Method for monitoring the filling process of a brush filling machine

Abstract

The invention relates to a method for monitoring the filling process of a brush filling machine (1), in particular a toothbrush filling machine, using an AI-based classifier (34), characterised in that: during the filling process, a time series (32) of a filling force (25) acting on a rammer tongue (3) when filling bristle bundles (4) is determined; and the AI classifier (34) uses a plurality of input parameters determined from this time series (25) and indicates, as an output, whether an error condition is present, wherein the AI classifier (34) has been trained using input parameters that represent an error-free filling process.

Description

[0001] PC 240230 J March 25, 2024 Method for monitoring the stuffing process of a brush stuffing machine. Method for monitoring the stuffing process of a brush stuffing machine, in particular a toothbrush stuffing machine, using an AI-based classifier. Brush stuffing machines are used to manufacture brushes and, specifically, to stuff bristle bundles into the bundle receiving holes of brush bodies, thus creating a bristle covering for a brush. For this purpose, brush stuffing machines have a stuffing tool with a pusher tongue, which is designed and intended to stuff bristle bundles into the bundle receiving holes of a brush body. As a rule, the bristle bundles are stuffed into the bundle receiving holes together with fastening anchors and anchored there. For a perfect stuffing result, it is necessary that various parameters are observed,so that flawless brushes are produced. These parameters include, for example, the length and alignment of the pusher tongue, the number of filaments, and the length of the armatures. If one of these parameters deviates from its norm, the produced brushes may be faulty or poorly equipped. The object of the invention is to create a method with which the stuffing process can be monitored and thus faulty and / or poorly manufactured brushes and / or errors in the brush stuffing machine can be detected. PC 240230 J 2 / 32 March 25, 2024 This object is achieved by a method having the features of claim 1. The method according to the invention is therefore characterized in that, during the stuffing process, a time series of a stuffing force acting on a pusher tongue during the stuffing of bristle bundles is determined, that the AI ​​classifier uses several input parameters determined from this time series and indicates as an output whether a fault condition exists,The AI ​​classifier was previously trained with input parameters that represent a defect-free tucking process. Monitoring the quality of the brushes would normally require an optical inspection, which is very complex and cost-intensive. The advantage of the invention is that the condition of the produced brushes can be determined from the measured values ​​of a simple force sensor using the AI ​​classifier. Monitoring the tucking process can thus be carried out simply and cost-effectively. In particular, such a force sensor can already be present on the brush tucking machine, so that its measured values ​​can be used directly. The AI ​​classifier is trained exclusively with input parameters that represent defect-free states, i.e., "good." In this way, a deviation of the input parameters from the training data can be easily identified as a defect state, i.e., "bad."This allows for a significant reduction in training time compared to training with poor values. To determine the time series, measured values ​​of the tamping force are recorded at a fixed time interval. The number of measured values ​​is determined by a sampling rate. PC 240230 J 3 / 32 March 25, 2024 The sampling rate is preferably in the double-digit kilohertz range. In one embodiment, a time series, particularly a univariate one, is determined from the tamping force for each tamping process. This means that one time series corresponds exactly to one tamping process at one tamping hole.thus reflects one cycle of the tamping machine. This number of measured values ​​could indeed be used as input parameters for the AI ​​classifier. However, the large number of input parameters can increase the complexity of the AI ​​classifier. Therefore, a large amount of memory and / or computing time may be required for the calculation. In one embodiment, the input parameters are determined from such a time series through feature extraction and / or data reduction. This reduces the number of input parameters, thereby reducing the complexity of the AI ​​classifier and allowing the calculation to be performed faster and with fewer resources. In one embodiment, a predetermined number of Fourier coefficients are determined as features during feature extraction. A Fourier transform or fast Fourier transform, i.e., a discrete Fourier transform, is applied to the time series.applied. The input parameters are limited to a specific number of Fourier coefficients, for example, the first 80. Alternatively or additionally, a predetermined number of wavelet coefficients can be determined as features during feature extraction, for example, by a discrete wavelet transform. Haar wavelets or Daubechies wavelets, in particular with values ​​for N between 4 and 20, or Coiflet wavelets with N between 2 and 4, are expediently used as wavelets. For all wavelets, for example, a level between 1 and 4 is used. The value N and the level are preferably chosen such thatthat the number of wavelet coefficients is approximately 80. In another alternative embodiment, or in addition, a distance to the median is determined as features during feature extraction for a predetermined number of intervals of the time series. Here, too, the predetermined number can be approximately 80. Through feature extraction, the number of input parameters can be significantly reduced without causing a significant loss of information. Alternatively or additionally, in an embodiment for data reduction, the number of measured values ​​in a time series can be reduced, in particular so that 200 measured values ​​remain per time series. In this embodiment, subsampling is performed, so to speak, which reduces the number of measured values ​​in the time series. For this purpose, for example, measured values ​​can simply be removed from the time series. However, it is also possible to calculate an average from several neighboring measured values.which is used as an input parameter. This allows direct use of the measured values ​​without the need for feature extraction calculations. This can lead to faster processing and reduced computational effort. In one embodiment, a multivariate time series is alternatively and / or additionally created from a predetermined number of time series. As already described above, one time series corresponds to a tamping process at a tamping hole. Several such time series are now combined into a multivariate time series. This means that, for example, the time series from 100 tamping holes are combined. The number of tamping holes per brush is irrelevant. For example, if a brush has 40 tamping holes, the multivariate time series comprises 2,5 brushes. Of course, these figures are only examples and can be easily adapted to the application. Such a multivariate time series can serve as the basis for determining the input parameters as described above. This means that features are extracted from such a multivariate time series. In an alternative embodiment, feature extraction can occur first, and the multivariate time series is formed from these features. Regardless of this, the multivariate time series is used as an input parameter. This way, an AI classifier calculation is not necessary for each hole individually. Instead, the measured values ​​from several plug holes are first collected and then considered as a unit. Although this does not allow for an error statement for each plug hole or brush, this is not absolutely necessary in practice to detect a deviation from normal operation during the operation of a brush tamping machine.i.e., to detect an error condition. Typically, an AI or an AI classifier is trained before its use. In one embodiment, a predetermined number of time series are determined during operation, and the resulting input parameters are defined as good, for example, by a user. In this way, the AI ​​classifier can be trained "on the fly" during operation. For example, when brush production starts up, a certain number of time series are recorded, which are then defined or confirmed as good by a user, for example, by viewing the brushes. The input parameters determined from these time series serve as training data for the AI ​​classifier. During further operation, the input parameters are tested against these previously defined input parameters. A continuous learning process,where the input parameters continuously contribute to the refinement of the AI ​​model is not required, but can be provided optionally. A significant advantage of this method is that it is not necessary to know in advance what the good input parameters are. These are initially determined and assumed to be good, and only subsequently defined or confirmed as good, for example, through user input. Thus, the method according to the invention can be used directly in production without a complex training process, even in the event of changes to the tamping process or brush geometry, or other changes. If a set of input parameters is rated as bad by the AI ​​classifier during operation, it can be provided that a user can determine whether the input parameters are good or bad. This means that a manual check of the result can be carried out. If the input parameters are nevertheless determined to be good,These input parameters can be additionally included in the training data, so that such input parameters are more likely to be rated as good in the future. PC 240230 J 7 / 32 March 25, 2024 In one embodiment, a sensor, for example a force sensor or a strain gauge, is used to determine the tamping force. In this way, the tamping force can be determined with little technical effort and cost-effectively. The tamping force can also be determined in other ways, for example from the power consumption of the tamping tool's drive motor or its torque. In one embodiment, it can be advantageous if the AI ​​classifier has at least one further input parameter derived from another sensor. In this way, the AI ​​classifier can access additional, ideally independent, information.so that better classification is possible. Such a second sensor can, for example, be a motion sensor for detecting vibrations or shocks. The input parameters for the second sensor can be determined in the same way as for the tamping force. This means that features can be extracted from time series. In a further embodiment, a second AI classifier can be used to additionally determine a type of fault condition. To train the second AI classifier, input parameters from time series of tamping forces, each of which is assigned to a specific fault condition, are used. In this way, it is not only possible to determine that a fault condition has occurred during operation, but it is also possible to determine what type of fault condition it is. To do this, however, it is necessary to train the AI ​​classifier with input parameters.that represent a fault condition. This fault condition must also be classified according to its type. In contrast to the training of the first AI classifier, training here is therefore carried out exclusively with "bad" data. The training of the second AI classifier is therefore preferably carried out in advance with controlled input parameters, for example, by deliberately provoking certain fault patterns. Especially when using a second AI classifier to detect fault types, additional input parameters derived from measured values ​​from a second sensor can be advantageous. In the best case, the second sensor provides measured values ​​that are independent of the measured values ​​of the first sensor, i.e., have no correlation. In a further embodiment,that after a fault condition occurs, the brush tamping machine is repaired. An operator determines the type of fault and a procedure for resolving the fault condition. After the repair, this information on the fault type and its resolution is stored, for example, in a database. If a fault condition occurs again, a hint for resolving the fault condition can be retrieved from this database by assigning the fault type. The database can, for example, be cloud-based. Alternatively or additionally, this information on the fault type and its resolution can be added to the training data of the AI ​​classifier, especially the second one. This has the advantage thatthat if an error condition occurs again, in addition to the error type, an indication for correcting the error condition can be output directly by the AI ​​classifier. Various models can be used as the AI ​​classifier, such as artificial neural networks or other methods. PC 240230 J 9 / 32 March 25, 2024 The AI ​​classifier is preferably a so-called one-class support vector machine, an isolation forest, an autoencoder, or a convolutional autoencoder. The invention also relates to a brush stuffing machine, in particular a toothbrush stuffing machine, with means by which the brush stuffing machine is configured to carry out the method according to the invention described above. As means for carrying out the method, the brush stuffing machine can have a pusher tongue for stuffing bristle bundles, a sensor, in particular a force sensor, for determining a stuffing force,which acts on the pusher tongue when stuffing bristle bundles, have a control unit and at least one AI classifier. The AI ​​classifier can be at least temporarily part of the brush stuffing machine. The control unit can have an interface for, preferably bidirectional, communication with the aforementioned database. The invention is described in more detail below using an exemplary embodiment, but is not limited to this exemplary embodiment. Further exemplary embodiments arise by combining the features of individual or multiple claims with one another and / or by combining individual or multiple features of the exemplary embodiment. Figure 1 shows a perspective view of a brush stuffing machine with a stuffing tool, wherein a pusher tongue of the stuffing tool is connected to a drive of the brush stuffing machine via a two-part tongue lever,wherein a force sensor is arranged between a drive lever and a force transmission lever PC 240230 J 10 / 32 25 March 2024 of the two-part tongue lever, against which the force transmission lever is preloaded in the tamping direction with a defined preload force, Figure 2 shows a diagram with a characteristic curve of the force sensor of the brush tamping machine shown in Figure 1, wherein the diagram shows a value of a sensor signal of the force sensor plotted against the force acting on the force sensor, Figure 3 shows a time series of the tamping force over four tamping processes, Figure 4 shows a schematic, side sectional view of a brush body in the area of ​​a tamping hole with a correctly positioned armature of the correct size, Figure 5 shows a schematic top view of a brush body in the area of ​​a tamping hole with a correctly positioned armature of the correct size, Figure 6 shows a superimposed representation of a multivariate time series of the tamping force consisting of several time series,which each reflect a fault-free state, Figure 7 shows a flow diagram of a method according to the invention for training an AI classifier, Figure 8 shows a flow diagram of a method according to the invention for monitoring the stuffing process, Figure 9 shows a schematic, lateral sectional view of a brush body in the area of ​​a stuffing hole with an anchor of the correct size that is positioned too high in the stuffing hole, Figure 10 shows a superimposed view of a multivariate time series of the stuffing force consisting of several time series, each of which reflects the faulty state of Figure 9, Figure 11 shows a schematic, lateral sectional view of a brush body in the area of ​​a stuffing hole with a PC 240230 J 11 / 32 March 25, 2024 anchor of the correct size that is positioned too low in the stuffing hole, Figure 12 shows a superimposed view of a multivariate time series of the stuffing force consisting of several time series,which each reflect the faulty state of Figure 11. Figure 13 is a schematic plan view of a brush body in the area of ​​a plug hole with a correctly sized armature positioned too high on the plug hole. Figure 14 is a schematic plan view of a brush body in the area of ​​a plug hole with a correctly sized armature positioned too far down on the plug hole. Figure 15 is a superimposed representation of a multivariate time series of the plugging force consisting of several time series, each reflecting one of the faulty states of Figures 13 and 14. Figure 16 is a schematic plan view of a brush body in the area of ​​a plug hole with a correctly sized armature positioned too far to the left on the plug hole. Figure 17 is a schematic plan view of a brush body in the area of ​​a plug hole with a correctly sized armature positioned too far to the right on the plug hole.Figure 18 is a superimposed representation of a multivariate time series of the stuffing force consisting of several time series, each reflecting one of the faulty states of Figures 16 and 17. Figure 19 is a schematic, top view of a brush body in the area of ​​a stuffing hole with an anchor that is too short and in the correct position. Figure 20 is a superimposed representation of a multivariate time series of the stuffing force consisting of several time series, each reflecting the faulty state of Figure 19. Figure 21 is a schematic, top view of a brush body in the area of ​​a stuffing hole with an anchor that is too long and in the correct position. Figure 22 is a superimposed representation of a multivariate time series of the stuffing force consisting of several time series, each reflecting the faulty state caused by too many filaments.Figure 23 shows a superimposed representation of a multivariate time series of the stuffing force consisting of several time series, each reflecting the faulty condition caused by too few filaments, and Figure 24 shows a superimposed representation of a multivariate time series of the stuffing force consisting of several time series, each reflecting the faulty condition caused by an impact of the stuffing tool. Figure 1 shows a brush stuffing machine, designated as a whole by 1, which is configured to carry out the method according to the invention. However, this method according to the invention is not limited to the brush stuffing machine shown as an example. Rather, the method can be easily transferred to other machines and other sensor arrangements. The brush stuffing machine 1 has a stuffing tool 2, which comprises a pusher tongue 3. The pusher tongue 3 is configured toBristle bundles 4, together with fastening anchors, are to be stuffed into bundle receiving holes 5 of a brush body 6. The brush stuffing machine 1 has a drive 7, which is connected to the pusher tongue 3 via a power transmission lever 8 and is configured to drive the pusher tongue 3 in a stuffing movement. Figure 1 shows that the power transmission lever 8 is preloaded against a sensor, namely a force sensor 10, in a stuffing direction of the pusher tongue 3 indicated by the arrow 9. The direction of a preload force acting on the force sensor 10 is illustrated by the arrow 24 in Figure 1. The force sensor 10, preloaded with the preload force 24, is configured to measure a force resulting from the preload force 24 and a stuffing force 25.which acts on the pusher tongue 3 when tufts of bristle 4 are tufted. The force sensor 10 is arranged in the brush tufting machine 1 shown in Figure 1 in such a way that it is relieved by the tufting force 25 acting on the force transmission lever 8 via the pusher tongue 3, opposite to the tufting direction 9. As the tufting force increases, the resulting force measurable with the force sensor 10 therefore becomes smaller. The relief of the force sensor 10 with increasing tufting force is made possible by a correspondingly movable mounting of the force transmission lever 8. The force transmission lever 8 is movably mounted and initially connected to the force sensor 10 in such a way that the force sensor 10 is relieved by the tufting force 25 acting on the force transmission lever 8 via the pusher tongue 3, opposite to the tufting direction 9 of the pusher tongue 3. If a stuffing force 24 acting against the stuffing direction 9 of the pusher tongue 3 is applied in the amount of an overload force which is so large,that it removes the preload of the force transmission lever 8 against the force sensor 10, a force transmission contact between the force transmission lever 8 PC 240230 J 14 / 32 March 25, 2024 and the force sensor 10 is removed. The relief of the load on the force sensor 10 that occurs when the push rod 3 is loaded is detected by the force sensor as a change in the resulting force and leads to the output of a corresponding sensor signal. The sensor signal represents the resulting force currently applied to the force sensor 10 and allows a conclusion to be drawn about the specifically applied stuffing force 25 and thus about the load to which the push rod 3 is exposed. Since the force sensor 10 is relieved by a load on the push rod 3 opposite to the stuffing direction 9, there is no risk of damage to the force sensor 10 if the push rod 3 is overloaded in this direction. The force sensor 10 is thus effectively protected from damage,without, however, impairing the monitoring of the pusher tongue 3. The brush tamping machine 1 has a two-part tongue lever 11, wherein one part of the tongue lever 11 is a drive lever 12 and a second part of the tongue lever 11 is the previously mentioned power transmission lever 8. The drive lever 12 is connected to the drive 7 of the brush tamping machine 1 via a connecting rod 13. The end of the drive lever 12 facing away from the connecting rod 13 is connected to the power transmission lever 8. The drive lever 12 and the power transmission lever 8 are articulated to one another in such a way that the force sensor 10 is relieved by the tamping force 25 acting on the power transmission lever 8 via the pusher tongue 3, opposite to the tamping direction 9. By means of the tamping force 25, the power transmission lever 8 can change its angular position to the PC 240230 J 15 / 32 25 March 2024 drive lever 12,by pivoting it slightly about a pivot axis 30 of the articulated connection between the drive lever 12 and the power transmission lever 8 relative to the drive lever 12. The articulated connection between the drive lever 12 and the power transmission lever 8 is formed by a pivot joint 14. The force sensor 10 is arranged on the drive lever 12 and between the drive lever 12 and the power transmission lever 8. By changing the angular position between the power transmission lever 8 and the drive lever 12, the force sensor 10 and the drive lever 12 can move away from the power transmission lever 8 with increasing tamping force 25 and can be released from it when the tamping force 25 is so great,that it cancels the pretensioning force 24. The pretensioning force 24 is generated by a pretensioning device 15 of the brush tamping machine 1. For this purpose, the force transmission lever 8 is pretensioned in the tamping direction 9 against the force sensor 10 by the pretensioning device 15. The pretensioning device 15 is arranged on the drive lever 12 and comprises a compression spring designated 16, which tensions the force transmission lever 8 against the force sensor 10 arranged on the drive lever 12. The brush tamping machine 1 has a control unit 17. The control unit 17 configures the brush tamping machine 1 to carry out the method according to the invention. The control unit 17 is configured to reduce the tamping speed of the pusher tongue 3 or to stop the pusher tongue 3.as soon as the resulting force measured by the force sensor 10 reaches or falls below a limit value. For this purpose, the control unit 17 can control the drive 7 of the PC 240230 J 16 / 32 March 25, 2024 brush tamping machine 1 depending on a sensor signal from the force sensor 10, for example, reduce the speed of the drive 7 or even stop the drive 7 as soon as the resulting force measured by the force sensor 10 reaches or falls below a limit value. The limit value of the resulting force, which must be reached or fallen below to reduce the tamping speed of the pusher tongue 3 and / or to reduce the speed of the drive 7, can be greater than the limit value,which must be reached or exceeded to stop the pusher tongue 3 and / or the drive 7. The brush manufacturing machine 1 has a bristle magazine 26 with bristle filaments arranged therein and a bundle separator 27 with a bundle removal notch 28. With the bundle removal notch 28, the bundle separator 27 is moved past the bristle magazine 26,to remove a bristle bundle 4 from the bristle magazine 26. The bundle separator 27 then transfers the removed bristle bundle 4 to the downstream stuffing tool 2. With the stuffing tool 2, the bristle bundle 4 is then stuffed into a bundle receiving hole 5 of a brush body 6 arranged on a holding device 29 of the brush stuffing machine 1. The previously explained brush stuffing machine 1 is designed for stuffing brushes. In this process, bristle bundles 4 are stuffed into bundle receiving holes 5 of a brush body 6 held ready by the pusher tongue 3 of the stuffing tool 2 of the brush stuffing machine 1. To determine a stuffing force 25 acting on the pusher tongue 3,a force PC 240230 J 17 / 32 March 25, 2024 resulting from the pretension force 24 and the tamping force 25 is measured with the force sensor 10, and the tamping force 25 is derived from this. The diagram in Figure 2 shows the characteristic curve of the force sensor 10 of the brush tamping machine 1. In the diagram, the sensor signal 18 output by the force sensor 10 is plotted against the force 19 acting on the force sensor 10. The diagram shows a measuring range start 20, a measuring range end 21, and an overload limit 22 of the force sensor 10. If the force sensor 10 is loaded beyond its overload limit 22, it can be damaged. Reliable measurements are only possible between the measuring range start 20 and the measuring range end 21. The method according to the invention is based on this arrangement and also uses the existing force sensor to monitor the tamping process according to the invention. As already mentioned above,The method according to the invention is not limited to the arrangement shown in Figure 1. Rather, the stuffing force can also be determined by a different mechanical sensor arrangement and / or by a different sensor, such as a strain gauge or by derivation from the power of the drive 7, which then functions as a force sensor. Completely different mechanical configurations are also possible. Figure 3 shows, by way of example, a time series with recorded measured values ​​of the stuffing force 25. The diagram in Figure 3 essentially comprises four stuffing processes. The beginning of a stuffing process begins with the force 23 set as the preload, as shown in Figure 2. Starting from this force, a stuffing process comprises one revolution of the drive 7. Figures 4 and 5 schematically show different views of a brush body 6 in the area of ​​a stuffing hole or bundle receiving hole 5.in which an anchor 31 is inserted in its normal or fault-free position. The filament bundles are not shown here for the sake of clarity. The curve of the stuffing force 25 in Figure 3 essentially shows the fault-free state, which includes an anchor position as shown in Figures 5 and 6. In the fault-free state, the stuffing force 25 has a characteristic curve in a time series that is essentially the same or at least similar for each bundle receiving hole 5. The invention has now recognized that deviations from this characteristic curve can be an indication of the presence of a faulty state. The finding of the invention is based on the idea of ​​detecting faults in the stuffing process based on deviations of the stuffing force from the characteristic curve. To detect a faulty state, an AI classifier 34 is used according to the invention, which is or has been essentially trained with time series.which result exclusively from error-free measurements. Figure 6 shows several such time series superimposed as an example. This clearly shows that the time series do exhibit local deviations, but essentially have a similar course. Such a set of time series can be used, for example, as a training data set for an AI classifier. During operation, a recorded time series is transferred to an AI classifier 34 trained in this way. The AI ​​classifier 34 then determines whether the transferred time series can be classified as good, taking into account the training data. The output of the AI ​​classifier can then be a binary value, i.e., good or bad, or a confidence value or a comparable value that allows an assignment to good or bad. In the example, a time series comprises, for example, 800 - 1000 measured values.The sampling rate is approximately 16 kHz to 20 kHz. However, this is only an example for this brush tamping machine. Other brush tamping machines can be operated with different tamping frequencies, which also changes the sampling rate and / or the number of measured values ​​per time series. This number of measured values ​​is quite high for use as input parameters. In practice, it can therefore be advantageous if the time series are not used directly as input parameters for the AI ​​classifier. For this reason, the time series can be preprocessed before use.for example, through feature extraction or data reduction. For example, for feature extraction, the time series can be transformed from time space to frequency space using a discrete Fourier transformation. The determined Fourier coefficients, or a portion of them, can then be used as input parameters for the AI ​​classifier. In this way, the temporal variance is removed from the measured values ​​during the stuffing process, so that features that are contained in the time series with a specific frequency can be considered more clearly and independently of time. In addition, for example, only a certain number,approximately the first 80 Fourier coefficients are used. In this way, the original signal can be reconstructed with sufficient accuracy while simultaneously significantly reducing the number of input parameters. Another option for feature extraction is the determination of wavelet coefficients. A discrete wavelet transform is suitable for this purpose. Similar to the discrete Fourier transform, several wavelet coefficients are determined that enable reconstruction of the output signal. A wavelet transform converts the signal from time-space to time-frequency space. These wavelet coefficients can be used alternatively or additionally as input parameters of the AI ​​classifier. Here, too, only a certain number of wavelet coefficients can be used to reduce the total number of input parameters. There are various wavelet families that can be used.for example, Haar, Daubechies, Coiflet 5, Symlet 6, biorthogonal, and reverse biorthogonal. Another possibility for feature extraction is the application of a distance measure. For example, a Euclidean distance or an elastic distance, such as dynamic time warping, can be used. For example, distances can be calculated locally by dividing the time series into intervals and then calculating the distance to a specific reference for each interval. The number of input parameters is determined by the number of intervals. Alternatively, it is also possible to reduce the time series data directly without extracting features. This can be done, for example, bythat only every second or every fourth measured value is taken into account. The unconsidered measured values ​​can be ignored or taken into account by averaging. There are also other options for feature extraction and / or data reduction that are not listed here, but which the person skilled in the art will readily consider. In principle, it is also possible to combine all of the above-mentioned or other methods for feature extraction and / or data reduction with each other, should this be necessary. The characteristic curve of the stuffing force depends on various parameters, for example, the number of filaments per bristle bundle,the diameter of the bundle receiving holes or the anchor dimensions. These parameters can, for example, be used as presets for each brush type. An AI classifier 34 would then have to be trained with good values ​​for each such preset. This would then have to be selected before commissioning the brush tamping machine. However, the invention has the advantage that training with the good values ​​is possible during operation, so to speak "on the fly." Figure 7 shows an example of a flow chart of such a training process 70. A brush tamping machine, such as that shown in Figure 1, is switched on in step 71, using the presets for a specific brush type. In a subsequent step 72, a time series of the tamping force is determined for a predetermined number of bundle receiving holes, and in a further step 73, input parameters for the AI ​​classifier 34 are determined from this.for example, as described above through feature extraction. PC 240230 J 22 / 32 March 25, 2024 These input parameters are passed in the next step 74 as training data to a previously untrained, so to speak "empty" AI classifier 34. It is initially assumed that all training data correspond to the state "Good." A check is then performed 75 to determine whether a predetermined number of training data sets has been determined. If the predetermined number has not been reached, the process is repeated in step 72. If sufficient training data is available, a user determines or confirms in a verification step 76 that the brushes produced up to that point were OK and thus all recorded training data can be classified as "Good." If errors have occurred, all training data is discarded in a further step 78, and training begins again with step 72. However, if a "Good" classification has been made,the training is completed 77. Immediately afterwards, the thus trained AI classifier 34 can be used in operation, for example according to the flow chart in Figure 8. It is also possible for the operation of the brush tamping machine to continue uninterrupted after training, i.e., even during the user confirmation in step 76. In operation according to the method 80, in a first step 81, analogous to step 72, a time series of the tamping force is determined for a predetermined number of bundle receiving holes, and in a further step 82, input parameters for the AI ​​classifier 34 are determined from this, for example, as described above by feature extraction. These input parameters are transferred to the previously trained AI classifier 34 in the next step 83. The AI ​​classifier 34 classifies the submitted input parameters based on the training data classified as good and thus determines,whether these are also good. In the subsequent test step 84, the result of the AI ​​classifier 34 is evaluated. If no error condition exists, i.e., the input parameters are classified as good, the process continues with step 81. If the result is classified as bad, an error condition may exist. In the example, in this error case, a further test step 85 is provided for a manual inspection of the previously produced brushes. A user can manually determine whether an error actually exists or whether the brushes are OK and therefore no error exists. If the input parameters are confirmed as good, the process can continue with step 81. Optionally, in a step 87, these input parameters can be added to the training data of the AI ​​classifier 34. If the error is confirmed,error correction 86 can take place. This can, for example, include detecting the type of error using a second AI classifier. This second AI classifier can have been trained in advance with input parameters that are assigned to typical error types. Figures 9 to 23 show, by way of example, various PC 240230 J 24 / 32 March 25, 2024 error types and associated, characteristic time series. A second AI classifier 34 can, for example, be trained with features that were extracted from these time series as described above. Based on the type of deviation in the characteristic time series, the second AI classifier 34 can assign input parameters to a type of error. Figure 9 shows the view of the brush body 6 in Figure 4, but the armature 31 is seated too high in the bundle receiving hole 5. This deviation leads to time series of the stuffing force 25, which are shown in Figure 10. Figure 11 shows the view of the brush body 6 of Figure 4,however, the armature 31 is seated too deep in the bundle receiving hole 5. This deviation leads to time series of the stuffing force 25, which are shown in Figure 12. Figures 13 and 14 each show a view of the brush body 6 of Figure 5, wherein the armature 31 is seated too far up or too far down in the bundle receiving hole 5. These deviations are symmetrical and lead to time series of the stuffing force 25, which are shown in Figure 15. Figures 16 and 17 each show a view of the brush body 6 of Figure 5, wherein the armature 31 is seated too far to the left or too far to the right in the bundle receiving hole 5. These deviations are symmetrical and lead to time series of the stuffing force 25, which are shown in Figure 18. Figure 19 shows the view of the brush body 6 of Figure 5, wherein the armature 31 is too short in relation to the diameter of the bundle receiving hole 5, but is seated in the correct position. This deviation leads to time series of the tamping force 25,PC 240230 J 25 / 32 March 25, 2024 shown in Figure 20. Another possible error is shown in Figure 21, which corresponds to the view of the brush body 6 in Figure 5, wherein the armature 31 is too long in relation to the diameter of the bundle receiving hole 5, but is in the correct position. However, no time series is shown for this. Other possible error types are that there are too many filaments in the filament bundle, which leads to time series of the stuffing force 25 according to Figure 22. If there are too few filaments in the filament bundle, time series of the stuffing force 25 can result according to Figure 23. Figure 24 shows time series of the stuffing force 25 that can arise when the stuffing tool is impacted. By training a second AI classifier 34 with such time series, the AI ​​classifier 34 can differentiate between and recognize such error types. It can be helpful to use independent input parameters as described above,for example, a second sensor, are additionally present. In a further development of the invention, the AI ​​classifier 34 can be trained for a fault type with additional repair instructions for correcting a specific fault, so that in addition to a fault type, instructions for correcting such a fault can be output by the AI ​​classifier 34. Alternatively or additionally, the repair instructions can also be stored in a, possibly cloud-based, database 33. The brush stuffing machine 1 is configured to carry out the method previously explained in PC 240230 J 26 / 32 March 25, 2024 and has corresponding means for this purpose. The brush stuffing machine uses the pusher tongue 3 for stuffing bristle bundles 4, the force sensor 10 for determining a stuffing force 25 that acts on the pusher tongue 3 when stuffing bristle bundles 4,the previously mentioned control unit 17 and the AI ​​classifier 34. / List of reference symbols,

[0002] PC 240230 J 27 / 32 March 25, 2024 List of reference symbols 1 Brush tucking machine 2 Tucking tool 3 Pusher tongue 4 Bristle bundle 5 Bundle receiving holes 6 Brush body 7 Drive 8 Power transmission lever 9 Tucking direction 10 Force sensor 11 Tongue lever 12 Drive lever 13 Connecting rod 14 Swivel joint between 8 and 12 15 Pretensioning device 16 Compression spring 17 Control unit 18 Sensor signal 19 Force acting on 10 20 Start of measuring range 21 End of measuring range 22 Overload limit 23 Pretension setting point 24 Pretension force 25 Tucking force 26 Bristle magazine 27 Bundle separator 28 Bundle removal notch 29 Holding device 30 Swivel axis 31 Armature 32 Time series 33 (Cloud) database PC 240230 J 28 / 32 March 25, 2024 34 AI Classifier / Claims

Claims

PC 240230 J 29 / 32 March 25, 2024 Claims 1. Method for monitoring the stuffing process of a brush stuffing machine (1), in particular a toothbrush stuffing machine, using an AI-based classifier, characterized in that during the stuffing process, a time series (32) of a stuffing force (25) acting on a pusher tongue (3) during the stuffing of bristle bundles (4) is determined, that the AI ​​classifier (34) uses a plurality of input parameters determined from this time series (25) and indicates as an output whether an error condition exists, wherein the AI ​​classifier (34) was trained with input parameters that represent an error-free stuffing process.

2. Method according to claim 1, characterized in that for each stuffing process, a time series (32), in particular a univariate one, is determined from the stuffing force (25). 3.Method according to one of the preceding claims, characterized in that a multivariate time series is formed from a predetermined number of time series.

4. Method according to one of the preceding claims, characterized in that the input parameters are determined by feature extraction and / or data reduction from a time series (32).

5. Method according to claim 4, characterized in that a predetermined number of Fourier coefficients and / or a predetermined number of wavelet coefficients and / or a distance to the median for a predetermined number of intervals of the time series (32) are determined as features during the feature extraction, in particular wherein the predetermined number is 80 in each case. PC 240230 J 30 / 32 March 25, 2024.

6. The method according to claim 4 or 5, characterized in that for data reduction, the number of measured values ​​in a time series (32) is reduced, in particular so that 200 measured values ​​remain per time series.

7. The method according to one of the preceding claims, characterized in that for training the AI ​​classifier (34) during operation, a predetermined number of time series (32) is determined and the input parameters determined therefrom are defined as good, for example by a user.

8. The method according to one of the preceding claims, characterized in that for input parameters rated as bad, a user can determine whether the input parameters are good or bad, in particular wherein, if rated as good, the input parameters are added to the training data of the AI ​​classifier (34). 9.Method according to one of the preceding claims, characterized in that a sensor, in particular a force sensor (10), is used to determine the stuffing force (25).

10. Method according to one of the preceding claims, characterized in that the AI ​​classifier (34) has at least one further input parameter which is derived from a further sensor.

11. Method according to one of the preceding claims, characterized in that a second AI classifier (34) is used to determine a type of error state, wherein input parameters from time series of stuffing forces, each of which is assigned to a specific error state, are used to train the second AI classifier (34). PC 240230 J 31 / 32 March 25, 2024.

12. Method according to one of the preceding claims, characterized in that after the occurrence of a fault condition, a repair of the brush tamping machine is carried out, for example by an operator; that after the repair, information on the type of fault and its rectification is stored in a, in particular cloud-based, database (33) and / or added to the training data of the, in particular second, AI classifier (34); and if a fault condition occurs again, in addition to the type of fault, an indication for rectifying the fault condition is output by the AI ​​classifier (34).

13. Method according to one of the preceding claims, characterized in that the AI ​​classifier (34) and / or a second AI classifier is a so-called one-class support vector machine, an isolation forest, an autoencoder, or a convolutional autoencoder. 14.Brush stuffing machine (1), in particular a toothbrush stuffing machine, with means by which the brush stuffing machine (1) is configured to carry out the method according to one of claims 1 to 13, in particular wherein the brush stuffing machine (1) has, as means for carrying out the method, at least one pusher tongue (3) for stuffing bristle bundles (4), a sensor, in particular a force sensor (10), for determining a stuffing force (25) acting on the pusher tongue (3) when stuffing bristle bundles (4), a control unit (17), and at least one AI classifier (34). / Summary.

Images