nth grinding process; - Reducing a respective number of extracted features by applying a Principal Component Analysis (PCA) to obtain a respective PCA feature cluster (400, 402, 404, 406) for the first and the Nth grinding process; - Comparing the obtained PCA feature cluster (400) for the first grinding operation with the PCA feature cluster (402, 404, 406) for the Nth grinding operation to determine wear of the grinding tool from the process data for the Nth grinding operation, in particular depending on a comparison and / or wear criterion; and - if, based on the comparison, it is determined that the grinding tool is considered worn, provide the PCA feature cluster (406) to the process data for the Nth grinding operation as training data for training a machine learning model to determine a tool condition. [12] Device (100) for training a machine learning model, wherein the device (100) has an evaluation and computing unit configured to perform the following steps: - Providing training data that includes a variety of PCA feature clusters (406) each generated by Principal Component Analysis (PCA) to process data for a grinding operation with a grinding tool determined to be worn, or that includes a variety of PCA feature clusters (400) each generated by Principal Component Analysis (PCA) to process data for a grinding operation with a grinding tool determined to be not worn; - Training the machine learning model, in particular an autoencoder, using the training data; and - Deploying the trained machine learning model. [13] Device (100) for determining a tool state and conditioning a grinding tool using a trained machine learning model, wherein the device (100) has an evaluation and computing unit configured to perform the following steps: - Providing pre-processed process data for a grinding operation with the grinding tool; - Extracting features from the pre-processed process data; - Reducing the number of extracted features by applying a Principal Component Analysis (PCA) to obtain a PCA feature cluster for the grinding process; - Determining a reconstruction error (500) in the reconstruction of the obtained feature cluster using the trained machine learning model to determine the tool condition of the grinding tool; and - Conditioning the grinding tool if the determined reconstruction error (500) meets a predetermined conditioning criterion.

Images