Computer-implemented method for controlling winding machine, computer program and winding machine

Abstract

The invention relates to a computer-implemented method for controlling a winding machine, a computer program and a winding machine, in particular to a computer-implemented method for controlling a winding machine comprising at least a winder and a rewinder, the method comprising the following steps: determining an actual speed of the winder during operation of the winding machine; performing signal processing on the determined actual speed in order to extract winder-related features, including subtracting the commanded speed from the determined actual speed; determining an envelope signal of the subtracted signal; and filtering the envelope signal, where the filtering retains amplitude-related information, using the filtered envelope signal as a winder-related feature and as an input to a trained machine learning algorithm, and operating the machine learning algorithm based on the winder-related feature and giving an anomaly flag as an output.

Description

Technical Field

[0001] The invention relates to a computer-implemented method for controlling a winding machine, a computer-implemented method for training a machine learning algorithm, a corresponding computer program and a winding machine. Background Art

[0002] Winders are installed in various applications related to the paper, textile or battery industries. Winders enable coils of different types of material to be wound around a target object. It is of utmost importance to ensure that the coil is wound tightly to guarantee a high quality of the wound component. For this purpose, complex control schemes have been developed for the winders.

[0003] Various parameters, such as friction or temperature, influence the web speed, which is compensated by the winder control. However, the web tension may show low-amplitude anomalies, and depending on the type of web material and the magnitude of the web tension variation, these anomalies are usually either detected during the product quality check or not detected. In both cases, product quality is affected, either the defect is detected late in the process, making the product have to be sorted out, or the defect is not detected at all, which can even lead to safety issues.

[0004] It is known that highly automated winders as well as control schemes ensure that the web is tightly wound around the object. To control the winding quality, different manual inspection methods are used. Visual inspection is known and can be checked by an experienced operator in cases where the magnitude of the anomaly in the applied web tension is quite high. In addition to the magnitude of the anomaly, the success of this method also depends on the material being wound. For example, it is quite simple for an experienced operator to determine if paper is tightly wound, while it requires additional effort when the web material is a strip roll or wire. The same is true when the anomaly occurs in the lower layers of the wound web and additional layers are wound before the visual inspection is performed.

[0005] Furthermore, manual spectrogram analysis is known for manual inspection: an operator monitors in online mode a spectrogram generated for analyzing the signal from the rewinder. In the manual case, the interpretation of the calculated spectrogram also requires a great deal of experience. The spectrogram is generated by a short-time Fourier transform of the signal acquired from the dancer. For example, a spectrogram generated by a short-time Fourier transform of the signal acquired from the dancer (e.g. the deviation of the dancer from a set point) is used and must be interpreted. The advantage of this method is that the quality control is performed online and the operator can stop the winding process if additional frequencies appear on the spectrogram and therefore an anomaly is assumed. Summary of the invention

[0006] In view of the shortcomings described in the prior art, the object of the present disclosure is to provide a computer-implemented method for controlling a winding machine and for training a machine learning algorithm with automatic anomaly detection, a corresponding computer program and a corresponding winding machine.

[0007] These objects are solved by the subject-matter of the independent claims. Advantageous embodiments are set forth in the dependent claims.

[0008] The present invention relates to a computer-implemented method for controlling a winding machine, the winding machine comprising at least a winder and a rewinder, the method comprising the following steps: determining an actual speed of the winder during operation of the winder, performing signal processing on the determined actual speed in order to extract winder-related features, including subtracting a command speed from the determined actual speed, determining an envelope signal of the subtracted signal, and filtering the envelope signal, wherein the filtering retains amplitude-related information, using the filtered envelope signal as a winder-related feature and as an input to a trained machine learning algorithm, and running the machine learning algorithm based on the winder-related feature and giving an abnormal flag as an output.

[0009] Winders can be used for any kind of material to be wound or any coil, such as paper, foil, fabric coil or the like. Furthermore, the target object on which the coil is wound can be round, prismatic, triangular or rectangular in shape. For all these special winders, a constant coil tension must be ensured as far as possible. As soon as the coil tension is not within a certain given range (depending on the specific equipment and construction), the coil is wound irregularly and the quality of the wound product is poor.

[0010] Therefore, different kinds of compensation mechanisms are known for winders to allow control of the web tension.In addition, different kinds of components are often used to introduce changes into the system or to modify parameters accordingly, such as in a specific dancer roll or web storage or load cell.

[0011] It is proposed to analyze the signal of the actual speed of the winder, which is the component that provides the coil for the rewinder strip coil, by means of a machine learning algorithm. The signal of the actual speed of the winder, processed in a correct manner, contains rich information for distinguishing between healthy and unhealthy winding states. The machine learning method for analyzing the winder signal works in a supervised or unsupervised setting and in real-time (i.e., during the operation of the winder and during the winding process) or in an ex post facto manner.

[0012] The design of the anomaly detection method is based on selecting signals with rich information content that can indicate whether an anomaly has occurred within a given period of time under investigation. This is achieved by utilizing the following signals: the signal of the actual speed of the winder and the signal of the commanded speed.

[0013] A process of feature extraction is performed, which includes the step of signal processing. In this way, the signal is converted and features are extracted, which include information that can be associated with the occurrence of anomalies. In particular, for feature extraction, the actual speed and command speed of the winder are processed together, as described below: if applicable, the positive offset and / or negative offset of the signal is removed. The signal of the winder command speed is subtracted from the signal of the actual speed. The resulting signal is used to determine the corresponding envelope signal. The envelope signal is filtered in a manner that retains information about the amplitude of the envelope signal. For example, depending on the noise of the signal, a low-pass filter with a short transition phase is used, such as a high-order Butterworth, Chebyshev, Bessel filter, preferably 3rd order and higher. By using a suitable filtering method, any interference or spikes or noise are filtered out while retaining any amplitude-related information of the signal.

[0014] The processed signals are used as winder-related features and input into unsupervised or supervised machine learning algorithms. Unsupervised machine learning algorithms indicate that an anomaly has occurred. Supervised machine learning algorithms indicate whether the process under investigation is of low or high quality at a given time window.

[0015] In an advantageous manner, an ML-based solution is provided that is able to detect anomalies in web tension by measuring and analyzing winder process parameters without the need for additional sensors.

[0016] Advantageously, there is no need for a person skilled in the art to manually perform a post hoc inspection of the wound material. This also reduces unnecessary material waste or rework.

[0017] Furthermore, in contrast to the well-known semi-automatic analysis based on the calculation of spectrograms, the described approach does not require an experience-based interpretation process.

[0018] The proposed fully automated solution is suitable for all materials and operating conditions and can be used for in-line quality control of the winding process. This also enables the method to be used in large-scale production with ex post quality assurance for all production units within a certain area.

[0019] According to one embodiment, the method further comprises the step of initiating a modification of one or more control parameters of the winder in case of an indicated anomaly. Detecting anomalies through a machine learning algorithm, the parameters of the winder, such as the dancer roller, can be adjusted to keep the quality of the product as high as possible.

[0020] With appropriate machine learning algorithms, an anomaly detector is implemented that outputs certain anomaly scores that are converted into parameters of the control system in order to reset the reference values ​​of certain components of the winder (e.g. winder or dancer roller) in order to keep the product of acceptable quality.

[0021] By feeding the output of the machine learning algorithm back into the control system, the winding process is improved and the web tension remains more stable.

[0022] According to one embodiment, the winder further comprises a coil storage, and coil storage related features are extracted from the coil storage actual position, and the coil storage related features are used as additional inputs to the trained machine learning algorithm, and the machine learning algorithm operates based on the winder related features and the coil storage related features.

[0023] In an advantageous manner, in a winding machine using a coil storage, signals and features related to the coil storage are also used to extract more or better information or information about anomalies in the winding process at an earlier point in time. In addition to the processed signal of the difference between the actual speed and the commanded speed, the processed signal of the actual position of the coil storage is also used, making the machine learning algorithm more robust.

[0024] The coil stocker is a linear motor based system whose operation aims to calibrate the disturbing behavior on the coil tension in a controlled way. However, the operation of the coil stocker is not efficient in cases where the anomalies are very unexpected, difficult to be captured by the reference signal or of short duration. In this case, the anomaly manifests itself in the coil stocker position which deviates from the equilibrium point due to the high speed or torque of the coil with the abnormal behavior. Therefore, the information content inherited from such a position signal is quite rich.

[0025] According to one embodiment, the coil storage related features are constructed based on the peak-to-peak values ​​of the actual position of the coil storage within a specified time window.

[0026] The peak-to-peak features of the actual position of the coil storage in a specified (e.g. user-defined) time window are used and further processed by means of statistical filtering. For example, a statistically based filtering method such as the median is used, for example as a median filter with a window size preferably of 5. The filtered peak-to-peak features not only show deviations in amplitude, i.e. deviation behavior, but also fluctuations associated with short-term changes. This allows capturing short amplitude fluctuations and allows detecting amplitude anomalies that evolve over time.

[0027] According to one embodiment, the time window is specified by the period of the signal of the actual position of the coil storage. This period and therefore the time window is known for example for a specific winding machine setting and is defined by the operator after an initial setting phase. Alternatively, the time window is automatically determined by the system based on the identified period.

[0028] According to one embodiment, the winder further comprises a floating roller, and floating roller related features are extracted from the actual position of the floating roller, and the floating roller related features are used as additional input to a trained machine learning algorithm, and the machine learning algorithm operates based on the winder related features and the floating roller related features or based on the winder related features, the web storage related features and the floating roller related features.

[0029] In a winder setup using a dancer, feature extraction from dancer-related signals is performed and used in addition to winder-related feature extraction to make the machine learning algorithm more robust. If both the coil storage and dancer are in place, feature extraction from all three signals of the winder, coil storage and dancer can even be used to feed the machine learning algorithm.

[0030] According to one embodiment, the dancer-related feature is constructed based on the waveform shape of the actual position of the dancer within a specified time window, in particular the crest factor of the signal of the actual position of the dancer. The structural change of the waveform shape can be used to obtain information about the health status of the dancer, and thus draw conclusions about the health status of the entire system.

[0031] According to one embodiment, the machine learning algorithm has been trained based on a supervised training method. Suitable supervised learning techniques for anomaly detection are the k-nearest neighbor algorithm (KNN), for example, k=5 or k=7, or methods based on logistic regression and decision trees, in particular methods with non-random partitioning.

[0032] According to one embodiment, the machine learning algorithm has been trained based on an unsupervised training method. For the unsupervised training method, the extracted features are introduced without additional information, and clusters are created by an algorithm, such as using a k-means method. For example, a k-means algorithm or a fuzzy c-means algorithm using k=2 or k=3 is used.

[0033] The present invention also relates to a computer-implemented method for training a machine learning algorithm, wherein the machine learning algorithm gives an abnormality flag as an output for controlling a winding machine and initiates a modification of one or more control parameters of the winding machine in the event of an indicated abnormality, wherein the winding machine includes at least a winder and a rewinder, the method comprising the following steps: determining the actual speed of the winder during operation of the winder, performing signal processing on the determined actual speed to extract winder-related features, including subtracting a command speed from the determined actual speed, determining an envelope signal of the subtracted signal, and filtering the envelope signal, wherein the filtering retains amplitude-related information, using the filtered envelope signal as a winder-related feature, and training the machine learning algorithm with the winder-related features.

[0034] According to one embodiment, an unsupervised training method is used to identify clusters and determine the degree of abnormality of the input data based on the corresponding clusters.

[0035] According to one embodiment, a supervised training method is used to identify categories based on a labeled training data set, and the degree of abnormality of the input data is determined based on the corresponding category.

[0036] In the case of supervised learning methods, it is necessary to estimate whether the web tension during operation is healthy or exceeds a threshold for normal behavior. This can be achieved by measuring the web tension and it can be classified into healthy and abnormal operating states. Those states associated with the training dataset are used as labels in the training phase of the machine learning algorithm.

[0037] According to one embodiment, the winder in the training phase comprises a web tension sensor, wherein the marking data are generated depending on the value of the web tension sensor.

[0038] In the given scope of investigation, sensors have been introduced into the winder to measure the web tension in real time. After signal processing, the amplitude of the web tension constitutes information that can be effectively used as a label for the corresponding training dataset.

[0039] The invention also relates to a computer program having instructions which, when executed by a computing device or system, cause the computing device or system to perform a method according to any one of the above-described method embodiments.

[0040] The invention also relates to a winding machine with a data processing system, the data processing system comprising means for carrying out a method according to any one of the embodiments of the method described above. The data processing system is for example integrated into a controller for controlling the winding machine. The controller controls the axes of the different rollers or the corresponding drives, in particular depending on the implementation of the winding machine. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In the following, different aspects of the invention are described in more detail with reference to the accompanying drawings.

[0042] Figure 1 Schematic diagram showing winder dependent curves of actual speed and command speed according to a first embodiment of the invention.

[0043] Figure 2 A schematic diagram showing a processed winder correlation curve for offset correction according to a first embodiment of the present invention.

[0044] Figure 3 Schematic diagram showing a winder related curve of winder command speed minus winder actual speed further processed according to the first embodiment of the present invention.

[0045] Figure 4 A schematic diagram of a filtered envelope signal extracted as a warp-related feature according to a first embodiment of the present invention is shown.

[0046] Figure 5 A schematic diagram showing the amplitude distribution of the actual speed of a winder in healthy operation and different unhealthy operations according to a first embodiment of the present invention is shown.

[0047] Figure 6 A schematic diagram of a winding machine according to a second embodiment of the present invention is shown.

[0048] Figure 7 Schematic diagram showing coil store correlation curves for different signal processing according to a second embodiment of the invention.

[0049] Figure 8 A schematic box diagram showing relevant features of a coil store according to a second embodiment of the invention.

[0050] Fig. 9 A schematic diagram showing a control mechanism for a winding machine according to a third embodiment of the present invention is shown.

[0051] Fig.10 A schematic diagram showing a dancer roller correlation curve according to a fourth embodiment of the present invention is shown.

[0052] Fig.11 shows the first time window during healthy operation Fig.10 An enlargement of the curve shown in .

[0053] Fig.12 The second time window during the unhealthy operation period is shown. Fig.10 An enlargement of the curve shown in .

[0054] Fig.13 shows the third time window during further unhealthy operation. Fig.10 An enlargement of the curve shown in . DETAILED DESCRIPTION

[0055] For a first embodiment of the invention, a winder is equipped with a machine learning algorithm that evaluates whether the winding process of winding a coil (e.g., a coil for a battery cell) from a winder (e.g., a round object) to a rewinder (e.g., a rectangular target object) is performed without disturbances that would result in poor battery cell quality. It is particularly important to ensure that a regular coil tension is applied to the coil during the winding process to ensure that there are no irregularities in the product that could result in battery damage. The machine learning algorithm analyzes the winding process based on winder-related features, which are extracted from the winding process parameters without the need for additional sensors. As output, the machine learning algorithm gives an indication whether the winding process is in healthy conditions or whether unhealthy conditions exist.

[0056] Figure 1 Two speed signals v are shown, the signal of the actual speed va of the winder in mm / ms being plotted against time in ms. Only the part of the repetitive forward movement of the winder is shown, the movement speed in the intervening time spans having been blanked out for clarity.

[0057] In addition, the command speed vc is shown in the same unit.

[0058] exist Figure 2 In the diagram shown, some signal processing has been performed, and Figure 1 The command velocity vc has now been offset corrected to become the offset corrected command velocity vc' plotted over time. Figure 1 The actual speed va of the same winder is also signaled in Figure 2 Plotted in for comparison.

[0059] In the next step, the actual speed va is subtracted from the offset-corrected command speed vc', and the subtracted signal dv is Figure 3 , and again plotted over time in the same units as in the previous figures.

[0060] As a next step, an envelope signal is determined which envelops the subtracted signal dv. The envelope signal is also filtered using a filtering algorithm which preserves amplitude information and removes noise. In the first embodiment, a third-order Butterworth algorithm is applied. The filtered envelope signal is used as a winder-related feature fw extracted from the process parameters of the winding process and is used in Figure 4 Shown in.

[0061] The winder related features fw are fed into the machine learning algorithm. As output, the machine learning algorithm according to the first embodiment gives a flag, i.e., whether the winding process in a given time window has been performed under normal conditions, so that no unhealthy state of the winder is expected, or whether there is a hint of the type of disturbance that occurred. In supervised machine learning algorithms, different degrees or types of disturbances, such as different sources of irregularities belonging to the web tension, can be trained with separately labeled training data sets.

[0062] Figure 4 The filtered envelope signal fw is shown in three different operating conditions, from 0-0.8 x10 6 The health status in the time period of ms, from 0.8-1.45 x10 6 ms has the first state of introducing interference, and from 1.45-2.25 x10 6 ms has a state of introducing a second interference. In this embodiment, the second interference is an additional interference that occurs after the first interference has occurred or an interference caused to the existing interference.

[0063] Figure 5 The distribution of the amplitude of the actual velocity va of the winder is shown, which is relatively sharply concentrated around a certain amplitude for the healthy state 10 and is more blurred for the unhealthy states 20 , 21 .

[0064] In a second embodiment, the winding process is assisted by a machine learning algorithm that takes into account not only winder-related features fw but also features extracted from the coil storage. Figure 6 A typical structure of a winder having a winder w and a rewinder r is shown.

[0065] The winder is a fully automatic system capable of winding a material web 1 (eg a textile web according to the second embodiment of the invention) around a target object. Figure 6 As shown, this is achieved by a system with a free or controlled rotor. In particular, a winder w is provided as a starting point, where the coil to be wound is stored. In an opposite position, a rewinder r is provided as a target object for receiving the coil. The target object in this embodiment has a prismatic shape.

[0066] Both the winder w and the rewinder r move in a controlled manner in order to wind the web tightly. For this purpose, two linear motors are deployed, which drive the floating roller d and the web storage wa, so that the web speed, torque and position can be calibrated according to the control scheme provided. In this control system, the mechanical system of the main shaft m is used to provide reference information about the web position and speed.

[0067] In order to utilize the coil storage related features as input to the machine learning algorithm, the signal of the actual position of the coil storage is analyzed. Figure 7 The signal swa of the actual coil storage position s plotted against the time t is shown. The actual coil storage position signal swa is filtered to remove noise etc. The resulting filtered signal swa' is processed to extract the peak-to-peak value PtP of the signal in a specified time window.

[0068] exist Figure 8 , box plots of different peak-to-peak values ​​PtP of the filtered coil storage actual position signal are shown. A first peak-to-peak value 10 is derived for the coil storage actual position signal swa during a time span during which the winding process is performed under normal or healthy conditions. The peak-to-peak value PtP is derived from the signal within a specified time window within this time span under normal conditions. The appropriate time window to be selected depends on the period of the actual coil storage position signal swa, wherein the period is formed by the compensating movement of the coil storage for compensating for irregularities in the coil tension. A second and a third peak-to-peak value 20, 21 are derived for the coil storage actual position signal swa during different time spans during which the winding process is performed under abnormal or unhealthy conditions.

[0069] Fig. 9 A scheme for a control mechanism located below a winding machine or used during the winding process is given. An anomaly detector 102 is provided, which works based on the extracted features fw, fwa, fd, which are described above or below and are extracted by a feature analyzer 101 based on various process parameters that can be derived from the winding machine process control system. As an output, the anomaly detector gives an anomaly flag j / n, which indicates whether the winding process is performed under normal conditions or under abnormal conditions. Preferably, additional information is provided to the control system C of the winding machine together with the anomaly flag j / n, which additional information has been derived in the training phase of the machine learning algorithm and allows conversion into parameters of the control system. For example, anomaly scores or anomaly degrees 20, 21 can be derived, for which appropriate reference values ​​of control parameters P1, P2, P3, ... of the winding control system have been derived in the training phase.

[0070] In an advantageous manner, the control mechanism is enriched with information about how to best adapt the control parameters in case of a specific unhealthy state of the winding machine. For example, the drive system 200 of the rewinder comprising the rewinder motor 201 and the load 202 provides the value of the actual speed 2v as well as the additional torque value 2t and the speed setpoint value 2vc to the current controller Ci and the speed controller Cv, both of which are influenced by the reference values ​​of the control parameters P1, P2, P3, ..., which are able to provide the current I and the speed V as correction variables to the drive system 200.

[0071] For example, an anomaly detector is implemented on an industrial edge system or industrial PC connected to a web application for training purposes.

[0072] Fig.10 FIG. 1 shows feature extraction in a winding process according to a fourth embodiment of the invention. Here, not only the input parameters of the winder of the winder are used for the purpose of anomaly detection, but also the dancer and its behavior during winding are checked. Thus, a signal of the actual position sd of the dancer is determined and recorded at time t.

[0073] from Fig.10 It can be seen from the figure that the waveform shape fd of the signal of the actual position sd of the floating roller has three different characteristic forms. The details of the first waveform shape fd1 are shown in Fig.11 is shown in more detail and zoomed in to Fig.10 Each rising edge belongs to a winding process in a winding program with repeated winding and unwinding (only the winding or forward movement is shown).

[0074] A signal of the actual position of the dancer roller having a waveform shape fd1 indicates normal behavior. As long as the machine learning algorithm gets such a waveform shape fd1 as input, no disturbance or anomaly is indicated.

[0075] Fig.12 The waveform shape fd2 is shown in more detail. The waveform shape fd2 is very different from the waveform shape fd1, which makes the signal particularly valuable for the machine learning algorithm. The machine learning algorithm is particularly suitable for having several extracted features as input, such as winder related features, dancer related features and optionally additional web storage related features. Taking into account these different extracted features makes the machine learning algorithm a more robust system and allows reliable anomaly detection also in cases where one of the components delivers wrong or misleading values, for example, these values ​​differ too much from the values ​​generated and collected as training data during the training phase.

[0076] To further illustrate, Fig.13The waveform shape fd3 is shown in more detail. The waveform shape fd3 also has a form that is significantly different from the waveform shapes fd1 and fd2. It is caused, for example, by a different kind of disturbance and indicates a specific kind of abnormality, for example, represented by an abnormality score or abnormality degree 21 that is different from the abnormality degree 20 in the case of the waveform shape fd2. For example, the abnormality score is represented by percentage information.

[0077] In an advantageous embodiment, each anomaly score is associated with a set of parameters, and the anomaly score is provided as feedback to a control mechanism for adjusting the controller, for example, adjusting the winder speed or the dancer force or counterforce according to reference values ​​from a design phase or a training phase.

[0078] Other possible implementations or alternative solutions of the present invention also cover combinations of features described above or below with respect to the embodiments that are not explicitly mentioned in this document. Those skilled in the art may also add separate or independent aspects and features to the most basic form of the present invention.

Claims

1. A computer-implemented method for controlling a winding machine (100), the winding machine comprising at least a winder (w) and a rewinder (r), the method comprising the following steps: - determining the actual speed (va) of the winder (w) during operation of the winder (100), - performing signal processing on said determined actual speed (va) in order to extract winder-related features (fw), comprising subtracting a commanded speed (vc, vc') of said winder from said determined actual speed (va), determining an envelope signal of the subtracted signal (dv), and filtering said envelope signal, wherein said filtering preserves amplitude-related information, - using the filtered envelope signal as said winder related feature (fw) and as input to a trained machine learning algorithm, and - Running the machine learning algorithm on the winder related features (fw) and giving anomaly flag (j / n) as output.

2. The method according to claim 1, further comprising the steps of: In case of an indicated anomaly, a modification of one or more control parameters (P1, P2, P3, ...) of the winding machine (100) is initiated.

3. The method according to claim 1 or 2, wherein: The winder (100) further comprises a coil storage (wa), and coil storage related features (fwa) are extracted from the coil storage actual position (swa), and the coil storage related features are used as additional inputs to the trained machine learning algorithm, and the machine learning algorithm operates based on the winder related features (fw) and the coil storage related features (fwa).

4. The method according to claim 3, wherein: The coil storage related features (fwa) are constructed based on peak-to-peak (PtP) values ​​of the coil storage actual position (swa) within a specified time window.

5. The method according to claim 4, wherein: The time window is specified by the period of the signal of the web storage actual position (swa).

6. The method according to claim 5, wherein: The winder (100) further comprises a floating roller (d), and a floating roller related feature (fd) is extracted from the floating roller actual position (sd), and the floating roller related feature is used as an additional input to the trained machine learning algorithm, and the machine learning algorithm operates based on the winder related feature (fw) and the floating roller related feature (fd) or based on the winder related feature (fw), the web storage related feature (fwa) and the floating roller related feature (fd).

7. The method according to claim 6, wherein: The dancer-related feature (fd) is constructed based on the waveform shape (fd1, fd2, fd3) of the dancer actual position (sd) within a specified time window, in particular the crest factor of the dancer-related actual position (sd) signal.

8. A method according to any one of the preceding claims, wherein: The machine learning algorithm has been trained based on a supervised training method.

9. A method according to any one of the preceding claims, wherein: The machine learning algorithm has been trained based on an unsupervised training method.

10. A computer-implemented method for training a machine learning algorithm, wherein: The machine learning algorithm gives an abnormality flag as an output for controlling a winding machine (100) and initiates a modification of one or more control parameters (P1, P2, P3, ...) of the winding machine (100) in case of an indicated abnormality, wherein the winding machine (100) comprises at least a winder (w) and a rewinder (r), the method comprising the following steps: - determining the actual speed (va) of the winder (w) during operation of the winder (100), - performing signal processing on said determined actual speed (vs) in order to extract winder-related features (fw), comprising subtracting a command speed (vc, vc') from said determined actual speed (va), determining an envelope signal of the subtracted signal (dv), and filtering said envelope signal, wherein said filtering preserves amplitude-related information, - using the filtered envelope signal as the winder related feature (fw), and - training the machine learning algorithm based on the winder related features (fw).

11. The method according to claim 10, wherein: Unsupervised training methods are used to identify clusters and determine the degree of abnormality of the input data based on the corresponding clusters (20, 21).

12. The method according to claim 10, wherein: Supervised training methods are used to identify categories based on labeled training datasets and determine the degree of abnormality of input data based on the corresponding categories (20, 21).

13. The method according to claim 12, wherein: The winder (100) in a training phase comprises a web tension sensor, wherein labeled data are generated depending on a value of the web tension sensor.

14. A computer program having instructions which, when executed by a computing device or system, cause the computing device or system to perform the method according to any one of claims 1 to 13.

15. A winding machine (100) having a data processing system, the data processing system comprising means for carrying out the method according to any one of claims 1 to 13.

Images