Method and device for detecting quality of synthetic thread

Abstract

The invention relates to a method for detecting the quality of a synthetic thread, in which at least one thread is extruded with a spinning device, the extruded thread is wound in a winding device into a thread reel, production and / or product parameters relating to the thread are recorded in the form of a time series before and / or during the winding of the thread into the thread reel, in order to detect the quality, the time series is analyzed by means of a data processing program for machine learning and / or for statistical time series analysis, and wherein at least one quality signal relating to the extruded and / or wound thread and / or relating to the extrusion and / or winding of the thread is generated on the basis of the result of the analysis of the data processing program.

Description

Technical Field

[0001] The invention relates to a method for detecting the quality of a synthetic thread. The invention also relates to an apparatus for detecting the quality of a synthetic thread. The invention also relates to an apparatus for producing a synthetic thread. Background Art

[0002] It is known to subject the synthesized threads or wound thread pools to a quality test after production. This testing step is carried out in particular in the form of a visual inspection, wherein the wound thread pools are tested for lint, incorrect construction, etc. However, this method only detects serious errors and quality problems.

[0003] It is known to carry out quality tests in laboratories on selected small quantities of silk or silk reels. For example, a single reel or a certain number of reels (particularly several reels of a single doff) are transported to the laboratory. There, relatively short sections of the wound silk are tested and checked. The data collected in the laboratory should then provide information on the quality of the reels produced during a whole day or a whole week.

[0004] However, quality fluctuations may occur during the production process, which cannot be reliably detected either by visual inspection or by testing random samples in the laboratory. Only a small number of produced wires or wire reels can be economically tested for their quality in the laboratory. Such random sampling may lead to quality problems going undetected and poor quality products being delivered to downstream customers. Therefore, quality monitoring in the laboratory after production does not guarantee a sufficiently high level of quality assurance. Moreover, such laboratory testing based on random samples can only detect quality problems at a later stage, if any. Until quality problems are detected, scrap or lower quality products are continuously produced and it is difficult to find any cause of the defect.

[0005] Furthermore, a method and a device for quality control when producing extruded polymer products are known from the prior art in WO 2006 / 013065A1. In the method according to WO 2006 / 013065A1, the production of the relevant polymer product is carried out continuously in a series of continuous process steps. For at least one of the process steps, the process parameters and / or product parameters are recorded and compared with a target value or a target value range. After the comparison, at least one quality signal for determining the quality of the polymer product and / or a control signal for controlling one of the process steps is generated. The process parameters and / or product parameters can also be assigned a measuring time and a conversion time, wherein the conversion time characterizes the material flow from the corresponding process step until the polymer product is completed.

[0006] However, by comparing the recorded process parameters and / or product parameters with target values ​​or target value ranges, only a limited amount of information about the respective process or the resulting product can be obtained. Furthermore, the target value or target value range must also be known, which can be particularly problematic in complex monitoring tasks. The quality signals generated in this way have a relatively low information content and therefore allow only limited quality monitoring or quality control in the production of extruded polymer products. Summary of the invention

[0007] In the above context, the object of the present invention is to define a method for quality detection of synthetic threads, by which quality detection is achieved with improved accuracy, informativeness or reliability during production under conditions of high productivity and cost-effectiveness. The object is also to define a device for detecting the quality of synthetic threads and a device for producing synthetic threads.

[0008] With regard to the method, this object is achieved by the subject matter of claim 1. A device according to the invention for testing the quality of a synthetic thread is defined in claim 24, and a device according to the invention for producing a synthetic thread is defined in claim 25. Advantageous embodiments are the subject matter of the respective dependent claims and are described below.

[0009] In a method for detecting the quality of a synthetic thread according to the invention, at least one thread is extruded by a spinning device and the extruded thread is wound into a reel in a winding device. Before and / or during the winding of the thread into the reel, production and / or product parameters related to the thread are recorded in the form of a time series. In addition, for quality detection, the time series is analyzed by a data processing program for machine learning and / or for statistical time series analysis, and at least one quality signal related to the extruded and / or wound thread and / or related to the extrusion and / or winding of the thread is generated based on the analysis results of the data processing program.

[0010] The recording of production and / or product parameters in the form of a time series makes it possible to assign a point in time (e.g. as a time stamp) to the respective data recorded in each case or an event related to production and / or product. Due to the recording and / or storage in the form of a time series, the recorded production and / or product parameters can be used in a particularly advantageous manner to determine the actual product quality.

[0011] Production and / or product parameters recorded in the form of a time series offer a particularly high information density and can therefore significantly increase the accuracy and informativeness of quality checks.

[0012] Furthermore, computer processing programs for machine learning and / or statistical time series analysis ensure the analysis of time series with particularly high reliability, security and informativeness. The use of machine learning also ensures a continuous improvement in the accuracy and reliability of the analysis with minimal expenditure.

[0013] In general, quality checks on extruded and / or wound threads and / or on the extrusion and / or winding of threads can be performed using a method according to the invention based on a large amount of available data and / or in real time, or within a short time after and / or during the production of synthetic threads. This ensures high accuracy, informativeness and reliability in quality checks.

[0014] By generating a corresponding quality signal, the quality and / or condition of the corresponding wire or the corresponding wire winding reel as well as the process quality of the production steps performed for the production can be reliably recorded and / or understood by the corresponding operating personnel. Thus, produced wires of lower quality can be detected particularly reliably. Thus, the risk of undesired use of inferior wires in the subsequent production process can be reduced or completely avoided.

[0015] According to a preferred embodiment of the method, the quality signal can include and / or represent the quality classification of the thread section, the thread, the thread reel and / or the doff consisting of a plurality of thread reels. In this way, depending on the quality classification, the corresponding thread, the corresponding thread reel or the doff consisting of a plurality of thread reels can continue to be used at a low cost.

[0016] Such a quality classification can be designed, for example, in such a way that one or all of the doffed thread reels are assigned to quality class "A", quality class "B" or quality class "C" by a data processing program or an algorithm implemented in a data processing program, depending on production and / or product parameters recorded in the form of a time series. Depending on the assignment to one of the categories, one or more reels can be passed on for appropriate further use or, if necessary, rejected as waste.

[0017] More preferably, the quality signal can be stored and / or assigned to a wire segment, a wire, a wire reel and / or a dolly consisting of a plurality of wire reels in the form of quality data. This enables appropriate recording, archiving and data use to be provided for subsequent process steps. In particular, the stored and specified quality data can be used to continuously improve the data processing program. Similarly, users, operators and / or customers can obtain from a database in order to retrieve the stored and specified quality data. The stored and specified quality data can be verified by the user and / or stored and / or specified again after verification. Subsequent laboratory tests can also be used as a basis for verification. In particular, the data verified by the user can be used to continuously improve the data processing program. The further use of the wire reel can be flexibly adjusted according to the stored and specified quality data, and the process organization can be improved.

[0018] According to another preferred embodiment, the time series analysis and / or the quality signal generation can be performed using or in particular using a predetermined and / or variable quality measure. The predetermined quality measure ensures a quality classification with sufficient reliability and safety. The ability to change the quality measure allows a continuous adaptation to the respective production conditions and / or a continuous improvement in the operation process. The quality measure can be defined and implemented according to the specific conditions of use.

[0019] Additionally or alternatively, the time series analysis and / or the quality signal generation can be performed using or in particular using at least one unified (undifferentiated) quality measure and / or at least one simple quality measure and / or at least one differentiated (differentiated) quality measure. Such quality measures can be realized with relatively little effort and ensure a sufficiently high level of information collection related to the silk thread, silk thread section or silk thread reel produced in each case.

[0020] More preferably, by the unified quality measurement, the classification of the quality signal can be performed according to the abnormality and / or normality of the analyzed time series. Such a unified quality measurement can achieve particularly reliable and effective classification. For example, in the unified quality measurement, a plurality of silk reels of a silk thread section, a silk thread reel or a doffing can be identified as "abnormal", wherein in this context, "abnormal" can mean that the product and / or production parameters recorded in the form of a time series deviate from normal expected values ​​or data to a certain extent. In particular, the monitored winding cycle can deviate from the mean value and therefore deviate from the normal winding cycle. Any deviation is captured by the product and / or production parameters recorded in the form of a time series.

[0021] For example, a wire reel classified as "abnormal" can be sent to the laboratory for further testing. This may show that in individual cases the deviations in the recorded parameters or data have no influence on the quality of the wire. It can also be determined during laboratory testing that the deviations lead to significant quality impairments and further use of the wire reel must be prevented. In the case of a wire reel classified as "normal", the observed winding cycles can correspond to average values ​​and therefore to normal winding cycles, and the product and / or production parameters recorded in the form of a time series contain at most minor deviations.

[0022] Still further preferably, using a simple quality measure, a classification of the quality signal can be performed based on a simple scale, in particular a simple scale with two categories. This allows obtaining more refined information about the produced thread or about the respective thread reel. A simple or binary grade scale can be defined, for example, by the grades "good" and "bad". Similarly, a simple grade scale for analyzing a thread reel or a plurality of thread reels of a doffing can be defined by classification or by the quality indicators "presence of broken thread" versus "absence of broken thread" or "presence of lint" versus "no lint".

[0023] Further preferably, using differentiated quality measurement, the classification of quality signals can be performed based on a multi-level scale. This allows obtaining even more sophisticated information about the produced silk thread or about the corresponding silk thread reel. In particular, the quality of a plurality of silk thread reels of a silk thread section, a silk thread reel or a doffing can be represented by a multi-level grade scale, in particular by a scale with more than two levels. For example, such a grade scale can be defined by dividing the grade scale into "A quality", "B quality" and "C quality". Being subdivided into "A quality", "B quality" and "C quality" is preferably based on the possibility of subsequent use of the corresponding silk thread reel. For example, "A quality" can represent unrestricted use in textile production. "B quality" can include restrictions, such as being only allowed to be used in selected use cases with lower quality requirements. "C quality" can, for example, represent waste.

[0024] According to another preferred embodiment, the data processing program can be designed to use empirical data, and / or labeled data, and / or processed and / or selected data to analyze the time series. Similarly, the data processing program can be designed to use unlabeled data to analyze the time series.

[0025] By including empirical data from the production of synthetic threads or thread reels wound with synthetic threads, a suitable basis can be provided for anomaly detection or for generating quality signals in future production processes. The empirical data can provide a relatively high information density about actual use and production conditions. During the production process of threads and / or thread reels, the empirical data can also be conveniently and comprehensively recorded and stored by continuous data acquisition. By marking, preparing and / or selecting data, especially empirical data, the information content inherent in the data can be further increased.

[0026] In an even more preferred manner, the data processing program can be trained with empirical and / or labeled and / or processed and / or selected and / or unlabeled data. By training the data processing program, a sufficiently high level of reliability in the generation of the information quality signal can be ensured at an early stage after the initial installation.

[0027] According to another preferred embodiment, the marked data may have at least one tag related to abnormal wire extrusion and / or abnormal spool winding. Additionally or alternatively, for marking the data, at least one result from the wire and / or spool test may be linked to the associated production and / or product parameters of the corresponding wire and / or the corresponding wire spool. In this way, the data set for generating quality signals in future production processes may be further improved.

[0028] For example, in the data, anomalies or exceptions related to wire extrusion and / or spool winding can be marked, for example, by a process expert for producing synthetic wire and / or wire spools of said wire. Data representing anomalies can also be selected. This allows for appropriate preparation of empirically collected data.

[0029] In the absence of such data preparation, selection or labeling, only unlabeled data is available. Unlabeled data can be generated in large quantities with less effort and used for further use or subsequent processing.

[0030] Further preferably, the data processing program can be trained with data of normal wire extrusion and / or normal spool winding. In particular, the data processing program can be trained exclusively with data of normal wire extrusion and / or normal spool winding. Thus, the corresponding data (in particular empirically recorded data) can be processed in such a way that only data without anomalies (i.e. data representing a normal or expected winding cycle or wire production process) are reused. Detection of deviations with respect to such processed or selected data can be achieved with a high degree of reliability.

[0031] Further preferably, the data processing program can be designed to perform supervised learning, in particular using labeled and / or processed data. By using supervised learning, new data points can advantageously be classified as abnormal. Particularly preferably, using supervised learning, new data points can be classified within simple and / or differentiated quality measures, as described above. This enables immediate and particularly reliable quality detection by the data processing program in production operation. Furthermore, the data processing program can be continuously developed in a particularly advantageous manner in this way.

[0032] Particularly preferably, the data processing program for supervised learning can be designed and / or constructed by logistic regression methods, random forest methods and / or gradient boosting tree methods. Such methods ensure high accuracy and reliability in the analysis of recorded data, thereby achieving particularly advantageous quality detection.

[0033] Further preferably, the data processing program can be designed for semi-supervised learning, in particular using processed and / or selected data. Such a design enables particularly good results to be achieved in quality testing with a relatively low amount of effort. In particular, it makes it possible that the training data can be processed and / or selected in such a way that it does not exhibit any anomalies. The semi-supervised method then learns the data distribution of normal or typical winding cycles and can use this to classify new data points as anomalies.

[0034] In a particularly preferred manner, the data processing program can be designed to perform semi-supervised learning using a one-class support vector machine method. This enables a relatively high degree of accuracy and reliability in analyzing the collected data and at the same time allows for appropriate refinement of the data processing program.

[0035] Still further preferably, the data processing program can be designed to perform unsupervised learning, in particular using unlabeled data. The design of unsupervised learning can be implemented with a particularly small amount of initial work, while at the same time ensuring a sufficiently high level of success in terms of quality detection. In the case of unsupervised learning, it can be assumed, for example, that the training data used contains a small proportion of abnormal data and thus a normal or typical distribution can be learned. Anomalies can then be identified using an anomaly measure or "anomaly scale" that indicates the deviation of new data points from this distribution. In this context, for example, isolation forests, autoencoders or various clustering algorithms can be applied.

[0036] In a particularly preferred manner, the data processing program for unsupervised learning can be designed using isolation forest methods, autoencoder methods and / or methods based on clustering algorithms. These embodiments and methods ensure high operational reliability and relatively high accuracy in the analysis of recorded data.

[0037] According to a further preferred embodiment, device data, in particular data relating to the spinning device and / or winding device, and / or device log data, and / or sensor data, and / or data relating to the doffing can be collected as production and / or product parameters. Particularly preferably, data relating to the entire production process can be collected starting from the feeding of the plastic granules and / or plastic material into the extruder, the polycondensation phase up to the winding phase, and taken into account by the data processing program for the analysis. This provides a particularly comprehensive database for the quality inspection of the yarn and / or the yarn winding reel.

[0038] Further preferably, performance data and / or power consumption data of the electric drive, torque data of the electric drive, pump data, data about the material melt for wire production, and / or data about melt pressure and / or melt viscosity, and / or data about wire tension can be collected as production and / or product parameters. Thereby the database for quality inspection of wire and / or wire winding reels can be further improved.

[0039] In an even more preferred manner, data related to extrusion and / or to the extrusion device can be collected as production and / or product parameters. In particular, this can include data related to the power consumption of the drive for the extruder or the extrusion screw, and / or data related to the drive speed of the extruder or the extrusion screw, and / or data related to the state of the extruder, the pressure in the extruder and / or the temperature in the extruder and / or the temperature in a section and / or in different sections in the extruder, and / or data related to the melt pressure in front of the spinning pump and / or the melt pressure behind the spinning pump, and / or the temperature in the spinning nozzle and / or in the spinning rod, and / or data related to the temperature in a ventilation container and / or the flow rate and / or the temperature of the coolant for the extruder and / or the spinning device.

[0040] In an even more preferred manner, data related to the melt spinning process and / or to the melt spinning device can be collected as production and / or product parameters. In particular, they can be data related to the spinning pump (in particular, related to the operating speed, power consumption, torque and / or operating state of the spinning pump), and / or data related to the preparation pump (in particular, related to the operating speed, power consumption, torque and / or operating state of the preparation pump), and / or data related to the godet roller (in particular, related to the power consumption, torque and / or operating state of the godet roller), and / or data related to the chuck device for the thread reel (in particular, related to the power consumption, torque and / or operating state of the chuck device), and / or data related to the speed of the pressure roller for the thread reel, and / or data related to the diameter of the thread reel, and / or data related to the winding duration and / or the remaining winding duration, and / or data related to the angle of winding on the thread reel, and / or data related to the transverse speed during winding.

[0041] By recording the above data or parameters, a particularly detailed database for detecting the quality of the wire and / or the wire winding reel can be provided, thereby further improving the success level of the quality detection.

[0042] According to a further preferred embodiment, plant external data can be collected as production and / or product parameters. Preferably, these can be environmental data, spatial data, temperature data and / or climate data. In particular, spatial data, temperature data and / or climate data of the production facility can be recorded. The database of the data processing program can be further refined and improved in this way, thereby ensuring a further improvement in the quality of the inspection.

[0043] According to a further preferred embodiment, the quality signal may have and / or represent a quality classification with a specified probability and / or a plurality of quality classifications each with a specified probability. The probability may preferably be quantified. Additionally or alternatively, the quality signal may have and / or represent a quality classification with a specified anomaly metric, wherein the anomaly metric is preferably a distance metric and / or a similarity metric associated with at least one cluster or a plurality of clusters, and / or a reconstruction error. This allows conclusions to be drawn about the quality of the respective thread and / or the respective thread reel with relatively high certainty and limited effort, and appropriate measures may be initiated based on the quality detection.

[0044] Such classification may be designed (by way of example only) such that the wire, wire spools and / or multiple wire reels of the dolly are assigned by the data processing program to quality category "A" with a probability of 80%, quality category "B" with a probability of 19%, and quality category "C" with a probability of 1%.

[0045] According to a further preferred embodiment, the data processing program can be trained with transformed time series data, in particular empirical time series data. Such transformation can preferably include fitting the data to a specific algorithm and / or removal of outliers and / or interpolation of missing data. This allows to improve the accuracy of the analysis performed by the data processing program in a particularly advantageous manner. The transformation can also be used to improve the model quality.

[0046] In an even more preferred manner, the data processing program can be designed for transforming the continuously recorded time series data, in particular for a transformation corresponding to the transformation of the training data performed in each case. This allows further improving the accuracy of the analysis performed by the data processing program during operation.

[0047] Further preferably, the time series data of continuous recording can be collected by data processing program and analyzed for quality detection subsequently.Additionally or alternatively, the time series data of continuous recording can be collected by data processing program and analyzed for quality detection simultaneously.Subsequent quality detection can be performed optionally in a series of iterative stages, and the accuracy of increase is provided thus.The analysis carried out while data collection or data acquisition enables early classification of silk thread or silk reel, and also enables early adjustment of process thus, or enables early decision to be made to the further use of corresponding silk thread or silk reel.

[0048] According to a further preferred embodiment, at least one current and / or future production parameter and / or product parameter can be changed based on the quality signal. Additionally or alternatively, the quality signal can be transmitted to the device controller and / or the device control system for adjusting at least one current and / or future production parameter and / or product parameter. In this way, the quality signal from the production process can be used particularly advantageously to improve the current and / or future production process. Thereby, the production can be improved immediately and continuously, so that the silk thread and / or silk thread reel can be produced with a permanently improved or at least constant quality.

[0049] According to a further preferred embodiment, a quality signal can be output, in particular as an optical and / or acoustic signal and / or a signal directed to an operator. For example, the signal output in this way can be a warning signal. Such a signal can be used to indicate a critical quality to remind the operator that the corresponding silk reel needs to be manually marked. Such a signal can also be used to indicate the need for further quality testing of the corresponding silk reel. The output of such a signal means that the decision on the further use of the correspondingly produced or wound spool can be simplified.

[0050] Another aspect of the present invention relates to a device for detecting the quality of synthetic yarns, in particular using the above-mentioned method, the device having at least one recording device for recording production and / or product parameters of extruded and / or wound yarns in the form of time series, the recording device having a data processing device for machine learning and / or for statistical time series analysis, wherein the data processing device is designed to analyze the time series recorded by the recording device for the purpose of quality detection, and wherein the data processing device generates at least one quality signal related to the extruded and / or wound yarn and / or related to the extrusion and / or winding of the yarn based on the analysis results.

[0051] Thus, by means of the recording device according to the invention, production and / or product parameters can be recorded in the form of a time series, and this enables a time (e.g. as a time stamp) to be assigned to the respective data or to the respective recorded production and / or product-related events. Due to the recording and / or storage in a time series, the production and / or product parameters recorded by the recording device can be used particularly advantageously to determine the actual product quality.

[0052] The production and / or product parameters recorded in time series by the recording device offer a particularly high information density and thus enable a significant increase in the accuracy and informativeness of the quality inspection.

[0053] Data processing programs for machine learning and / or statistical time series analysis enable the analysis of time series to be performed with a high degree of reliability, security, and informativeness. The use of machine learning can also continuously improve the accuracy and reliability of the analysis with only minimal effort.

[0054] In general, the device according to the invention enables quality detection of extruded and / or wound yarns and / or quality detection related to the extrusion and / or winding of yarns based on a large number of available data points. The quality detection can be carried out in real time, also during and / or shortly after the production of the synthetic yarn. This ensures high accuracy, effectiveness and reliability of the quality detection.

[0055] Another independent aspect of the invention relates to an apparatus for producing synthetic threads, comprising a spinning device for extruding synthetic threads, a winding device for winding the extruded threads into a reel, and the above-mentioned apparatus for quality inspection of synthetic threads.

[0056] The preferred embodiments and advantages described above with respect to the method for detecting the quality of synthetic threads also apply in the same manner to the apparatus for detecting the quality of synthetic threads and the apparatus for producing synthetic threads according to the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In the following, the invention is explained in more detail based on examples and with reference to the accompanying drawings, in which in each case:

[0058] Figure 1 shows a device for detecting the quality of a synthetic thread according to an exemplary embodiment of the present invention;

[0059] Figure 2 shows an apparatus for producing synthetic yarn according to an exemplary embodiment of the present invention;

[0060] Figure 3 A method for detecting the quality of a synthetic thread according to an exemplary embodiment of the present invention is shown;

[0061] Figure 4 shows a diagram of a plurality of recorded time series according to an exemplary embodiment of the present invention;

[0062] Figure 5 shows a diagram of a plurality of recorded time series according to an exemplary embodiment of the present invention;

[0063] Figure 6 A diagram showing transformation and data processing of a plurality of recorded time series according to an exemplary embodiment of the present invention;

[0064] Figure 7 A diagram showing transformations and data processing when applying a trained data processing program to produce wire and / or wire spools is shown. DETAILED DESCRIPTION

[0065] Figure 1 A schematic block diagram of an apparatus 10 for detecting the quality of synthetic silk threads according to an exemplary embodiment of the present invention is shown. The apparatus 10 may be designed and / or constructed in particular for carrying out the method described below.

[0066] The device 10 has at least one recording device 12 for recording production and / or product parameters 14 of the extruded and / or wound thread in the form of a time series. The recording device 12 may include storage means 16 for storing the recorded production and / or product parameters 14 in the form of a time series.

[0067] Furthermore, the device 10 comprises a data processing device 18 for machine learning and / or for statistical time series analysis. In this case, the data processing device 18 can be configured to analyze the time series recorded by the detection device 12 for quality detection. Furthermore, the data processing device 18 can be designed and / or configured to generate at least one quality signal 20 related to the extruded and / or wound wire and / or related to the extrusion and / or winding of the wire based on the analysis results.

[0068] The data processing device 18 may be designed and / or constructed to execute a data processing program 22 .

[0069] The data processing device 18 and / or the data processing program 22 may include a data preprocessing unit 24 for preprocessing the recorded time series. Such preprocessing may include preparing and / or selecting data from the recorded time series.

[0070] In addition, the data processing device 18 and / or the data processing program 22 may include a data analysis unit 26, which is designed and / or constructed to further analyze or process the pre-processed time series data. In particular, machine learning can be implemented and executed by the data analysis unit 26 in operation.

[0071] Figure 2 A device 100 for producing a synthetic thread according to an exemplary embodiment of the invention is shown. The device 100 comprises at least one spinning device 28 for extruding the synthetic thread and at least one winding device 30 for winding the extruded thread into a reel. In addition, the device 100 is equipped with a device 10 for detecting the quality of the synthetic thread as described above.

[0072] The apparatus 100 may also include a godet device 32, which is arranged between the spinning device 28 and the winding device 30. Furthermore, the apparatus 100 may include an extruder 34. A discharge pump 36 with a pump drive 38 and a further end filter device 40 may be located downstream of the extruder 34.

[0073] The device 10 for testing the quality of synthetic yarns can be constructed and / or designed to record and / or analyze production and / or product parameters 14 of different areas or parts of the device 100. These parameters include (only by way of example) production and / or product parameters 14 related to the spinning device 28, and / or to the winding device 30, and / or to the godet device 32, and / or to the extruder 34, and / or to the discharge pump 36 with the pump drive 38, and / or to the end filter device 40.

[0074] Figure 3 A method for detecting the quality of a synthetic thread according to an exemplary embodiment of the present invention is shown.

[0075] In step 200 of the method, at least one filament may be extruded using spinning device 28 .

[0076] In a further step 202 of the method, the extruded filament can be wound onto a reel in a winding device 30 .

[0077] In a further step 204 of the method, production and / or product parameters 14 related to the filament can be recorded in the form of a time series. The recording of the production and / or product parameters 14 in the form of a time series in step 204 can be performed before and / or during the winding of the filament into a reel. Furthermore, the recording of the production and / or product parameters 14 in the form of a time series in step 204 can be performed before and / or during the extrusion and / or melt spinning of a melt, in particular a polymer melt, to form a synthetic filament.

[0078] Thus, when recording the production and / or product parameters 14 in the form of a time series in step 204 , source systems for extruding the melt and / or for melt spinning and / or for winding the filament can be accessed.

[0079] In a further step 206 , the production and / or product parameters 14 recorded in the form of a time series can be stored and / or buffered in the data memory 16 .

[0080] In a further step 208 of the method, the recorded and / or stored time series can be analyzed by a data processing program 22 for machine learning and / or statistical time series analysis in order to perform a quality check.

[0081] Step 208 of the method may also include steps 210 and 212 .

[0082] In step 210, the recorded and / or stored time series may be preprocessed, ie by a data preprocessing unit 24. Preprocessing may include marking, preparing and / or selecting data, in particular recorded production and / or product parameters 14.

[0083] In step 212, the recorded and / or stored and / or preprocessed time series may be analyzed by the data analysis unit 26. In this case, step 212 may be a method step in which an analysis using statistical time series analysis and / or machine learning is performed by the data analysis unit 26. In step 212, in particular an inference procedure or model may be executed.

[0084] In step 214 at least one quality signal related to the extruded and / or wound wire and / or related to the extrusion and / or winding of the wire may be generated based on the analysis results of the data processing program 22 or based on the analysis results from step 208 .

[0085] Step 214 may include step 216 and / or step 218 .

[0086] In step 216, a quality signal may be output to the operator. This may be an optical and / or acoustic quality signal. Such a quality signal output to the operator may also additionally or alternatively be a digital work instruction.

[0087] In step 218, the quality signal can be output to a machine and / or process controller. Thus, in step 218, for example, a signal can be output that can have an influence on step 200 in which the filament is extruded by means of the spinning device 28 and / or step 202 in which the filament is wound onto a reel in the winding device 30.

[0088] In the following, reference is made to Figures 4 to 7Further aspects of the invention are described, in particular relating to the training of the data processing device 18 and / or the data processing program 22 and to data analysis performed by the data processing device 18 and / or the data processing program 22 before and / or during yarn production.

[0089] As part of the model development, machine data, machine log data and / or further sensor data may be collected over a long period of time, and in the case of supervised learning, quality data may be collected. The collected data may be empirical data, in particular data in the form of time series data. In time series data, each data collection may have a timestamp assigned to it.

[0090] The raw data collected or recorded in this way can be converted into a form suitable for model training and optionally cleaned of outliers. Additionally or alternatively, missing data can be interpolated. Furthermore, the data preprocessed in this way can be converted into a form in which basic features are calculated in a way that meets the requirements of the algorithm to be used and / or improves the quality of the model. This process can be called "feature engineering".

[0091] The model in the form of a data processing program 22 or data processing device 18 can be trained on the basis of training data and validated on the basis of test data. The aim is to achieve high model quality on the test data so that the model can also classify observations with high quality that are not part of the training data set.

[0092] In production, the model in the form of a data processing program 22 or a data processing device 18 can be used for analysis and / or inference (e.g. for classifying the doffing into a plurality of quality classes). It may be important here to apply the data preprocessing (including "feature engineering") defined during development and training. For example, if the doffing occurs at time t, the data from the time period tp to tq (where p>q) can be used to assign the doffing to the corresponding quality class.

[0093] Based on an exemplary embodiment, the data preparation for data of a single position is described below. In addition, an example of supervised learning for dolly classification will be studied in more detail.

[0094] In accordance with Figure 4 In the illustration of , a time series S is collected for a location or a product and / or production parameter 14. The product and / or production parameter 14 can be, for example, the current intensity of the corresponding godet rollers, their torque or the rotational speed of the spinning pump.

[0095] Each time series S is represented in a row. For each time series S there is a data point X ST Data point XST represents a specific value of a specific time series at time T. For simplicity reasons, for example, assume that for all time series S, there are available data points X ST The numbers are the same and therefore the subscript can be omitted.

[0096] These multivariate time series S have a task-related structure in that they can be assigned to individual wrapping cycles.

[0097] In accordance with Figure 5 In the exemplary embodiment of , products and / or production parameters 14 are shown for a total of four dollies D for the considered time period. For each dollie D, there may be a plurality of time series S. For simplicity, it is assumed that there are no runtime differences (eg due to different products or thread damage).

[0098] The further steps of data processing vary depending on the method used for feature engineering or machine learning. In the following, the data can first be transformed from the form S x T to the form (T+D) x S. This now allows the time series S to be defined by column and the doffing D and time to be defined by row. This transformation to the so-called "long format" is a common method for representing time series data. This means that a so-called multi-index can be placed on the row representing the corresponding time T and doffing D or winding cycle.

[0099] In this step, standardized time indexes can be used instead of the original timestamps. For example, if dolly D is wound or produced at time T, the previous times can be assigned indices T-1, T-2, ... Figure 6 This transformation is shown on the left side of . Based on this dataset, so-called features can be calculated that can be used for training and evaluation of the model.

[0100] In this example, during the feature engineering FE, two features can be calculated for each of the nine time series S1 to S9, where each time series has seven observations at time T (e.g. mean and standard deviation as a measure of dispersion). This results in a final dataset with D rows (4 rows in the exemplary embodiment, corresponding to the number of droppers) and S xp columns (18 columns in the exemplary embodiment, from the nine time series S, two features being calculated for each time series in each case). Figure 6 Shown on the right.

[0101] Again, for simplicity, it can be assumed that the number of features does not differ between the time series S, thus being able to omit appropriate indexing.

[0102] Since in this case we are considering an example of supervised learning, we can now assign a corresponding quality indicator Q to each feature vector. This can also be generally referred to as a target. Depending on the framework used, this target can also be represented as a matrix. The quality indicator Q or target can be output in the form of a quality signal 20.

[0103] This process can be carried out in multiple winding cycles, which has been Figure 6 denoted by N. This data can then be fed to a training and evaluation process TE. The final model M can be serialized and stored in a file, either in the format of the data processing program 22 and / or in the format of the data processing device 18.

[0104] Reference Figure 7 , the following describes the application of the model M or the data processing program 22 and / or the data processing device 18 developed in this way in production. In production, the model M trained in the above-described manner can be used for reasoning or analysis. The time series S can be recorded in the same way as the time series S used for model development. For simplicity, it is assumed that the final model M requires all time series S, and it is assumed that no further selection is made during model development.

[0105] If the quality of dolls D, in particular dolls D with multiple reels, is to be predicted, data from the beginning to the end of the respective winding cycle can be recorded and collected in the form of a time series S. These data can then be processed using the processing steps also used in the model development and provided to the model M or the data processing program 22 and / or the data processing device 18 for inference and / or analysis.

[0106] The processing and / or preparation of the data results in the generation of a vector V indicating or representing the probability of membership in a particular class, or a quality indicator Q having a probability. Data points may be assigned to a class or a quality indicator Q by the model M or by the data processing program 22 and / or the data processing device 18 based on boundary values ​​to be defined or other different criteria.

[0107] This enables a quality inspection of synthetic threads and / or thread spools to be carried out with a high degree of reliability and with relatively little effort.

[0108] The source systems, i.e. for example the spinning device 28 and / or the winding device 30, can be diverse and complex. Depending on the conditions of use, the recorded products and / or production parameters 14 can be different. Furthermore, the recording device 12 can be complex in design, in particular having access to a plurality of sensors, not shown in detail here, in order to acquire a sufficiently high amount of data.

[0109] Furthermore, as part of supervised and / or unsupervised learning, it may be necessary to set thresholds for membership probabilities and / or anomaly scores, which are not explained in detail here. Additionally or alternatively, the results of unsupervised learning can be used for supervised learning.

[0110] The data processing device 18 and / or the data processing program 22 or the model M according to the invention can be continuously evaluated and improved, which is also not explained in detail here. Updating the model M by retraining the model can be performed in a particularly advantageous manner.

[0111] List of Reference Numerals

[0112] 10 Equipment for testing the quality of synthetic yarn

[0113] 12 Recording Device

[0114] 14 Product and / or production parameters

[0115] 16 Storage Device

[0116] 18 Data processing device

[0117] 20 Quality Signal

[0118] 22 Data Processing Procedures

[0119] 24 Data preprocessing unit

[0120] 26 Data Analysis Unit

[0121] 28 Spinning device

[0122] 30 Winding device

[0123] 32 Godet Roller Device

[0124] 34 Extruder

[0125] 36 Discharge pump

[0126] 38 Pump driver

[0127] 40 End filter

[0128] 100 Equipment for producing synthetic yarn

[0129] 200 Steps of a method for testing the quality of synthetic yarn

[0130] 202 Steps of a method for testing the quality of synthetic yarn

[0131] 204 Steps of a method for testing the quality of synthetic yarn

[0132] 206 Steps of a method for testing the quality of synthetic yarn

[0133] 208 Steps of a method for testing the quality of synthetic yarn

[0134] 210 Steps of a method for testing the quality of synthetic yarn

[0135] 212 Steps of a method for testing the quality of synthetic yarn

[0136] 214 Steps of a method for testing the quality of synthetic yarn

[0137] 216 Steps of a method for testing the quality of synthetic yarn

[0138] 218 Steps of a method for testing the quality of synthetic yarn

[0139] S Time Series

[0140] T Time (point)

[0141] X ST Data Points

[0142] D Dropping

[0143] M Model

[0144] FE Feature Engineering

[0145] Q Quality indicator

[0146] TE Training and Assessment Process

[0147] V Vector.

Claims

1. Method for testing the quality of synthetic yarn, -in, extruding at least one filament using a spinning device (28), - wherein the extruded filament is wound onto a reel in a winding device (30), - wherein production and / or product parameters (14) related to the thread are recorded in the form of a time series (S) before and / or during winding the thread onto the reel, - wherein, in order to perform the quality check, the time series (S) is analyzed by a data processing program (22) for machine learning and / or statistical time series analysis, and - wherein at least one quality signal (20) is generated on the basis of the analysis result of the data processing program (22) related to the extruded and / or wound thread and / or related to the extrusion and / or winding of the thread.

2. The method according to claim 1, characterized in that: The quality signal (20) comprises and / or represents a quality classification of a thread section, a thread, a thread reel and / or a dob (D) consisting of a plurality of thread reels.

3. The method according to claim 1 or 2, characterized in that: The quality signal (20) is stored in the form of quality data and / or is assigned to a thread section, a thread, a thread reel and / or a dob (D) consisting of a plurality of thread reels.

4. The method according to any one of the preceding claims, characterized in that The analysis of the time series (S) and / or the generation of the quality signal (20) is performed by using or in particular using a predetermined and / or variable quality metric, preferably based on at least one unified quality metric and / or at least one simple quality metric and / or at least one differentiated quality metric.

5. The method according to claim 4, characterized in that Using the unified quality metric, classification of the quality signal (20) is performed according to abnormality and / or normality of the analyzed time series (S).

6. The method according to claim 4 or 5, characterized in that: Classification of the quality signal (20) is performed using a simple quality metric based on a simple scale, in particular a simple scale with two categories.

7. The method according to any one of claims 4 to 6, characterized in that Classification of the quality signal (20) is performed based on a multi-level scale using a differentiated quality metric.

8. The method according to any one of the preceding claims, characterized in that The data processing program (22) is designed to analyze the time series (S) using empirical data and / or labeled data and / or processed and / or selected data and / or unlabeled data.

9. The method according to any one of the preceding claims, characterized in that The data processing program (22) is trained with empirical and / or labeled and / or processed and / or selected and / or unlabeled data.

10. The method according to claim 8 or 9, characterized in that: The marked data has at least one tag related to abnormal wire extrusion and / or abnormal spool winding and / or in order to mark the data, at least one result from a wire and / or spool test is linked to a relevant production and / or product parameter (14) of the corresponding wire and / or the corresponding wire spool.

11. The method according to any one of the preceding claims, characterized in that The data processing program (22) is trained with data from normal wire extrusion and / or normal spool winding, in particular trained exclusively with data from normal wire extrusion and / or normal spool winding.

12. The method according to any one of the preceding claims, characterized in that The data processing program (22) is designed for performing supervised learning, in particular using labeled and / or processed data, preferably by using methods for logistic regression, random forest methods and / or gradient boosted tree methods.

13. The method according to any one of the preceding claims, characterized in that The data processing program (22) is designed to perform semi-supervised learning, in particular using processed and / or selected data, preferably using a one-class support vector machine method.

14. The method according to any one of the preceding claims, characterized in that The data processing program (22) is designed to perform unsupervised learning, in particular using unlabeled data, preferably using an isolation forest method, an autoencoder method and / or a clustering algorithm.

15. The method according to any one of the preceding claims, characterized in that Device data and / or device log data and / or sensor data are recorded as production and / or product parameters (14), wherein the device data are in particular data related to a spinning device (28) and / or a winding device (30) and / or an extruder (34) and / or a discharge pump (36) of an extruder (34).

16. The method according to any one of the preceding claims, characterized in that External data of the plant are collected as production and / or product parameters (14), the external data preferably being environmental data, spatial data, temperature data and / or climate data.

17. A method according to any one of the preceding claims, characterised in that Performance data and / or power consumption data of the electric drive, torque data of the electric drive, pump data, data about the material melt used to produce the wire, and / or data about melt pressure and / or melt viscosity, and / or data about wire tension are recorded as production and / or product parameters (14).

18. A method according to any one of the preceding claims, characterised in that The quality signal (20) has and / or represents a quality classification with a specified probability and / or a plurality of quality classifications each having a specified probability, wherein the probability is preferably quantified, and / or the quality signal has and / or represents a quality classification with a specified anomaly metric, wherein the anomaly metric is preferably a distance metric and / or a similarity metric associated with at least one cluster or a plurality of clusters, and / or a reconstruction error.

19. The method according to any one of the preceding claims, characterized in that The data processing program (22) is trained with transformed time series data, in particular empirical time series data, wherein the transformation preferably includes fitting a specific algorithm and / or removing outliers and / or interpolating missing data.

20. The method according to any one of the preceding claims, characterized in that The data processing program (22) is designed for the transformation of continuously recorded time series data, in particular for a transformation corresponding to the transformation of training data performed in each case.

21. The method according to any one of the preceding claims, characterized in that The data processing program (22) collects the continuously recorded time series data and analyzes them subsequently and / or simultaneously for quality inspection.

22. The method according to any one of the preceding claims, characterized in that At least one current and / or future production and / or product parameter (14) is changed based on the quality signal (20), and / or the quality signal (20) is transmitted to a device controller and / or a device control system for adjusting at least one current and / or future production and / or product parameter (14).

23. A method according to any one of the preceding claims, characterised in that A quality signal (20) is output, in particular as an optical and / or acoustic signal and / or as a signal directed to an operator.

24. A device (10) for testing the quality of a synthetic thread, in particular using a method according to any one of the preceding claims, - having at least one recording device (12) for recording production and / or product parameters (14) of the extruded and / or wound thread in the form of a time series, - having data processing means (18) for machine learning and / or statistical time series analysis, -in, The data processing device (18) is configured to analyze the time series (S) recorded by the recording device (12) for quality detection, and -in, The data processing device (18) generates at least one quality signal (20) related to the extruded and / or wound thread and / or related to the extrusion and / or winding of the thread based on the analysis result.

25. A device (100) for producing synthetic yarn, - having a spinning device (28) for extruding synthetic yarns, - having a winding device (30) for winding the extruded filament into a reel, and A device (10) for testing the quality of synthetic threads according to claim 24 is provided.

Images