DE102023200348A1 - Method for training a predictive model to monitor a conditioning process - Google Patents

Abstract

According to various embodiments, a method for training a predictive model is described, comprising obtaining time series sensor data of a conditioning process, labeling the time series sensor data by labels, wherein the labels indicate sensor data elements of the time series sensor data that reflect that a quality criterion has been met in the conditioning process, training a generative model using the labeled time series sensor data to generate additional time series sensor data of the conditioning process for labels of the additional time series sensor data, wherein the labels indicate sensor data elements of the additional time series sensor data that reflect that a quality criterion has been met in the conditioning process, and training a predictive model using the labeled time series sensor data, the additional time series sensor data, and the labels of the additional time series sensor data by supervised learning.

Description

State of the art

The present disclosure relates to apparatus and methods for training a predictive model to monitor a conditioning process.

Grinding tools used in production processes are subject to wear. Thus, they must be conditioned (or dressed) regularly, i.e. a conditioning process must be carried out. To avoid missing the point in time when the grinding tool has been sufficiently conditioned and the conditioning can be stopped (or even should be stopped to avoid wasting grinding tool material), the conditioning process itself should be monitored. Accordingly, effective procedures for monitoring a conditioning process are desired.

Disclosure of the invention

According to various embodiments, a method is provided for training a predictive model for monitoring a conditioning process, comprising obtaining time series sensor data of a conditioning process, labeling the time series sensor data by labels, wherein the labels indicate sensor data elements of the time series sensor data that reflect that a quality criterion has been met in the conditioning process, training a generative model using the labeled time series sensor data to generate additional time series sensor data of the conditioning process for labels of the additional time series sensor data, wherein the labels indicate sensor data elements of the additional time series sensor data that reflect that a quality criterion has been met in the conditioning process, and training a predictive model for monitoring the conditioning process using the labeled time series sensor data, the additional time series sensor data, and the labels of the additional time series sensor data by supervised learning.

Generating additional time series sensor data for labels of the additional time series sensor data, i.e. generating additional sensor data items, each sensor data item being generated for a specific label, may be achieved by training a conditioned type of regenerative model such that the additional time series sensor data can be conditioned to the required specification, e.g., the time at which the conditioning should be stopped (e.g., the stroke at which the conditioning should be stopped).

According to the method described above, real sensor data are enriched by artificial sensor data generated by a generative model. Thus, training data sets with sufficiently representative properties can be provided, which enables meeting the needs of the training procedure used to train the predictor.

Various examples are given below.

Example 1 is the procedure as described above.

Example 2 is the method of Example 1, wherein the predictive model is trained to indicate for an input sensor data item whether the input sensor data item reflects that the quality criterion has been met in the conditioning process from one or more features extracted from the input sensor data item and input to the predictive model.

Using features (e.g. statistical features) as input to the prediction model instead of full sensor datasets reduces the size of the prediction model while maintaining predictive capabilities.

Example 3 is the method of example 1 or 2, wherein training the generative model using the labeled time series sensor data comprises extracting one or more features of the sensor data item and training the generative model using the one or more features and the label of the sensor data item for each sensor data item of the time series sensor data.

Example 4 is the method of any of Examples 1 to 3, which includes training a plurality of prediction models to monitor the conditioning process using the labeled time series sensor data, the additional time series sensor data, and the labels of the additional time series sensor data by supervised learning using different proportions of labeled time series sensor data and additional time series sensor data, evaluating the trained plurality of prediction models, and selecting one of several prediction models as a prediction model depending on the rating.

In other words, multiple prediction models are trained, each trained with a respective proportion (i.e., ratio) of real to artificial sensor data (i.e., the number of sensor data items holding measured sensor data to the number of sensor data items holding sensor data generated by the generative model). This allows determining optimal training in terms of the proportion of real to artificial sensor data. Finally, a more effective prediction model can be provided compared to using only the real sensor data for training.

Example 5 is a method of controlling a conditioning process, comprising training a predictive model according to any one of Examples 1 to 4, obtaining time series sensor data of a conditioning process to be observed comprising a series of sensor data items, inputting information about the sensor data item to the predictive model for each of the sensor data items, and stopping the conditioning process in response to the predictive model indicating for a sensor data item of the sensor data items that the quality criterion has been met in the conditioning process.

Example 6 is a conditioning process control system configured to perform a method according to any of Examples 1 to 5.

Example 7 is a computer program comprising instructions that, when executed by a computer, cause the computer to perform a method according to any of Examples 1 to 5.

Example 8 is a computer-readable medium comprising instructions that, when executed by a computer, cause the computer to perform a method according to any of Examples 1 to 5.

• 1 das Konditionierung eines Schleifwerkzeugs;
• 2 einen Ablaufplan, der eine Prozedur für die Konditionierungssteuerung darstellt;
• 3 ein typisches Signal von einem AE-Sensor (Sensor für akustische Emission);
• 4 eine Signalanalyseprozedur, die auf die Sensordaten aus 3 angewendet wird;
• 5 ein Beispiel künstlicher Sensordaten, die durch ein generatives Modell erzeugt werden, (auch als digitaler Zwilling bezeichnet);
• 6 ein Beispiel des Einflusses des Verhältnisses realer zu künstlichen Daten in einem Trainingsdatensatz auf die Leistungsfähigkeit eines Vorhersagemodells;
• 7 Ergebnisse eines vereinfachten Anwendungsbeispiels zum Überwachen und Steuern eines Konditionierungsprozesses; und
• 8 einen Ablaufplan, der ein Verfahren zum Trainieren eines Vorhersagemodells zum Überwachen eines Konditionierungsprozesses gemäß einer Ausführungsform darstellt.

In the drawings, like reference characters generally refer to the same parts throughout the several views. The drawings are not necessarily to scale, emphasis generally being placed upon illustrating the principles of the invention. In the following description, various aspects are described with reference to the following drawings, in which:

• 1 the conditioning of a grinding tool;
• 2 a flow chart representing a procedure for conditioning control;
• 3 a typical signal from an AE sensor (acoustic emission sensor);
• 4 a signal analysis procedure based on the sensor data from 3 is applied;
• 5 an example of artificial sensor data generated by a generative model (also called a digital twin);
• 6 an example of the influence of the ratio of real to artificial data in a training dataset on the performance of a predictive model;
• 7 Results of a simplified application example for monitoring and controlling a conditioning process; and
• 8th a flowchart illustrating a method for training a predictive model to monitor a conditioning process according to an embodiment.

The following detailed description refers to the accompanying drawings which show, by way of example, specific details and aspects of this disclosure in which the invention may be practiced. Other aspects may be utilized and structural, logical, and electrical changes may be made without departing from the scope of the invention. Since some aspects of this disclosure may be combined with one or more other aspects of this disclosure to form new aspects, the various aspects of this disclosure are not necessarily mutually exclusive.

Various examples are described in more detail below.

The grinding of components (as defined by DIN 8589, for example) is divided into different variants. The grinding process is characterized by the use of geometrically undefined cutting edges in order to process components with partially complex geometries and high hardness (e.g. > 55 HRc) with high precision.

Grinding processes must always be specifically adjusted to the respective machining task in such a way that the optimum production quality and the optimum costs are achieved. The grinding tools used are generally subject to load- and time-dependent wear during continuous production operations. This wear can be caused by fluctuating component properties such as dimensional changes or different material properties (e.g. hardness). In addition, the progression of wear causes, for example, higher process forces, process temperatures and/or vibrations. To ensure consistent quality of the manufactured product, grinding tools are conditioned at regular intervals, which is colloquially referred to as dressing.

1 represents the conditioning of a grinding tool (or dressing tool) 101, in this example a grinding wheel.

Conditioning is a generic term for the preparation of grinding tools. The generic term includes the processes of dressing and cleaning, with dressing being further divided into profiling and sharpening. The aim is to bring the grinding tool 101 into a defined state (i.e. such that it meets a certain quality criterion) which enables the grinding-related processing of the product within the required tolerances. For this purpose, any tool geometry errors must be eliminated, the required cutting ability must be generated and any smearing must be removed from the pore spaces of the grinding wheel 101. The conditioning process is carried out with a separate tool 102, the conditioning tool (or dressing tool), and usually occurs in a sequence of strokes.

In the example from 1 the conditioning tool 102 is a surface plate, but it may also be another fixed dressing tool (e.g. a single-grain or multi-grain dressing tool) or a rotating dressing tool (such as a dressing pot or a dressing roller).

• • Feste Abrichtintervalle, üblicherweise eine Anzahl bearbeiteter Werkstücke, nach denen das Schleifwerkzeug durch Ausführen einer festen Anzahl von Konditionierungswiederholungen konditioniert wird. Sowohl das Intervall, das auf der Anzahl der Werkstücke beruht, als auch die Anzahl der Konditionierungswiederholungen (z. B. Konditionierungs-„Hübe“), die ausgeführt werden, beruhen auf der Erfahrung eines Prozessexperten, die so weit wie möglich alle Schwankungen von dem Prozess, von dem Werkstückmaterial, von Vorprozessen usw. konservativ ausgleichend umfasst. Der Typ und der Bereich der Einstellungsparametervariablen beruhen auf der Erfahrung des jeweiligen Prozessexperten. Falls der Prozessexperte beschränkte Erfahrung hat, z. B. bei der Inbetriebnahme einer neuen Serie von Schleifprozessen, kann der Prozess ineffizient sein oder kann der Einstellprozess zeitaufwendig sein. Außerdem müssen selbst im Fall kleiner Änderungen der Serienprozesse eine große Anzahl von Experimenten erneut ausgeführt werden, indem z. B. die Materialeigenschaften oder das Schneidvolumen geändert werden, um zu möglichen, dass der Prozessexperte den Konditionierungsprozess optimal einstellt.
• • Zusätzlich zur Erfahrung eines Experten können (lineare oder nichtlineare) Regressionsmodelle verwendet werden, die mögliche Ursachen einer geänderten Konditionierungsqualität auf der Grundlage historischer Daten identifizieren können. In diesen Regressionsmodellen werden üblicherweise nur eindimensionale Beziehungen betrachtet und werden Wechselwirkungen zwischen den verschiedenen Parametern und Einflussfaktoren selten untersucht. Darüber hinaus erfordern entsprechende Regressionsmodelle eine umfassende Datenbank, die eine große Vielfalt verschiedener Abrichtsituationen umfasst.
• • Verfahren des überwachten oder unüberwachten maschinellen Lernens (ML), die historische Daten der Messung des Konditionierungsprozesses betrachten und mögliche Ursachen einer geänderten Konditionierungsqualität identifizieren können. Diese Verfahren können lineare oder nichtlineare Abhängigkeiten modellieren und werden unter Verwendung eines realen Datensatzes von Beobachtungen mit beschränkter Größe trainiert, die nicht eine Überfülle möglicher Varianten der Messung eines realen Prozesses umfassen.
• • Die Spezifikation von Prozess- und/oder Qualitätsgrenzwerten (z. B. Prozess- und Konditionierungskräften, Leistung, Temperatur, Fluidströmung oder Qualitätsparametern wie etwa maximale Oberflächenrauheit, obere und untere Komponentendimension usw.), die mittels technischer Systeme (z. B. Sensoren oder Messvorrichtungen) bestimmt werden, werden überwacht (Einhüllendenkurvenprinzip). Falls einer oder mehrere dieser Prozess- oder Qualitätsgrenzwerte verletzt werden, wird ein automatisch oder manuell initiierter Konditionierungsprozess ausgeführt. In diesem Fall kann der Konditionierungsprozess entweder auf der Grundlage der Erfahrung oder variabel auf der Grundlage der Empfehlung des technischen Überwachungssystems ausgeführt werden. Da eine klare Beschreibung und Überwachung von Grenzwerten, insbesondere für komplexe Schleifwerkzeuggeometrien, schwierig sein kann, ist die Definition der Prozess- und/oder Qualitätsgrenzwerte in dem Konditionierungsprozess herausfordernd. Folglich sind eine Vielzahl von Experimenten erforderlich, um die Grenzwerte zu identifizieren.
• • In Forschungsumgebungen wird eine optische Beobachtung (wie etwa Messung von Graustufen, topographischen Daten oder Rasterelektronenmikroskopie) des Schleifwerkzeugprofils ausgeführt, um feine topographische Informationen der Schleifschicht zu extrahieren. Diese Technologie liefert eine gute Darstellung der realen Topographie, ist aber im Allgemeinen wegen des hohen Messaufwands nicht prozessfähig.
• • Außerdem wird nur in Forschungsumgebungen eine physikalische Modellierung von Abrichtverfahren ausgeführt, um die Schleifschicht nach dem Abrichten in Abhängigkeit von bestimmten Abrichtparametern vorherzusagen. Da diese Verfahren die Körner auf einer mikroskopischen Ebene modellieren, sind sie einerseits sehr komplex. Andererseits sind die Ergebnisse bisher nicht im industriellen Schleifenprozess anwendbar, da die Eingangsvariablen (wie etwa die Topographie, der genaue Verschleißzustand sowohl des Schleif- als auch des Abrichtwerkzeugs usw.) nicht genau bekannt sind.

The conditioning process is a secondary process that can be time-consuming, which reduces productivity and increases costs. It is therefore important to manage the conditioning process optimally and reduce its frequency and duration to a minimum. In industrial production, various concepts for carrying out and monitoring the conditioning process are established. These are as follows:

• • Fixed dressing intervals, typically a number of workpieces machined after which the grinding tool is conditioned by performing a fixed number of conditioning iterations. Both the interval based on the number of workpieces and the number of conditioning iterations (e.g. conditioning “strokes”) performed are based on the experience of a process expert, which conservatively compensates as far as possible for all variations in the process, workpiece material, previous processes, etc. The type and range of setting parameter variables are based on the experience of the specific process expert. If the process expert has limited experience, e.g. when starting up a new series of grinding processes, the process may be inefficient or the setting process may be time consuming. In addition, even in the case of small changes to the series processes, a large number of experiments must be re-run, e.g. by changing the material properties or cutting volume, to ensure that the process expert optimally sets the conditioning process.
• • In addition to the experience of an expert, regression models (linear or non-linear) can be used, which can identify possible causes of a changed conditioning quality based on historical data. In these regression models, usually only one-dimensional relationships are considered and interactions between the various parameters and influencing factors are rarely investigated. In addition, corresponding regression models require a comprehensive database that includes a wide variety of different dressing situations.
• • Supervised or unsupervised machine learning (ML) methods that can look at historical data of the measurement of the conditioning process and identify possible causes of a change in conditioning quality. These methods can model linear or nonlinear dependencies and are trained using a real dataset of observations of limited size that do not include a profusion of possible variants of the measurement of a real process.
• • The specification of process and/or quality limits (e.g. process and conditioning forces, power, temperature, fluid flow or quality parameters such as maximum surface roughness, upper and lower component dimensions, etc.) determined by technical systems (e.g. sensors or measuring devices) are monitored (envelope curve principle). If one or more of these process or quality limits are violated, an automatically or manually initiated conditioning process is carried out. In this case, the conditioning process can be carried out either on the basis of experience or variably on the basis of the recommendation of the technical monitoring system. Since While a clear description and monitoring of limit values can be difficult, especially for complex grinding tool geometries, the definition of process and/or quality limits in the conditioning process is challenging. Consequently, a large number of experiments are required to identify the limit values.
• • In research environments, optical observation (such as grayscale measurement, topographic data or scanning electron microscopy) of the grinding tool profile is performed to extract fine topographic information of the grinding layer. This technology provides a good representation of the real topography, but is generally not process-capable due to the high measurement effort.
• • In addition, physical modelling of dressing processes is only carried out in research environments to predict the grinding layer after dressing depending on certain dressing parameters. On the one hand, since these processes model the grains at a microscopic level, they are very complex. On the other hand, the results are not yet applicable to the industrial grinding process because the input variables (such as the topography, the exact wear state of both the grinding and dressing tools, etc.) are not precisely known.

The above concepts for executing and monitoring the conditioning process to correct wear-related and situation-related process result changes are intended to repeat the conditioning process with a defined frequency based on the knowledge of process experts or process and/or quality limits. However, especially for complex grinding tool geometries, changing grinding tool wear behavior or the wear of the conditioning feature, the above concepts for executing and monitoring the conditioning process reach their technological limits due to their rigid implementation, i.e. economic savings potential remains unused.

According to various embodiments, approaches are described for improving the efficiency of the conditioning process state detection and thus of the conditioning process itself in case of a situation- and/or time-related change in grinding tool wear.

More specifically, according to various embodiments, a monitoring system 103 (including one or more sensors) is provided and a controller 104 determines whether the conditioning process should be stopped using sensor data provided by the monitoring system 103.

• • die Verringerung notwendiger Abrichthübe, was:
  • ◯ die Ausfallzeit verringert und die Produktivität des jeweiligen Bearbeitungsprozesses erhöht;
  • ◯ die Verwendung der Schleifwerkzeugs-Schleifschicht erhöht;
  • ◯ den Abrichtwerkzeugverschleiß verringert.
• • flexible Reaktion auf Änderungen in der Konditionierungssituation (z. B. auf Änderungen der Materialeigenschaften des Schleifwerkzeugs) und ermöglicht somit, dass der Konditionierungsprozess individuell ausgeführt wird.

For example, the control unit 104 performs a target-actual comparison of the sensor data recorded by the monitoring system 103, e.g. after each conditioning stroke, to determine whether the grinding tool surface meets previously defined requirements. This enables:

• • the reduction of necessary dressing strokes, which:
  • ◯ reduces downtime and increases the productivity of the respective machining process;
  • ◯ increases the use of the grinding tool abrasive layer;
  • ◯ reduces dressing tool wear.
• • flexible response to changes in the conditioning situation (e.g. changes in the material properties of the grinding tool) and thus enables the conditioning process to be carried out individually.

According to various embodiments, the controller implements a machine learning model 105, hereinafter also referred to as a predictor or prediction model, which receives sensor data obtained by the monitoring system 103 and determines (“predicts”) from this input whether a certain quality criterion is met (i.e., whether dressing can be stopped).

Furthermore, according to various embodiments, a generative machine learning model 106 generates training data for training the prediction model 105, also referred to as a “digital twin” for the conditioning process, i.e., the generative machine learning model 106 is used to generate training data which, as indicated by the dashed arrow in 1 specified, can be used to train the prediction model 105.

The generative machine learning model 106 is itself trained using recorded data from the monitoring system 103 (and/or from one or more other monitoring systems) using generative AI training techniques. This in turn enables the extensive generation of conditioning process (monitoring sensor) data (e.g., signal trends) to train the predictor 105 as if a conditioning process were occurring in the digital space. This makes it possible to circumvent inefficiencies of a real-world dataset, mainly its limited variability. Various families of ML-supervised methods for training the predictor 105 take advantage of this increased variability and size of the training dataset. This leads to better training and thus better performance of the predictor 105 (i.e., the conditioning process state detector).

For example, the control unit 104 evaluates the grinding tool state after each individual conditioning stroke using the predictor 105 and the specification of a termination criterion (which typically includes a quality criterion) for terminating the conditioning process. The conditioning process may be controlled by a computer-based algorithm (e.g., implemented at least partially in the control unit) that includes recording and evaluating, which may in particular include a specification of the termination criterion (e.g., quality criterion). The algorithm processes signals from the conditioning process to make a prediction using the predictor 105. Signal sources, i.e., sensors of the monitoring system 103, for providing sensor data used by this algorithm may be sensors for providing acoustic emission signals, process forces, vibration signals such as electrical power, electrical current or electrical voltage from drive components, or other signals for monitoring the conditioning process.

The control unit 105 may issue a conditioning process stop signal after each conditioning stroke. Depending on the type of action and on the maturity of the conditioning process monitoring based on the algorithm (e.g. with regard to the training of the predictor 105), any controls by a process expert may be omitted entirely. In addition to the measurement resolution of the sensors, this depends on the learning evaluation of the sensor data in correlation with existing quality data.

• • Während des Konditionierungsprozesses werden Prozesssignale (d. h. Sensordaten) über mehrere Sensoren (z. B. des Überwachungssystems 103 oder eines anderen Überwachungssystems) aufgezeichnet.
• • Die Prozesssignale werden digitalisiert, vorverarbeitet und daraufhin in charakteristische Merkmale umgesetzt (wobei z. B. durch die Steuereinheit 104 eine entsprechende Vorverarbeitung ausgeführt wird).
• • Diese Merkmale dienen als ein Bild des Konditionierungsprozesses und ein durch überwachtes Lernen trainiertes Vorhersagemodell (z. B. der Prädiktor 105) macht aus diesen Merkmalen eine Aussage (Vorhersage), ob der Konditionierungsprozess fortgesetzt oder angehalten werden sollte.

According to various embodiments, the following is carried out:

• • During the conditioning process, process signals (ie, sensor data) are recorded via multiple sensors (e.g., the monitoring system 103 or another monitoring system).
• • The process signals are digitized, preprocessed and then converted into characteristic features (whereby, for example, appropriate preprocessing is carried out by the control unit 104).
• • These features serve as a picture of the conditioning process, and a prediction model trained by supervised learning (e.g., predictor 105) uses these features to make a statement (prediction) as to whether the conditioning process should be continued or stopped.

More specifically, a procedure or algorithm for conditioning process monitoring and conditioning process control according to various embodiments is based on the use of statistical or ML methods in three phases as described in 2 are shown.

2 shows a flow chart 200 representing a procedure for conditioning control.

In a first phase 201, the digital twin 106 of the real conditioning process, as represented by the recorded data obtained by a monitoring system (e.g. sensors), is trained using a generative AI approach.

In a second phase 202, the conditioning process monitoring system (i.e., the predictor 105) is trained and continuously used by the digital twin using real and artificially generated data.

In a third phase 203, the trained predictor 105 is used to monitor the conditioning process. For this, the first two phases must have been at least partially executed, although they can be continued to increase the robustness of the predictor 105.

The first phase 201 includes data collection 204, signal analysis and feature extraction 205, training 206 of the generative machine learning model, and evaluating (or evaluating) 207 the generative machine learning model. These steps are described in more detail below.

Step 204: Data collection (the training phase of the generative model)

Data acquisition involves the use of a measurement chain of sensors, charge amplifiers, filters and analog-to-digital converters to collect sensor data about the conditioning process. Depending on the object or process phenomenon for monitoring, different sensors (e.g. sensors for detecting acoustic emission, force, vibration, and the electrical power, current or voltage of drive components, as well as other suitable sensors) may be used (but this is not necessary). The corresponding measurement signals from the sensors are amplified and filtered (if necessary) before being digitized (if analogue) with an analogue-to-digital converter in accordance with the WKS sampling theorem. In the case of conditioning process monitoring and conditioning process control, acoustic emission, electrical power, vibration and current signals may preferably, but not exclusively, be used. The sampling rate of the acoustic emission signal may preferably, but not exclusively, be 0.1 to 2 MHz and the sampling rate of the electrical power and current may preferably, but not exclusively, be 50 to 100 kHz. The acoustic emission sensor and the vibration sensor should be mounted as close as possible to the contact zone between the grinding tool and the dressing tool. In addition, the number of coupling points between components (e.g. bearings, screw connections) should be as low as possible. Possible locations are in or on the tool spindle, in or on the work spindle, on the tailstock or at the tool spindle or tailstock center point or other components. The electrical power and current should be measured on the three conductors of the tool 101, the dresser 102 and/or the work spindle.

Step 205: Signal analysis and feature extraction (the training phase of the generative model)

Signal evaluation involves processing the recorded sensor data, i.e. the raw sensor measurement signals. For this purpose, these are adapted as required in such a way that the measurement range is of interest. Different methods can be used to evaluate the raw sensor measurement signals, which can include a transformation of the signal in the time and frequency domain in order to reduce the raw data to significantly changing variables.

3 shows a typical signal from an acoustic emission sensor (AE sensor), corresponding to a sensor data item, which is further processed to extract a feature set.

4 issues a response to the sensor data 3 applied signal analysis procedure, which includes the construction of a low frequency envelope 401 of the signal and the detection of the signal part 402 related to the conditioning process (also referred to as the conditioning zone).

For example, the sensor signal in the conditioning zone 402 may correspond to a conditioning stroke (or other unit of the conditioning process). When actual monitoring is performed, features are derived per conditioning stroke and evaluated per conditioning stroke.

After the conditioning zone 402 has been detected by the algorithm, feature extraction is performed, i.e., values of features (e.g., statistical features) of a feature set are determined. The feature set includes, e.g., statistical characteristics of the conditioning signal (such as rolling mean, standard deviation, minimum, maximum, skewness, kurtosis, absolute value, etc. and their compositions). Labels indicating the optimal stop of the conditioning process are manually prepared by a process expert. That is, the set of feature values (such as mean, etc.) derived from the sensor signal corresponding to a conditioning stroke is provided with a label indicating whether the conditioning should be terminated after this conditioning stroke (or another unit of the conditioning process) or should instead be continued. Each set of feature values and a label forms a training data item.

Step 206: Training the generative AI model 106

The generative AI model 106 is trained on the basis of the training data items to generate artificial training data for different labels, ie sensor data items (such as a 3 shown time series, from which features can then be extracted as training input data for the predictor). The training is carried out using an ML training algorithm in accordance with the generative ML model used, which can be e.g. a Generative Adversarial Network (GAN), a Variational Autoencoder (VAE), Variational Recurrent Autoencoders, recurrent neural networks or their conditional counterparts. This approach includes the application of deep AI architectures to capture an efficient latent variable representation of the process under consideration and nonlinear mapping from this representation to observed data. A recurrent structure of neural network configurations enables the inclusion of temporal dependencies over time steps. This leads to the modeling of a nonlinear behavior of considered time series measurement data of the real observed process. A conditional type of generative Al model may be used if it is of interest to learn a digital twin 106 parameterized according to the desired needs to generate additional training data items and thus an artificial training data set with specified characteristics, e.g. points indicating the dressing process stop signal appearance or required specifics of the observed structure and dynamics of the considered measurement data. Thus, the obtained digital twin 106 can be used to generate data sets in accordance with probabilistic distributions of interest for further successful training of the supervised classifier in order to prepare for dealing with highly variable measurement data of the real process.

Thus, time series data labelled by an expert (i.e. one or more experts) (such as those in 3 shown) to train (using a conditional generative model) a generative AI that creates a digital twin of the production process.

5 shows an example of artificial sensor data generated by the digital twin 106.

A first diagram 501 shows the evolution of the conditioning process (in terms of acoustic emission sensor data in absolute values). The index of the conditioning stroke is given along the y-axis. The time in seconds is given along the x-axis. The line 503 indicates the optimal stop of the conditioning (specified by a label defined by an expert). Thus, each row shown in the first diagram 501 (i.e., the values for each conditioning stroke) is a sensor data item generated by the digital twin 106, where the digital twin 106 provides the sensor data items including a sensor data item for a stroke at which the conditioning should or can be stopped and sensor data items for strokes at which the conditioning should be continued. For example, the generative model can be trained and viewed to generate sensor data conditioned in a specific stroke at which the conditioning should be stopped (i.e., a “hold stroke”), i.e. i.e., it generates the entire section of dressing cycle sensor data (the entire sequence of strokes within the same dressing cycle) for the specific holding stroke. Accordingly, the label of the sensor data item marked by line 503 (stroke 10), which in this example is the holding stroke, is set to indicate that the conditioning process should be stopped, while the label for strokes 0 through 9 is set to indicate that conditioning should continue. For strokes 11 through 20, the label can also be set to indicate that conditioning should be stopped.

A second diagram 502 shows the absolute value (given along the y-axis) over time (given along the x-axis in seconds) for conditioning stroke 1 (dotted line), for conditioning stroke 10 (dashed line) and for conditioning stroke 20 (solid line).

Step 207: Assessment/evaluation of the generative model 207

Based on a test data set, a final selection can be made as to which configuration of the generative AI leads to the best digital twin 106 with sufficient correspondence to the real conditioning process as represented by the measured sensor data. In this way, the relevance (i.e. quality) of the generative model is finally assessed, taking random influences into account, and the best model is saved as digital twin 106. The assessment of which generative model with which hyperparameters provides the best performance can be done, for example, by comparative testing using the lower bound of the evidence, whereby, depending on the required control of the quality of the digital twin, different metrics (e.g. correlation-based metrics including Pearson, Spearman rank correlations, mean square errors, maximum errors, adapted metrics that show whether the correct process development is preserved in the artificially generated data) can also be applied.

The second phase 202 includes generating artificial sensor data using the digital twin 208, combining the artificial sensor data with real (measured) sensor data 209, preparing features 210, training the predictive model 211, and evaluating (i.e., evaluating) the predictive model 212. These steps are described in more detail below.

Step 208: Generating artificial sensor data using the digital twin 106

The digital twin (i.e. the appropriate parameterized generative model generated in the first phase 201 in steps 206 and 207) for the conditioning process of a grinding tool 102 (with respect to the generated sensor data) is used to generate artificial sensor data of specified sizes and properties, which is possible due to the conditional type of the underlying generative AI model. Depending on the generative model used, this can be done by sampling from a latent space and causing the generative model to generate a sensor data item (such as the one in 3 shown). In the case of a conditional generative model (such as a conditional GAN), the generative model may also be input with a specification of a desired characteristic of the generated sensor data item. The desired characteristic may also be the label for which a sensor data item should be generated, ie, which sensor data item in a sequence of sensor data items (corresponding to a conditioning) is the one at which the conditioning should stop.

Step 209: Combining artificial sensor data with real sensor data

To prepare a training dataset, real (i.e. measured) sensor data is combined with the artificial data in accordance with a specified proportion (or multiple proportions, if different ones should be tried). Thus, real sensor data with limited size and variability is enriched by the sensor data (represented by the digital twin 106) of the digital conditioning process. This leads to fulfilling the needs of ML training methods to train using datasets with sufficiently representative properties. For example, the test dataset for the predictor 105 may contain only real sensor data.

Step 210: Preparing features

For each real and artificial sensor data element (such as the 3 shown), a feature set is extracted (as training input of each training data item). This feature set includes statistical characteristics of the conditioning sensor signal of the sensor data item (such as rolling mean, standard deviation, minimum, maximum, skewness, kurtosis, absolute value, etc. and their compositions). (This step may also be performed before forming training data sets in accordance with different proportions of artificial and real sensor data in step 209.)

Step 211: Training the prediction model 105

The prediction model 105 is preferably, but not exclusively, trained using an ML algorithm such as representation learning techniques, feature importance estimation and feature selection algorithms, supervised approaches including, but not limited to, deep learning architectures and ensemble methods. Since the digital twin 106 enables the flexible generation of conditioning process sensor data (for desired labels, i.e., in other words, for desired classes), in this step one or more ML models can be trained using supervised training per each proportion of real and artificial data in the training data sets prepared in steps 209 and 210. Changing this proportion influences the performance of the resulting models. Applying data generated by the digital twin improves the performance of training routines for various families of supervised ML algorithms including, but not limited to, decision trees, fully connected neural networks and convolutional neural networks, ensemble methods.

6 shows an example of the influence of the ratio of real to artificial data in the training dataset (indicated along the x-axis) on the performance of the trained predictive model in terms of the Spearman correlation (indicated along the y-axis; if each of the variables is a perfectly monotonic function of the other, a perfect Spearman correlation of +1 or -1 occurs) between the output of the ML-based detector (i.e., using the trained predictor) to halt the conditioning process and labels by a process expert.

The dashed line 601 indicates the performance for using only real sensor data. The upper graph 602 is for a 10-fold benchmark training average and the lower graph 603 is for a 10-fold benchmark test average.

As can be seen, adding a moderate amount of artificial data increased the quality of the predictor 105, while adding too much artificial data resulted in a deterioration in quality because artificial data are imperfect and their inherent noise exceeded the benefits from variability.

Step 212: Assessment and evaluation of the prediction model 105

If several predictors have been trained (including using different proportions of real and artificial data), a selection can be made on the basis of a test data set as to which predictive model is best for the given process boundary conditions leads to the best performance. In this way, the relevance (and thus quality) of the prediction model is finally assessed taking random influences into account and the best model is stored and entered into the control unit 104 as a predictor 105. The evaluation of which prediction model with which hyperparameters provides the best performance can be done by comparing with various metrics (e.g. Spearman rank correlation, Kendall rank correlation, accuracy, balanced accuracy, F-score, ROC-AUC, R ² value, explained variance, maximum error). The application of steps 211 and 212 leads to the determination of the optimal proportion of real to artificial data in the training data set.

The third phase 202 includes data collection 213, signal analysis and feature extraction 214, and application of the trained prediction model 215. These steps are described in more detail below.

Step 213: Data acquisition for monitoring a conditioning process This corresponds to step 204. The data acquisition carried out, for example, by the monitoring system 103 comprises the measurement chain consisting of sensors, charge amplifiers, filters and analog/digital conversion. Depending on the object or objects being monitored, different sensors (e.g. sensors for detecting structure-borne sound, force, vibration as well as the electrical power, electrical current and electrical voltage of drive components as well as other suitable sensors) can be used. The corresponding measurement signals of the sensors are amplified and filtered (as required) before they are digitized (if they are analog) with an analog/digital converter in accordance with the WKS sampling theorem. In the case of conditioning process monitoring and conditioning process control, structure-borne sound signals, electrical power, vibration and current can preferably be used, but not exclusively. The sampling rate of the structure-borne sound signal should preferably, but not exclusively, be 1 to 2 MHz and the sampling rate of the electrical power and current should preferably, but not exclusively, be 50 to 100 kHz. The structure-borne sound sensor and the vibration sound sensor should be mounted as close as possible to the contact zone between the tool and the workpiece. In addition, the number of coupling points between components (e.g. bearings, screw connections) should be as low as possible. Possible positions are in or on the tool spindle, in or on the workpiece spindle, on the tailstock or at the tool spindle or tailstock center point or other components. The electrical power and current should be measured on the three conductors of the tool and/or workpiece spindle.

Step 214: Signal analysis and feature extraction

This corresponds to step 205. The signal analysis (or signal evaluation) carried out e.g. by the control unit 104 comprises the processing of the raw sensor measurement signals. For this purpose, these are adapted as required in such a way that the measurement range is of interest. For the evaluation of the raw sensor measurement signals, different methods can be used, which can include a transformation of the signal in the time and frequency domain in order to reduce the raw data to significantly changing variables. According to various embodiments, a low frequency envelope of the signal is constructed and a signal part relating to the conditioning process is identified. Statistical characteristics of the conditioning signal are then calculated as features (e.g. rolling mean, standard deviation, minimum, maximum, skewness, kurtosis, absolute value, etc. and their compositions).

Step 215: Applying the prediction model 105

The control unit 104 uses the predictive model 105 trained in phase 2 (steps 211 and 212) to signal whether the conditioning process can be completed based on the data generated in step 214. The control unit 104 controls the dressing tool 102 accordingly (by sending an appropriate signal to the dressing tool 102).

• • Messung von Sensorsignalen
• • Detektieren des Hubs (d. h. Detektieren des Teils der Sensordaten, der einem Konditionierungshub entspricht)
• • Vorbereitung von Merkmalen für den Hub
• • Klassifizieren der Merkmale (in die Klassen „Konditionierung anhalten“ und „Konditionierung fortsetzen“)
• • Anhalten oder Fortsetzen der Konditionierung in Übereinstimmung mit der Klassifizierung der Merkmale

Thus, Phase 3 implements condition process monitoring, which includes:

• • Measurement of sensor signals
• • Detecting the stroke (ie detecting the part of the sensor data that corresponds to a conditioning stroke)
• • Preparing features for the hub
• • Classifying the features (into the classes “pause conditioning” and “continue conditioning”)
• • Pause or continue conditioning in accordance with the classification of characteristics

7 shows results of a simplified application example for monitoring and controlling a conditioning process. In the starting position, a convex grinding wheel profile was conditioned. The conditioning process was originally carried out conventionally by a pro process experts so that a total of 20 fixed conditioning strokes have to be carried out so that the grinding tool again corresponds to the required condition. The application of the methodology described above shows that after six conditioning strokes an optimum number of conditioning strokes has already been achieved again.

The diagram from 7 shows the development of the conditioning process (in terms of acoustic emission sensor data generated by the digital twin in absolute values). The absolute value is given (along the y-axis) over time (along the x-axis in seconds) for conditioning stroke 1 (dotted line), for conditioning stroke 6 (dashed line) and for conditioning stroke 20 (solid line).

Learning the digital twin using generative Al methods provides an efficient way to generate observations of digital conditioning processes of grinding tools, which could be represented in the form of extensive data sets with desired properties. Since these data sets are required by the training algorithms and lead to the increased performance of ML models, a more optimal control of the real production process becomes possible.

• • Reinigung/Vorbereitung von Filtermaterialien und Situationsvervollständigung des Prozesses.
• • Reinigung/Vorbereitung von Katalysatoren und der Situationsabschluss des Prozesses.

In addition, the application of the algorithm is not limited to conditioning processes for grinding tools, but can also be used to monitor the tool condition in other machining processes (especially conditioning processes) (e.g. from the main groups of DIN 8580 such as primary shaping, forming, separating, joining, coating and changing material properties). If required, the measurement technology may need to be adapted to the conditioning of grinding tools. This means that the use of the procedure described above in other technical systems in which a conditioning process can be controlled/regulated depending on the situation is also conceivable. Possible areas of application are:

• • Cleaning/preparation of filter materials and situational completion of the process.
• • Cleaning/preparation of catalysts and the situational completion of the process.

In summary, according to various embodiments, a 8th The method shown is provided.

8th shows a flowchart 800 illustrating a method for training a predictive model for monitoring a conditioning process (e.g., a grinding tool).

In 801, time series sensor data of a conditioning process is obtained.

In 802, the time series sensor data is labeled by labels (e.g., by a predetermined set of labels corresponding to, e.g., a predetermined set of classes), wherein the labels indicate sensor data elements of the time series sensor data that reflect that a quality criterion has been met in the conditioning process.

In 803, using the labeled time series sensor data, a (e.g., conditional) generative model is trained to generate additional time series sensor data of the conditioning process for labels of the additional time series sensor data (e.g., the predetermined set of labels), wherein the labels indicate sensor data elements of the additional time series sensor data that reflect that a quality criterion has been met in the conditioning process.

In 804, a predictive model is trained for supervising the conditioning process using the labeled time series sensor data, the additional time series sensor data, and the labels of the additional time series sensor data through supervised learning (wherein the labels of the time series sensor data and the additional time series sensor data are used as ground truth, and information about the sensor data items of the time series sensor data and the additional time series sensor data (e.g., features) is used as training input of the predictive model).

For example, the default set of labels used as ground truth for the prediction model can simply be “pause conditioning” and “continue conditioning”.

The prediction model is trained to correctly predict the labels for the received sensor data and for the additional sensor data, i.e., the prediction model is trained by supervised learning to correctly classify the sensor data items, i.e., into sensor data items that reflect that the quality criterion has been met and into sensor data items that the quality criterion has not been met (if this can be achieved in training). A sensor data item (e.g., time series data for a part of the conditioning process such as for a hub as in the example from 3 ), which reflects that the quality criterion has been met, means that the sensor data element contains sensor data that were obtained (ie recorded or measured) at a time in the conditioning process at which the quality criterion is or has been met.

Meeting the quality criterion may mean that the conditioning process can be stopped. It may mean, for example, that a dressing tool being conditioned is sufficiently smooth and/or does not have a hole, etc.

The label for a sensor data item of the time series sensor data may be determined using one or more features extracted from the sensor data item. The label determined using one or more features (i.e., determined from the one or more features) is then associated with the one or more features, i.e., it may be considered a label for the one or more features. Since the one or more features are features of a sensor data item, the label may equivalently also be considered a label of the sensor data item itself (i.e., the label labels the sensor data item).

According to one embodiment, the method is computer-implemented.

Although specific embodiments have been shown and described herein, it will be apparent to those of ordinary skill in the art that a variety of alternative and/or equivalent implementations may be substituted for the specific embodiments shown and described without departing from the scope of the present invention. This application is intended to cover any adaptations or variations of the specific embodiments discussed herein. Thus, this invention is intended to be limited only by the claims and their equivalents.

Claims (8)

A method for training a predictive model to monitor a conditioning process, the method comprising: obtaining time series sensor data of a conditioning process; labeling the time series sensor data by labels, wherein the labels indicate sensor data elements of the time series sensor data that reflect that a quality criterion has been met in the conditioning process; training a generative model using the labeled time series sensor data to generate additional time series sensor data of the conditioning process for labels of the additional time series sensor data, wherein the labels indicate sensor data elements of the additional time series sensor data that reflect that a quality criterion has been met in the conditioning process; training a predictive model to monitor the conditioning process using the labeled time series sensor data, the additional time series sensor data, and the labels of the additional time series sensor data by supervised learning. Procedure according to Claim 1 wherein the prediction model is trained to indicate for an input sensor data element whether the input sensor data element reflects that the quality criterion has been met in the conditioning process from one or more features extracted from the input sensor data element and input to the prediction model. Procedure according to Claim 1 or 2 wherein training the generative model using the labeled time series sensor data comprises extracting one or more features of the sensor data item and training the generative model using the one or more features and the label of the sensor data item for each sensor data item of the time series sensor data. Procedure according to one of the Claims 1 until 3 which comprises training a plurality of prediction models to monitor the conditioning process using the labeled time series sensor data, the additional time series sensor data, and the labels of the additional time series sensor data through supervised learning using different proportions of labeled time series sensor data and additional time series sensor data, evaluating the trained plurality of prediction models, and selecting one of the plurality of prediction models as a prediction model depending on the evaluation. A method for controlling a conditioning process, the method comprising: training a prediction model according to one of the Claims 1 until 4 ; obtaining time series sensor data of a conditioning process to be observed comprising a series of sensor data items; inputting information about the sensor data item into the prediction model for each of the sensor data items; and stopping the conditioning process in response to the prediction model indicating for a sensor data item of the sensor data items, that the quality criterion has been met in the conditioning process. Conditioning process control system configured to perform a method according to any of the Claims 1 until 5 to execute. A computer program comprising instructions which, when executed by a computer, cause the computer to perform a method according to any of the Claims 1 until 5 executes. A computer-readable medium comprising instructions that, when executed by a computer, cause the computer to perform a method according to any of the Claims 1 until 5 executes.

Images