Method for monitoring filling process of brush filling machine
By using an artificial intelligence classifier in the brush stuffing machine to analyze the stuffing force time series, combined with feature extraction and data reduction techniques, the problem of defect identification in the brush stuffing process was solved, and low-cost and efficient monitoring and prevention were achieved.
Patent Information
- Application Number
- CN202480010457.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2023-03-28
- Filing Date
- 2024-03-26
- Publication Date
- 2025-09-12
AI Technical Summary
The existing technology makes it difficult to effectively monitor the stuffing process of the brush stuffing machine, resulting in the possibility of defects in the produced brushes, and visual inspection is time-consuming and cost-intensive.
By determining the time series of the filling force during the filling process, using artificial intelligence classifiers to analyze input parameters and identify defect states, combined with feature extraction and data reduction techniques, the computational complexity and resource requirements are reduced.
It achieves low-cost and efficient brush stuffing process monitoring, can identify and prevent the production of defective brushes, and reduces training time and resource consumption.
Smart Images
Figure CN120641008A_ABST
Abstract
Description
Technical Field
[0001] Method for monitoring the stuffing process of a brush stuffing machine, in particular a toothbrush stuffing machine, using an artificial intelligence-based classifier. Background Art
[0002] Brush stuffing machines are used for producing brushes and, in particular, for stuffing bristle bundles into bundle receiving openings of a brush body and thereby producing a bristle stuffing for the brush.
[0003] For this purpose, brush stuffing machine has stuffing tool, and this stuffing tool has pushing tongue, and this pushing tongue sets up and is arranged to be used for stuffing bristle bundle in the bundle receiving hole of brush body.Usually bristle bundle is stuffed in the bundle receiving hole together with fixed anchor and is anchored there.
[0004] For a flawless filling result, different parameters must be observed so that a flawless brush is produced. These parameters include, for example, the length and orientation of the push tongues, the number of filaments, and the length of the anchors.
[0005] If one of these parameters deviates from its normal value, the produced brush may be defective or poorly configured. Summary of the Invention
[0006] The object of the present invention is to create a method with which the stuffing process can be monitored and thus defective and / or poorly made brushes and / or defects of the brush stuffing machine can be detected.
[0007] This object is achieved by a method having the features of claim 1 .
[0008] Accordingly, the method according to the invention is characterized in that a time series of the filling force acting on the push tongue when filling the bristle bundle is determined during the filling process, an artificial intelligence classifier uses multiple input parameters determined from the time series and outputs whether a defective state exists, wherein the artificial intelligence classifier has been trained in advance using input parameters representing a defect-free filling process.
[0009] Typically, visual inspections are required to monitor brush quality, which is very complex and cost-intensive. The advantage of the present invention is that, with the help of an artificial intelligence classifier, the condition of the produced brushes can be inferred from the measured values of a simple force sensor. This allows for simple and cost-effective monitoring of the stuffing process. In particular, such a force sensor may already be present on the brush stuffing machine, allowing its measured values to be used directly.
[0010] The artificial intelligence classifier is trained only with input parameters that represent a non-defective state, or "good." This makes it easy to identify deviations from the training data as a defective state, or "bad." This significantly shortens training time compared to training with bad values.
[0011] To determine the time series, measured values of the packing force are determined within fixed time intervals. The number of measured values is determined by the sampling rate. The sampling rate is preferably in the double digit kilohertz range.
[0012] In one embodiment, for each stuffing process, a time series, in particular a univariate time series, is determined from the stuffing force. This means that a time series corresponds exactly to a stuffing process at a stuffing hole and thus reflects a cycle of the stuffing machine.
[0013] Although this number of measured values can be used as input parameters for an artificial intelligence classifier, the complexity of the artificial intelligence classifier may increase due to the large number of input parameters. Therefore, the calculation may require a lot of memory and / or computing time.
[0014] Therefore, in one embodiment, the input parameters are determined from such a time series by feature extraction and / or data reduction. In this way, the number of input parameters is reduced, thereby reducing the complexity of the artificial intelligence classifier and the calculation can be performed faster and with fewer resources.
[0015] In one embodiment, a predetermined number of Fourier coefficients are determined as features during feature extraction. A Fourier transform, or a fast Fourier transform, i.e., a discrete Fourier transform, is applied to the time series. The input parameters are limited to a predetermined number of Fourier coefficients, for example, the first 80 Fourier coefficients.
[0016] Alternatively or additionally, a predetermined number of wavelet coefficients can be determined as features during feature extraction, for example, by discrete wavelet transformation. Advantageously, Haar wavelets or Daubechies wavelets are used as wavelets, in particular wavelets with N values between 4 and 20 or Coiflet wavelets with N values between 2 and 4. For all wavelets, for example, levels between 1 and 4 are used. The value N and the levels are preferably selected such that the number of wavelet coefficients is approximately 80.
[0017] In another alternative embodiment or in addition, as a feature, during feature extraction, a distance from the median is determined for each of a predetermined number of intervals of the time series. Here, the predetermined number can also be approximately 80.
[0018] Through feature extraction, the number of input parameters can be significantly reduced without causing relevant information loss.
[0019] Alternatively or additionally, in one embodiment, to reduce data, the number of measured values in the time series is reduced, in particular, to retain 200 measured values per time series. In this embodiment, subsampling is performed, thereby reducing the number of measured values in the time series. This can be accomplished, for example, by simply removing measured values from the time series. Alternatively, an average value can be calculated from multiple adjacent measured values and used as the input parameter. This allows the measured values to be used directly without requiring calculations for feature extraction. This results in faster processing and reduced computational effort.
[0020] In one embodiment, a multivariate time series is formed alternatively and / or additionally from a predetermined number of time series. As already described above, a time series corresponds to a filling process at a filling hole. Now, a number of such time series are combined to form a multivariate time series. This means, for example, that the time series of 100 filling holes are combined. The number of filling holes per brush is not important. If a brush has, for example, 40 filling holes, the multivariate time series includes 2.5 brushes. Of course, these numbers are only exemplary and can be easily adapted to the application.
[0021] Such a multivariate time series can be the basis for determining the input parameters, as described above. This means extracting features from such a multivariate time series.
[0022] In an alternative embodiment, feature extraction may be performed first, and a multivariate time series may be formed from these features.
[0023] Independently of this, a multivariate time series is used as an input parameter. In this way, the artificial intelligence classifier does not have to be calculated separately for each hole. Instead, the measured values of a plurality of filled holes are first collected and then considered as a whole unit.
[0024] Although it is not possible to make a defect statement for each filling hole or brush in this way, in practice, this is not mandatory for determining deviations from normal operation, i.e. defective states, during the operation of the brush filling machine.
[0025] Artificial intelligence or artificial intelligence classifiers are typically trained before they are used.
[0026] In one embodiment, in order to train the artificial intelligence classifier during operation, a predetermined number of time series are determined and the input parameters determined therefrom are defined as good, for example by the user. In this way, the training of the artificial intelligence classifier can be carried out during operation, so to speak "on the fly". For this purpose, for example, at the start of brush production, a certain number of time series defined or confirmed as good by the user are recorded. For example, by viewing the brushes. The input parameters determined from these time series are used as training data for the artificial intelligence classifier. In subsequent runs, the input parameters are tested against these previously defined input parameters. A continuous learning process, in which the input parameters continuously contribute to the refinement of the artificial intelligence model, is not required but can be set as optional.
[0027] A significant advantage of this method is that it is not necessary to know in advance what good input parameters look like. These good input parameters are first determined and considered good, and only then defined or confirmed as good, for example, by user input. Therefore, even if changes occur in the filling process, the brush geometry, or other variations, the method according to the present invention can be put into production directly without the need for a complex training process.
[0028] If an AI classifier classifies a set of input parameters as poor during operation, it can be provided that the user can determine whether the input parameters are good or poor. This allows for manual review of the results. If the input parameters are still determined to be good, they can be included in the training data, making it more likely that such input parameters will be rated as good in the future.
[0029] In one embodiment, a sensor, such as a force sensor or a strain gauge, is used to determine the packing force. This allows the packing force to be determined cost-effectively and with minimal technical effort. The packing force can also be determined in other ways, such as from the power consumption or torque of the packing tool's drive motor.
[0030] In one embodiment, it may be advantageous for the artificial intelligence classifier to have at least one additional input parameter derived from another sensor. This allows the artificial intelligence classifier to utilize additional, preferably independent, information, thereby achieving better classification. Such a second sensor could, for example, be a motion sensor for detecting vibrations or oscillations. The input parameters for this second sensor can be determined in the same manner as for the filling force. This means, for example, that features can be extracted from a time series.
[0031] In another embodiment, a second artificial intelligence classifier can be used to additionally identify a fault condition. Input parameters from time series of the filling forces associated with each identified fault condition are used to train the second artificial intelligence classifier. This allows not only the occurrence of a fault condition during operation to be determined, but also the type of fault condition involved.
[0032] However, to do this, the artificial intelligence classifier must be trained using input parameters that represent defect states. Furthermore, these defect states must be classified according to their type. Unlike the training of the first artificial intelligence classifier, training here is therefore performed exclusively using "bad" data. Therefore, training the second artificial intelligence classifier is preferably performed in advance using controlled input parameters, for example by deliberately generating specific defect images.
[0033] It is when using the second artificial intelligence classifier for identifying defect types that additional input parameters derived from the measurement values of the second sensor can be advantageous. In the best case, the measurement values provided by the second sensor are independent of the measurement values of the first sensor, ie, have no correlation.
[0034] Another embodiment provides for the brush stuffing machine to be serviced after a defect occurs. Here, an operator determines the defect type and the method for correcting the defect. After the repair, this information about the defect type and its correction is stored, for example, in a database. If the defect reoccurs, instructions for correcting the defect can be retrieved from the database by assigning the defect type. This database can be cloud-based, for example.
[0035] Alternatively or additionally, this information about the defect type and its elimination can be added to the training data of an artificial intelligence classifier, in particular a second artificial intelligence classifier. This has the advantage that, when a defect state reappears, the artificial intelligence classifier can directly output instructions for eliminating the defect state in addition to the defect type.
[0036] As artificial intelligence classifiers, different models can be used, such as artificial neural networks or other methods.
[0037] Preferably, the artificial intelligence classifier is a so-called one-class support vector machine, an isolation forest, an autoencoder or a convolutional autoencoder.
[0038] The present invention also relates to a brush stuffer, in particular a toothbrush stuffer, which is equipped with a device for the brush stuffer to implement the method according to the invention. As a device for implementing the method, the brush stuffer can include a push tongue for stuffing the bristle bundle, a sensor for determining the stuffing force acting on the push tongue when stuffing the bristle bundle, in particular a force sensor, a control unit and at least one artificial intelligence classifier. The artificial intelligence classifier can at least sometimes be part of the brush stuffer. The control unit can have an interface for communication with the aforementioned database, preferably for bidirectional communication. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] The present invention will be described in detail below based on an embodiment, but the present invention is not limited to this embodiment. Further embodiments can be obtained by combining the features of one or more protective claims with each other and / or by combining one or more features of the embodiments.
[0040] In the picture:
[0041] Figure 1 A perspective view of a brush stuffer with a stuffing tool is shown, wherein a push tongue of the stuffing tool is connected to a drive of the brush stuffer via a two-part tongue lever, a force sensor being arranged between a force transmission lever and a drive lever of the two-part tongue lever, the force transmission lever being preloaded against the force sensor with a defined preload force in the stuffing direction,
[0042] Figure 2 Shown in Figure 1 , a diagram showing a characteristic curve of a force sensor of a brush stuffing machine, in which the magnitude of the sensor signal of the force sensor is plotted against the force acting on the force sensor,
[0043] Figure 3 shows the time series of the packing force over four packing processes,
[0044] Figure 4 A schematic side cross-section of the brush body in the area of the plugged hole is shown, with the anchors positioned precisely and of the correct size.
[0045] Figure 5 A schematic top view of the brush body in the area of the plugged hole is shown, with the anchors being precisely positioned and of the correct size.
[0046] Figure 6 An overlay view of a multivariate time series of the packing force consisting of a plurality of time series reflecting a defect-free state,
[0047] Figure 7 A flow chart showing a method for training an artificial intelligence classifier according to the present invention is shown.
[0048] Figure 8 A flow chart showing a method according to the invention for monitoring a filling process is shown,
[0049] Figure 9 A schematic side cross-sectional view of the brush body in the area of the plugged hole is shown, wherein the anchor is the correct size, but is positioned too high in the plugged hole.
[0050] Figure 10 Shows an overlay view of the multivariate time series of the packing force consisting of multiple time series reflecting Figure 9 The defective state,
[0051] Figure 11 A schematic side cross-sectional view of the brush body in the area of the plugged hole is shown, wherein the anchor is the correct size, but is positioned too low in the plugged hole.
[0052] Figure 12 Shows an overlay view of the multivariate time series of the packing force consisting of multiple time series reflecting Figure 11 The defective state,
[0053] Figure 13 A schematic top view of the brush body in the area of the plugging hole is shown, wherein the anchor is of the correct size, but is positioned too far upwards at the plugging hole.
[0054] Figure 14 A schematic top view of the brush body in the area of the plugging hole is shown, wherein the anchor is correctly sized, but is positioned too far below the plugging hole.
[0055] Figure 15 Shows an overlay view of the multivariate time series of the packing force consisting of multiple time series reflecting Figure 13 and Figure 14 The defective state,
[0056] Figure 16 A schematic top view of the brush body in the area of the plugging hole is shown, wherein the anchor is correctly sized, but is positioned too far to the left at the plugging hole.
[0057] Figure 17 A schematic top view of the brush body in the area of the plugging hole is shown, wherein the anchor is correctly sized, but is positioned too far to the right at the plugging hole.
[0058] Figure 18 A view showing the superposition of a multivariate time series of packing force consisting of multiple time series reflecting Figure 16 and Figure 17 One of the defective states,
[0059] Figure 19A schematic top view of the brush body in the area of the plugged hole is shown, wherein the anchor which is too short is in the correct position.
[0060] Figure 20 Shows an overlay view of the multivariate time series of the packing force consisting of multiple time series reflecting Figure 19 The defective state,
[0061] Figure 21 A schematic top view of the brush body in the area of the plugged hole is shown, with the overlong anchor in the correct position.
[0062] Figure 22 An overlay view of a multivariate time series of the packing force consisting of a plurality of time series, each of which reflects a defective state due to an excess of wire,
[0063] Figure 23 An overlay view of a multivariate time series of the packing force consisting of a plurality of time series reflecting a defective state due to too little wire, and
[0064] Figure 24 Shown is an overlay of a multivariate time series of the filling force consisting of a plurality of time series, each of which reflects a defective state caused by an impact of the filling tool. DETAILED DESCRIPTION
[0065] Figure 1 A brush stuffer, generally designated 1, is shown, which is configured to carry out the method according to the present invention. However, the method according to the present invention is not limited to the brush stuffer shown as an example. Rather, the method can be easily transferred to other machines and other sensor arrangements.
[0066] The brush stuffing machine 1 has a stuffing tool 2 which comprises a push tongue 3. The push tongue 3 is provided for stuffing a bristle bundle 4 together with a fixing anchor into a bundle receiving opening 5 of a brush body 6.
[0067] The brush stuffing machine 1 comprises a drive device 7 which is connected to the push tongue 3 via a force transmission rod 8 and is provided for driving the push tongue 3 to perform a stuffing movement.
[0068] Figure 1 It is shown that the force transmission rod 8 is preloaded against a sensor, namely the force sensor 10, in the filling direction indicated by the arrow 9 of the push tongue 3. The direction of the preload force acting on the force sensor 10 is indicated by the arrow 24. Figure 1 As shown in the diagram.
[0069] The force sensor 10 , which is preloaded with a preload force 24 , is provided for measuring a force resulting from the preload force 24 and a tampon force 25 acting on the push tongue 3 when the bristle bundle 4 is being padded.
[0070] exist Figure 1 In the brush stuffing machine 1 shown, the force sensor 10 is arranged so that it is relieved by the stuffing force 25 acting on the force transmission rod 8 via the push tongue 3 against the stuffing direction 9. As the stuffing force increases, the total force measurable by the force sensor 10 decreases.
[0071] As the filling force increases, the force sensor 10 is relieved, i.e., by a correspondingly movable mounting of the force transmission rod 8. The force transmission rod 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 filling force 25 acting on the force transmission rod 8 via the push tongue 3 counter to the filling direction 9 of the push tongue 3.
[0072] If the filling force 24 acting against the filling direction 9 of the push tongue 3 is at the level of an overload force which is so great that it cancels the preload of the force transmission rod 8 on the force sensor 10, the force transmission contact between the force transmission rod 8 and the force sensor 10 will be lost.
[0073] The unloading of the force sensor 10 that occurs when the push tongue 3 is loaded is detected by the force sensor as a change in the resultant force and results in the output of a corresponding sensor signal. The sensor signal represents the resultant force currently present at the force sensor 10 and allows inferences to be drawn about the specific filling force 25 and, therefore, the load on the push tongue 3.
[0074] Since the force sensor 10 is unloaded by the load of the push tongue 3 against the stuffing direction 9, there is no need to worry about damage to the force sensor 10 when the push tongue 3 is overloaded in this direction. Therefore, the force sensor 10 is effectively protected from damage, but the monitoring of the push tongue 3 is not hindered.
[0075] The brush stuffer 1 has a two-part tongue 11 , wherein one part of the tongue 11 is the drive rod 12 and the second part of the tongue 11 is the aforementioned force transmission rod 8 .
[0076] The drive rod 12 is connected to the drive device 7 of the brush stuffer 1 via a connecting rod 13. The end of the drive rod 12 facing away from the connecting rod 13 is connected to the force transmission rod 8.
[0077] The drive rod 12 and the force transmission rod 8 are connected to each other in an articulated manner such that the force sensor 10 is relieved of stress by a filling force 25 acting on the force transmission rod 8 via the push tongue 3 counter to the filling direction 9. The filling force 25 causes the force transmission rod 8 to change its angular position relative to the drive rod 12 by pivoting it relative to the drive rod 12 about a pivot axis 30 of the articulated connection between the drive rod 12 and the force transmission rod 8.
[0078] The articulated connection between the drive rod 12 and the force transmission rod 8 is formed by a pivot joint 14. The force sensor 10 is arranged on the drive rod 12 and here between the drive rod 12 and the force transmission rod 8. By changing the angular position between the force transmission rod 8 and the drive rod 12, the force sensor 10 and the drive rod 12 can be moved away from the force transmission rod 8 as the filling force 25 increases and can be separated from the force transmission rod when the filling force 25 is so great that it counteracts the preload force 24.
[0079] The preload force 24 is generated by the preload device 15 of the brush stuffer 1. For this purpose, the force transmission rod 8 is preloaded in the stuffing direction 9 against the force sensor 10 by means of the preload device 15. The preload device 15 is arranged on the drive rod 12 and comprises a compression spring, designated 16, which tensions the force transmission rod 8 against the force sensor 10, which is arranged on the drive rod 12.
[0080] The brush stuffing machine 1 has a control unit 17. The brush stuffing machine 1 is configured to carry out the method according to the invention via the control unit 17. The control unit 17 is configured to reduce the stuffing speed of the push tongue 3 or to stop the push tongue 3 as soon as the resultant force measured by the force sensor 10 reaches or falls below a threshold value.
[0081] For this purpose, the control unit 17 can control the drive 7 of the brush stuffing machine 1 based on the sensor signal of the force sensor 10, that is, for example, as soon as the resultant force measured by the force sensor 10 reaches or falls below a limit value, the speed of the drive 7 is reduced or even stopped.
[0082] Here, the resultant force limit value that must be reached or lowered in order to reduce the filling speed of the push tongue 3 and / or to reduce the speed of the drive device 7 can be greater than the limit value that must be reached or lowered in order to stop the push tongue 3 and / or the drive device 7.
[0083] The brush making machine 1 has a bristle magazine 26 in which bristle filaments are arranged and a bundle splitter 27 with a bundle removal groove 28. The bundle splitter 27 moves past the bristle magazine 26 using the bundle removal groove 28 to remove bristle bundles 4 from the bristle magazine 26. The bundle splitter 27 then transfers the removed bristle bundles 4 to the downstream stuffing tool 2. The stuffing tool 2 then stuffs the bristle bundles 4 into the bundle receiving holes 5 of the brush body 6 arranged on the holding device 29 of the brush stuffing machine 1.
[0084] The brush stuffing machine 1 described above is provided for stuffing brushes. Here, the bristle bundles 4 are stuffed into the bundle receiving holes 5 of the prepared brush body 6 using the push tongues 3 of the stuffing tool 2 of the brush stuffing machine 1 .
[0085] In order to determine the stuffing force 25 acting on the push tongue 3 , the force resulting from the preload force 24 and the stuffing force 25 is measured by means of the force sensor 10 and the stuffing force 25 is derived therefrom.
[0086] Figure 2 The diagram in shows the characteristic curve of the force sensor 10 of the brush stuffing machine 1. In the diagram, the sensor signal 18 output by the force sensor 10 is plotted against a force 19 acting on the force sensor 10.
[0087] The diagram shows the measuring range start 20, the measuring range end 21, and the overload limit 22 of the force sensor 10. If the force sensor 10 is loaded beyond its overload limit 22, it could be damaged. Reliable measurements are only possible between the measuring range start 20 and the measuring range end 21.
[0088] The method according to the invention is now based on this arrangement and uses the existing force sensors also for monitoring the filling process according to the invention. As already mentioned above, the method according to the invention is not limited to Figure 1 Rather, the filling force can also be determined by other mechanical sensor arrangements and / or by other sensors (e.g., by strain gauges or by power extraction from the drive 7), which then serve as force sensors. Completely different mechanical configurations are also possible.
[0089] Figure 3 By way of example, a time series of recorded measured values of the filling force 25 is shown. Figure 3 The diagram of the embodiment of the present invention basically includes four filling processes. The starting point of the filling process begins when the force 23 set as the preload force is set, as shown in Figure 2 As shown in . Starting from this force, one filling process includes one rotation of the drive device 7.
[0090] Figure 4 and Figure 5 Schematically, different views of the brush body 6 are shown in the region of the filling hole or bundle receiving hole 5, in which the anchor 31 is located in its normal or non-defective position. The bundle is not shown here for the sake of clarity. Figure 3 The course of the packing force 25 in the embodiment of FIG. 1 shows a substantially flawless state, which includes the following: Figure 5 and 6 Anchor position shown.
[0091] In a defect-free state, the packing force 25 has a characteristic course over a time series that is essentially the same or at least similar for each beam receiving opening 5. The present inventors have recognized that deviations from this characteristic course can be indicative of a defective state. This understanding is based on the concept of identifying defects in the packing process based on deviations of the packing force from the characteristic course.
[0092] To detect defective states, an artificial intelligence classifier 34 is used according to the invention, which is essentially trained or has been trained using time series derived exclusively from defect-free measurements.
[0093] Figure 6 A plurality of such superimposed time series are shown as an example. It becomes clear here that the time series, although definitely having local deviations, nevertheless have essentially similar courses.
[0094] Such a set of time series can be used, for example, as a training data set for an artificial intelligence classifier. During operation, the recorded time series are transferred to the artificial intelligence classifier 34 trained in this manner. The artificial intelligence classifier 34 then determines whether the transferred time series can be classified as "good" based on the training data. The output of the artificial intelligence classifier can then be a binary value, i.e., "good" or "bad," or, for example, a confidence value or comparable value that allows for the assignment of "good" or "bad."
[0095] In this example, a time series includes, for example, 800 to 1000 measured values, with a sampling rate of approximately 16 kHz to 20 kHz. However, this is only one example for this brush stuffer. Other brush stuffers can operate with different stuffing frequencies, thereby also varying the sampling rate and / or the number of measured values per time series.
[0096] This number of measurements is quite high for use as input parameters. Therefore, in practice it can be advantageous not to use the time series directly as input parameters for artificial intelligence classifiers. For this reason, the time series can be preprocessed before use, for example by feature extraction or data reduction.
[0097] For example, for feature extraction, a time series can be transformed from the time domain into the frequency domain using a discrete Fourier transform. The specific Fourier coefficients, or parts thereof, can then be used, for example, as input parameters for an artificial intelligence classifier. This removes the time variable from the measured values during the filling process, allowing features contained in the time series at a specific frequency to be considered more clearly and independently of time.
[0098] Furthermore, for example, only a certain number of Fourier coefficients, for example the first 80, can be used. In this way, the original signal can be reconstructed with sufficient accuracy while significantly reducing the number of input parameters.
[0099] Another possibility for feature extraction is to determine wavelet coefficients. The discrete wavelet transform is suitable for this. In this case, similar to the discrete Fourier transform, multiple wavelet coefficients are determined that allow reconstruction of the original signal. The wavelet transform converts the signal from the time domain to the time-frequency domain. These wavelet coefficients can alternatively or additionally be used as input parameters for an artificial intelligence classifier.
[0100] Here, only a certain number of wavelet coefficients can be used to reduce the total number of input parameters.
[0101] There are different wavelet families available here, such as Haar, Daubechies, Coiflet 5, Symlet 6, Biorthogonal, and Revers-Biorthogonal.
[0102] Another possibility for feature extraction is to use a distance metric. Here, for example, Euclidean distance or elastic distance, such as dynamic time warping, can be used.
[0103] For example, the distance can be calculated locally by dividing the time series into a plurality of 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.
[0104] Alternatively, the time series data can be directly reduced without extracting features. This can be done, for example, by taking only one out of every two or four measured values into account. The unaccounted-for measured values can be ignored or taken into account by forming an average.
[0105] Furthermore, there are further possibilities for extracting features and / or reducing data, which are not listed here but which the skilled person can easily consider.
[0106] In principle, all of the above-mentioned or further methods for feature extraction and / or data reduction can also be combined with one another in any desired manner, if this should be necessary.
[0107] The characteristic course of the packing force depends on different parameters, such as the number of filaments per bristle bundle, the diameter of the bundle receiving hole or the anchor size. These parameters can be used as presets for each brush type, for example.
[0108] Now, for each such presetting, an artificial intelligence classifier 34 with good values must be trained. This artificial intelligence classifier must then be selected before the brush stuffing machine is put into operation.
[0109] However, the present invention has the advantage that good value training can be utilized during operation, so to speak "on the fly." Figure 7 The flow chart of such a training process 70 is shown as an example. Figure 1 , is switched on in step 71, wherein the presetting for the specific brush type is used. In a subsequent step 72, a time series of the filling force is determined for each of the predetermined number of beam receiving holes, and in a further step 73, input parameters for the artificial intelligence classifier 34 are determined from this, for example, by feature extraction as described above.
[0110] In a next step 74, these input parameters are transmitted as training data to a previously untrained, so to speak "empty," artificial intelligence classifier 34. In this case, it is initially assumed that all training data correspond to good conditions.
[0111] It is then checked 75 whether a predetermined amount of training data has been determined.
[0112] If the predetermined number has not been reached, the method in step 72 is repeated.
[0113] If there is sufficient training data, it is determined or confirmed by the user in a check step 76 that the brushes produced so far are acceptable and therefore all recorded training data can be classified as "good".
[0114] If a defect occurs, then in a continuing step 78 all training data will be discarded and training will restart with step 72 .
[0115] However, if the classification is “good”, the training ends 77.
[0116] The artificial intelligence classifier 34 trained in this way can then be used in operation, for example according to Figure 8 Flowchart in . It is also possible here that after the training, ie also during the user confirmation in step 76, the operation of the brush stuffing machine continues without interruption.
[0117] During operation according to method 80 , in a first step 81 , a time series of the filling forces is determined for a predetermined number of beam receiving openings, similarly to step 72 , and then in a subsequent step 82 , input parameters for artificial intelligence classifier 34 are determined therefrom, for example by feature extraction as described above.
[0118] In a next step 83 these input parameters are passed to the previously trained artificial intelligence classifier 34. The artificial intelligence classifier 34 classifies the submitted input parameters based on the training data that was rated as good and thereby determines whether these parameters are also good.
[0119] In the following checking step 84 , the result of the artificial intelligence classifier 34 is now evaluated. If there is no defect state, ie if the input parameter is classified as “good”, the method continues with step 81 .
[0120] If the result is classified as bad, a defective state may exist. In this example, in the case of a defect, provision is made for a manual inspection of the previously produced brush in a subsequent inspection step 85. Here, the user can manually determine whether a defect actually exists or whether the brush is normal and therefore not defective.
[0121] If the input parameters are deemed good, the method can continue with step 81 . Optionally, these input parameters can be added to the training data of the artificial intelligence classifier 34 in step 87 .
[0122] If a defect is confirmed, defect elimination 86 may be performed. This defect elimination may include, for example, identifying the defect type using a second artificial intelligence classifier. The second artificial intelligence classifier may have been previously trained using input parameters that match typical defect types.
[0123] Figures 9 to 23 Different defect types and their associated characteristic time series are shown as examples. For example, the second artificial intelligence classifier 34 can be trained using the features extracted from these time series as described above. Based on the types of deviations in the characteristic time series, the second artificial intelligence classifier 34 can match input parameters to defect types.
[0124] Figure 9 Show Figure 4 The brush body 6 is shown in FIG. 3 , but the anchor 31 is positioned too high in the bundle receiving hole 5. This deviation results in a packing force 25 of Figure 10 The time series shown.
[0125] Figure 11 Show Figure 4 The brush body 6 is shown in FIG. 3 , but the anchor 31 is positioned too low in the bundle receiving hole 5. This deviation results in a packing force 25 of Figure 12 The time series shown.
[0126] Figure 13 and Figure 14 Shown separately Figure 5 6, wherein the anchor 31 is positioned too high or too low in the bundle receiving opening 5. These deviations are symmetrical and result in a packing force 25 of the same Figure 15 The time series shown.
[0127] Figure 16 and Figure 17 Shown separately Figure 5 6, wherein the anchor 31 is positioned too far to the left or too far to the right in the bundle receiving opening 5. These deviations are symmetrical and result in a filling force 25 of the same magnitude as Figure 18 The time series shown.
[0128] Figure 19 Show Figure 5 6, wherein the anchor 31 is too short relative to the diameter of the bundle receiving hole 5, but is located in the correct position. This deviation leads to a decrease in the filling force 25. Figure 20 The time series shown in .
[0129] Figure 21 Another possible defect is shown in Figure 1. Figure 5 The illustration of the brush body 6 corresponds to FIG, wherein the anchor 31 is too long relative to the diameter of the beam receiving opening 5, but is located in the correct position. However, the time sequence is not shown for this purpose.
[0130] Other possible defect types include: too many filaments in the tow, which results in a decrease in the packing force 25. Figure 22 If there are too few filaments in the tow, the packing force may be reduced by 25%. Figure 23 time series.
[0131] Figure 24 The time sequence of the tamping force 25 which may occur during an impact of the tamping tool is shown.
[0132] By training the second artificial intelligence classifier 34 with such a time series, the artificial intelligence classifier 34 can distinguish and recognize such defect types. It may be helpful here if, as described above, independent input parameters are additionally present, for example input parameters of the second sensor.
[0133] In a further development of the invention, the artificial intelligence classifier 34 can be trained with elimination instructions for eliminating specific defects in addition to the defect type, so that in addition to the defect type, instructions for eliminating such defects can also be output by the artificial intelligence classifier 34.
[0134] Alternatively or additionally, maintenance instructions can also be stored in a database 33 , which may be cloud-based.
[0135] The brush stuffing machine 1 is configured to implement the method described above and has a corresponding device. As a device for implementing the method, the brush stuffing machine comprises a push tongue 3 for stuffing the bristle bundle 4, a force sensor 10 for determining the stuffing force 25 acting on the push tongue 3 when stuffing the bristle bundle 4, the control unit 17 already mentioned above, and an artificial intelligence classifier 34.
[0136] Reference Signs List
[0137] 1 Brush Stuffing Machine
[0138] 2. Caulking tools
[0139] 3. Tongue Push
[0140] 4 bristle bundles
[0141] 5 beam receiving holes
[0142] 6 brush body
[0143] 7 drive unit
[0144] 8 force transmission rods
[0145] 9 Stuffing direction
[0146] 10 force sensors
[0147] 11 Tongue bar
[0148] 12 drive rods
[0149] 13 connecting rod
[0150] 14 pivot hinges between 8 and 12
[0151] 15 Preload device
[0152] 16 pressure spring
[0153] 17 Control Unit
[0154] 18 sensor signals
[0155] The force 19 acts on 10
[0156] 20 Starting point of measurement range
[0157] 21 End point of measuring range
[0158] 22 Overload limit
[0159] 23 Preload setting point
[0160] 24 preload
[0161] 25 packing force
[0162] 26 bristle library
[0163] 27 beam splitter
[0164] 28 beam removal slots
[0165] 29 holding device
[0166] 30 pivot axis
[0167] 31 Anchor
[0168] 32 time series
[0169] 33 (Cloud) Database
[0170] 34 Artificial Intelligence Classifier
Claims
1. A method for monitoring the stuffing process of a brush stuffing machine (1), in particular a toothbrush stuffing machine, using an artificial intelligence-based classifier, characterized in that During the filling process, a time series (32) of the filling force (25) acting on the push tongue (3) when filling the bristle bundle (4) is determined, and an artificial intelligence classifier (34) uses multiple input parameters determined from the time series (25) and outputs whether a defective state exists, wherein the artificial intelligence classifier (34) has been trained using input parameters representing a defect-free filling process.
2. The method according to claim 1, characterized in that For each packing process, a time series (32), in particular a univariate time series, is determined from the packing force (25).
3. The method according to any one of the preceding claims, characterized in that Form a multivariate time series from a predetermined number of time series.
4. The method according to any one of the preceding claims, characterized in that The input parameters are determined from the time series (32) by feature extraction and / or data reduction.
5. The method according to claim 4, characterized in that When extracting features, a predetermined number of Fourier coefficients and / or a predetermined number of wavelet coefficients are determined as features, and / or a distance from the median is determined for a predetermined number of intervals of the time series (32), in particular the predetermined number is 80.
6. The method according to claim 4 or 5, characterized in that To reduce the data, the number of measurements in the time series (32) was reduced, in particular such that 200 measurements were retained per time series.
7. The method according to any one of the preceding claims, characterized in that In order to train the artificial intelligence classifier (34) on the fly, a predetermined number of time series (32) are determined and the input parameters determined therefrom are defined as good, for example by a user.
8. The method according to any one of the preceding claims, characterized in that For input parameters that are rated as bad, the user can determine whether the input parameter is good or bad, and in particular, add the input parameter to the training data of the artificial intelligence classifier (34) when it is rated as good.
9. The method according to any one of the preceding claims, characterized in that To determine the filling force ( 25 ), a sensor, in particular a force sensor ( 10 ), is used.
10. The method according to any one of the preceding claims, characterized in that The artificial intelligence classifier (34) has at least one further input parameter derived from a further sensor.
11. The method according to any one of the preceding claims, characterized in that The type of the defect state is determined using a second artificial intelligence classifier (34), wherein input parameters are determined for training the second artificial intelligence classifier (34) from time series of the filling forces respectively associated with the determined defect state.
12. The method according to any one of the preceding claims, characterized in that After a defective state occurs, the plugging machine is repaired, for example, by an operator, and after the repair, information about the defect type and its elimination is stored in a database (33), in particular a cloud database, and / or added to the training data of an artificial intelligence classifier, in particular a second artificial intelligence classifier (34), and when the defective state reappears, in addition to the defect type, the artificial intelligence classifier (34) also outputs a prompt about eliminating the defective state.
13. The method according to any one of the preceding claims, characterized in that The artificial intelligence classifier (34) and / or the second artificial intelligence classifier is a so-called one-class support vector machine, an isolation forest, an autoencoder or a convolutional autoencoder.
14. A brush stuffing machine (1), in particular a toothbrush stuffing machine, comprising a device, wherein the brush stuffing machine (1) is configured to implement the method according to one of claims 1 to 13 by means of the device, in particular the brush stuffing machine (1) comprising at least one push tongue (3) for stuffing a bristle bundle (4), a sensor for determining a stuffing force (25) acting on the push tongue (3) when stuffing the bristle bundle (4), in particular a force sensor (10), a control unit (17) and at least one artificial intelligence classifier (34) as a device for implementing the method.