Gluing detection system and method

Abstract

The invention provides a gluing detection system which is used for detecting a gluing layer coated on an object to be detected. The gluing detection system comprises a bearing device and a detection device. The bearing device is used for unfolding and folding the to-be-tested object. The detection device comprises a radio module, a memory and a processor. The sound receiving module is used for receiving sound when the to-be-detected object is unfolded and generating first audio data. The memory is used for storing the classification model. The processor is used for preprocessing the first audio data to generate second audio data; performing feature extraction on the second audio data to generate feature data; inputting the feature data into a classification model to judge whether the feature data belongs to normal gluing distribution or abnormal gluing distribution; and outputting an abnormal warning in response to the situation that the glue coating distribution is judged to be abnormal. Embodiments of the present disclosure may utilize a support vector machine to achieve high accuracy of gluing anomaly detection with a small number of samples.

Description

Technical Field

[0001] This invention provides a glue coating detection system, and more particularly a system and method for detecting glue coating anomalies using machine learning. Background Technology

[0002] In the production process of radio-frequency identification (RFID) tags or other adhesive tags, uneven coating of the adhesive layer may occur. During stacking or winding and storage, the pressure between the layers may cause the adhesive layer to exceed the area that should be coated, resulting in adhesive overflow.

[0003] If glue overflow occurs, in some applications where the label and backing paper need to be rolled up together, such as when using an automatic labeling machine, when the rolled-up label and backing paper are unrolled, the overflowing glue will cause the backing papers to stick together, causing the automatic labeling machine to jam and fail to operate normally.

[0004] Traditionally, the determination of adhesive overflow is usually done manually. For example, quality inspectors would directly touch the surface of the label with their hands to determine whether there was adhesive overflow at the bond between the label and the backing paper. This method relies on the experience of the quality inspectors, which is not only labor-intensive but also inefficient and inaccurate, and may even damage the label itself.

[0005] Therefore, an efficient detection method is needed to determine whether adhesive overflow has occurred. Automating the entire detection process without human intervention can improve efficiency, save manpower, and increase the accuracy of the detection. Summary of the Invention

[0006] This invention provides an adhesive coating detection system for detecting adhesive coatings applied to a test object. The system includes a support device and a detection device. The support device is used to unfold and retract the test object. The detection device includes a sound receiving module, memory, and a processor. The sound receiving module is used to receive sound when the test object is unfolded, generating first audio data. The memory is used to store a classification model. The processor is coupled to the sound receiving module and the memory, and is used to: preprocess the first audio data to generate second audio data; extract features from the second audio data to generate feature data; input the feature data into the classification model, and use the classification model to determine whether the feature data belongs to a normal or abnormal adhesive coating distribution; and, in response to the determination of an abnormal adhesive coating distribution, output an abnormal warning.

[0007] In some embodiments, when the processor performs the preprocessing, it is further configured to: convert the first audio data from the time domain to the frequency domain to obtain first frequency domain data; adjust the gain of the first frequency domain data in each of the multiple frequency bands to obtain second frequency domain data; and convert the second frequency domain data to the time domain to obtain the second audio data.

[0008] In some embodiments, when the processor performs the feature extraction, it is further configured to: calculate a plurality of audio features of the second audio data, the plurality of audio features including the zero-crossing rate, energy, energy entropy, spectral centroid or spectral slip of the second audio data, to generate the feature data; and store the feature data in the memory.

[0009] In some embodiments, the plurality of audio features further include the spectral entropy, spectral transition, spectral dispersion, Mel frequency cepstral coefficients, chromaticity, or chromaticity difference of the second audio data.

[0010] In some embodiments, the classification model has been trained by a computing device before being stored in the memory. The training includes: obtaining multiple normal audio samples and multiple abnormal audio samples; inputting the multiple normal audio samples and the multiple abnormal audio samples as a training dataset into the classification model by the computing device; adjusting multiple parameters of the classification model for the training dataset to optimize the objective function, wherein the objective function depends on the multiple parameters of the classification model and the training dataset; and storing the multiple parameters of the classification model in the memory in response to the optimization of the objective function.

[0011] In some embodiments, the test object includes a face material to be bonded and a backing paper, and the face material to be bonded is coated with an adhesive layer to adhere to the backing paper.

[0012] In some embodiments, the carrier is a roll-to-roll device, comprising: a first roll for unfolding the object to be tested; and a second roll at a first distance from the first roll for retracting the object to be tested; wherein the sound receiving module is at a second distance from the unfolded part of the object to be tested, and the second distance is less than the first distance.

[0013] In some embodiments, the first roll further includes a first clip and a second clip, and the object to be tested is located between the first clip and the second clip.

[0014] This disclosure provides a method for detecting adhesive coating, applicable to the test object. The method includes: unfolding the test object by a support device; recording sound by a sound receiving device while the test object is unfolded to generate first audio data; preprocessing the first audio data by a processor to generate second audio data; extracting features from the second audio data by the processor to generate feature data; inputting the feature data into a classification model by the processor, and using the classification model to determine whether the feature data belongs to a normal adhesive coating distribution or an abnormal adhesive coating distribution; and outputting an abnormality warning by the processor in response to the determination that the adhesive coating is abnormal.

[0015] In some embodiments, the step of preprocessing the first audio data by the processor to generate the second audio data includes: converting the first audio data from the time domain to the frequency domain by the processor to obtain first frequency domain data; adjusting the gain of the first frequency domain data in each of a plurality of frequency bands by the processor to obtain second frequency domain data; and converting the second frequency domain data to the time domain by the processor to obtain the second audio data.

[0016] In some embodiments, the classification model has been trained before being stored in memory. The training includes: obtaining multiple normal audio samples and multiple abnormal audio samples; inputting the multiple normal audio samples and the multiple abnormal audio samples as a training dataset into the classification model; adjusting multiple parameters of the classification model for the training dataset to optimize an objective function, wherein the objective function depends on the multiple parameters of the classification model and the training dataset; and storing the multiple parameters of the classification model in memory in response to the optimization of the objective function.

[0017] The embodiments disclosed herein can utilize support vector machines to achieve high accuracy in detecting adhesive coating anomalies with a small number of samples. Simultaneously, by using fewer important audio features, the efficiency of model training and operation can be increased. By automating the entire detection process without human intervention, efficiency can be improved, manpower can be saved, and detection accuracy can be increased. Attached Figure Description

[0018] Figure 1A This is a schematic diagram of an adhesive coating detection system according to a first embodiment of the present invention;

[0019] Figure 1B This is a cross-sectional view of the test object according to the first embodiment of the present invention;

[0020] Figure 1C This is a perspective view of the first roll according to a first embodiment of the present invention;

[0021] Figure 2 This is a functional block diagram of the detection device according to the first embodiment of the present invention;

[0022] Figure 3 This is a flowchart of the adhesive coating detection method according to the second embodiment of the present invention;

[0023] Figure 4 This is a detailed flowchart of the steps of the adhesive coating detection method according to the second embodiment of the present invention.

[0024] [Symbol Explanation]

[0025] 100: Adhesive Application Inspection System

[0026] 110: Test Item

[0027] 120: Detection device

[0028] 130: Bearing device

[0029] 131: First Roll

[0030] 131a: Expanded section

[0031] 131b: Clip

[0032] 132: Second roll

[0033] D1, D2: Distance

[0034] 111: Base paper

[0035] 112: Adhesive coating layer

[0036] 113: Surface material

[0037] 121: Radio module

[0038] 122: Processor

[0039] 123: Memory

[0040] 124: Classification Model

[0041] 300: Method

[0042] S310~S370: Steps Detailed Implementation

[0043] The embodiments of the present invention will be described below with reference to the accompanying drawings. In the drawings, the same reference numerals denote the same or similar components or method flows.

[0044] A first embodiment of the present invention provides an adhesive coating detection system 100. (See reference...) Figure 1A ,in Figure 1A This is a schematic diagram of an adhesive coating detection system 100 according to an embodiment of the present invention. The adhesive coating detection system 100 is used to detect a test object 110. In some embodiments, the adhesive coating detection system 100 is used to check whether the distribution of the adhesive layer coated on the test object 110 conforms to production specifications (e.g., whether it is uniformly distributed, or whether it exceeds the required coating range).

[0045] Please refer to the above. Figure 1B ,in Figure 1B This is a cross-sectional view of the test object 110 in the first embodiment. For example... Figure 1B As shown, the test object 110 includes a backing paper 111, an adhesive layer 112, and a top layer 113, and when the test object 110 is rolled up, the test objects 110 overlap with each other.

[0046] For example, the object under test 110 can be a radio-frequency identification (RFID) tag or other adhesive label. Taking an RFID tag as an example, the object under test 110 may contain an antenna pattern (not shown) printed on the face material 113, and an adhesive layer 112 is applied to the face material 113, so that the face material 113 is attached to the backing paper 111 through the adhesive layer 112.

[0047] In practical applications, if the adhesive layer 112 on the test object 110 is unevenly distributed, it may extend beyond the surface material 113, resulting in adhesive overflow. If overflow occurs, the rolled-up test objects will stick together when unrolled, which is undesirable in some applications. For example, in automatic labeling machines, if overflow causes the rolled-up labels to stick together, the machine may jam and malfunction. Therefore, an efficient detection method is needed to determine whether adhesive overflow has occurred on the adhesive layer 112 distribution on the test object 110.

[0048] In some embodiments, the adhesive coating detection system 100 includes a detection device 120 and a carrier device 130. The carrier device 130 is used to unfold and retract the test object 110. In some embodiments, the carrier device 130 includes a first roll 131 and a second roll 132, wherein the first roll 131 and the second roll 132 are separated by a first distance D1 (e.g., one meter to two meters). The test object 110 is unfolded on the unfolded portion 131a of the first roll 131 and retracted on the second roll 132. During the unfolding and retraction of the test object 110 by the carrier device 130, the detection device 120 is used to automatically distinguish the adhesive coating distribution of the test object 110.

[0049] Please refer to the following: Figure 2 ,in Figure 2 This is a functional block diagram of a detection device 120 according to an embodiment of the present invention. The detection device 120 includes a sound receiving module 121, a processor 122, and a memory 123. The sound receiving module 121 is used to receive sound and generate first audio data when the object to be tested 110 is unfolded. The memory 123 is used to store a classification model 124. The processor 122 is coupled to the sound receiving module 121 and the memory 123.

[0050] In some embodiments, such as Figure 1C As shown, the first reel 131 further includes a pair of clips 131b, including a first clip and a second clip. The clips 131b are used to clamp the object to be tested 110, which is beneficial for transmitting the sound emitted when the object to be tested 110 is unfolded, making the sound reception clearer.

[0051] In some embodiments, the sound receiving module 121 of the detection device 120 may include a microphone, microphone array, or other sensing circuit with sound receiving function. The detection device 120 and its sound receiving module 121 are disposed adjacent to the unfolded portion 131a. The sound receiving module 121 is used to collect the sound emitted when the test object 110 is unfolded. The processor 122 of the detection device 120 analyzes the collected sound and determines whether the adhesive layer 112 on the test object 110 is normally distributed or abnormally distributed. For example, if the sound is relatively low when the test object 110 is unfolded, it may indicate that the adhesive coating meets the production specifications (i.e., no adhesive overflow has occurred); if the test object 110 emits a loud and high-frequency sound when unfolded, it may indicate that the adhesive coating does not meet the production specifications (i.e., adhesive overflow has occurred), etc.

[0052] To ensure accurate sound reception, the detection device 120 and its sound receiving module 121 are positioned near the unfolded portion 131a of the object under test 110. As shown in the embodiment of FIG1, the detection device 120 is placed at a second distance D2 (e.g., 20 to 30 centimeters) from the unfolded portion 131a of the object under test 110, wherein the second distance D2 is less than the first distance D1. In other words, the detection device 120 is positioned closer to the first reel 131 used to unfold the object under test 110 and relatively farther away from the second reel 132 used to retract the object under test 110.

[0053] refer to Figure 3 ,in Figure 3 This is a flowchart of an adhesive coating detection method 300 according to an embodiment of the present invention. In some embodiments, Figure 1A and Figure 2 The adhesive application detection system 100 shown can be used to perform... Figure 3 The adhesive coating test method 300 is shown.

[0054] First, in step S310, the object under test 110 is unfolded, generating a sound. In step S320, the sound from the unfolded object 110 is received and converted by the receiving device 121 to generate first audio data. In step S330, the processor 122 preprocesses the first audio data to generate second audio data. In some embodiments, the preprocessing stage may include adjusting the gain of the first audio data in each frequency band (i.e., adjusting the EQ). In some embodiments, the sound from the unfolded object may correspond to certain specific frequency bands, and thus the recognition accuracy can be improved by increasing the gain of these frequency bands. In some embodiments, noise (e.g., noise emitted by factory machinery) may have relatively fixed frequencies, and the noise can be reduced by decreasing the gain of these frequency bands.

[0055] refer to Figure 3 and Figure 4 ,in Figure 4This is a detailed flowchart of step S330 of the adhesive coating detection method 300 according to a second embodiment of the present invention. Specifically, in step S331, the processor 122 first converts the first audio data from the time domain to the frequency domain to obtain first frequency domain data; in step S332, the processor 122 adjusts the gain corresponding to each frequency band of the first frequency domain data to obtain second frequency domain data; in step S333, the processor 122 converts the second frequency domain data back to the time domain to obtain second audio data. In some embodiments, the purpose of preprocessing is to eliminate background noise from machine operation or noise such as human voices. Since these noises typically have relatively fixed frequency bands, noise suppression can be achieved by reducing the gain of the corresponding frequency bands.

[0056] In step S340, processor 122 performs feature extraction on the second audio data to generate feature data. The feature data includes multiple features calculated directly or indirectly from the second audio data. In some embodiments, these multiple features include zero-crossing rate, energy, energy entropy, spectral centroid, spectral rolloff, spectral entropy, spectral flux, spectral spread, Mel-frequency cepstral coefficients, chroma vector, chromadeviation, and / or any other features calculated directly or indirectly from the audio data. In other embodiments, these multiple features may include only a few more important features, such as zero-crossing rate, energy, energy entropy, spectral centroid, and spectral rolloff. Selecting to retain fewer, more important features helps to effectively increase the training and operating speed of the model without compromising its discriminative ability.

[0057] Zero-crossing rate can be used to describe changes in sound. When adhesive is peeled off, the sound changes drastically, causing an increase in the number of zero-crossing points in the audio. The zero-crossing rate can effectively capture this drastic change, thus helping to identify adhesive aberrations. The zero-crossing rate is commonly used in audio event detection to distinguish between abnormal audio events and a static background.

[0058] Energy represents the intensity of an audio signal over time, while energy entropy provides a measure of the randomness of energy distribution, that is, the degree of variation in the audio signal. The sound produced when adhesive is torn off typically has an energy spike, followed by a rapid drop in energy. Therefore, energy entropy can effectively capture this variation.

[0059] The centroid of the spectrum is commonly used to assess the brightness of a sound. When adhesive is peeled off, the centroid value may be higher because the sound contains high-frequency components; conversely, when there is no sound or the sound does not contain high-frequency components, the centroid value may be lower. Therefore, it helps distinguish the sound of peeling adhesive from background noise. Additionally, the centroid can be used to differentiate sound patterns and helps to isolate noise.

[0060] Spectral sliding can be used to help capture changes in the high-frequency components of a sound. When the adhesive is peeled off, the high-frequency components suddenly intensify, and this abrupt energy change can be reflected in a measurement of spectral sliding because this metric changes as the sound changes.

[0061] In step S350, the processor 122 inputs the feature data into the classification model 124 and uses the classification model 124 to determine whether the feature data belongs to a normal or abnormal distribution of glue application. The classification model 124 is pre-trained and stored in memory 123, so it can be directly provided to the processor 122 for use.

[0062] Next, in step S360, if the classification model 124 determines that the feature data belongs to a normal adhesive distribution, the process ends; if the classification model 124 determines that the feature data belongs to an abnormal adhesive distribution, then proceed to step S370, where the processor 122 outputs an abnormality warning. In some embodiments, this abnormality warning can be provided to an alarm, causing the alarm to activate and emit an alarm sound to warn the user. In some embodiments, the abnormality warning can also be provided to a server via a communication interface (not shown in the figure). After receiving the abnormality warning, the server can transmit a message to the user's device via the network to notify the user of the abnormality.

[0063] The following describes the training of classification model 124. Training of classification model 124 can be performed on an external computing device or on processor 122. First, classification model 124 needs to obtain multiple normal audio samples and multiple abnormal audio samples. These normal and abnormal audio samples must undergo preprocessing and feature extraction as described in steps S330-S340. After these audio features are calculated into a format acceptable to classification model 124, these audio features can be selectively normalized, limiting the value range of all features to the same or at least approximately the same range (e.g., between zero and one). This is particularly important for models whose optimization function involves distance calculations, such as support vector machines (SVM), linear regression, and k-nearest neighbors (KNN) algorithms. In this case, without normalization, due to differences in feature value ranges, some features with larger value ranges will inherently have larger weights, thus overestimating the importance (i.e., the degree of influence) of those features.

[0064] Next, the processed normal and abnormal audio samples are used as training data and input into classification model 124. For this training dataset, several parameters of classification model 124 are adjusted to optimize a manually defined objective function, which depends on the parameters of classification model 124 and the training dataset. Typically, this objective function is set to be minimized; this function is called the loss function or cost function. Then, in response to the optimization of the objective function, the parameters of the classification model are stored in memory. Thus, classification model 124 is considered to have completed training.

[0065] Typically, because anomalous events are not so frequent (hence the term "anomaly"), the number of normal audio samples will exceed the number of anomalous audio samples. In this case, the proportion of the loss due to incorrect judgment in the loss function can be increased. In some embodiments, since the patterns of normal audio samples are relatively fixed, a strategy similar to anomaly detection can be used to identify outlier anomalous audio samples. In other embodiments, a semi-supervised learning approach can be adopted, using the training dataset as labeled samples, while the model can continue to be trained using audio data as unlabeled samples during the operational phase.

[0066] In some embodiments, the classification model 124 can be a supervised support vector machine (SVM). Since an SVM finds a hyperplane in a feature space that best separates two classes of samples, where optimization (maximization) is applied to the distance between the nearest data sample point to the hyperplane in the feature space and the hyperplane itself, this is a model whose loss function involves distances in the feature space. Standardization techniques can be used to improve accuracy. SVMs offer significant advantages over deep learning, which requires a large number of samples for training, when the number of samples is small and the feature dimension is large (i.e., a large number of features are used for judgment). Therefore, it is particularly suitable for applications involving the detection of adhesive application anomalies based on audio signals.

[0067] In summary, the embodiments of the present invention can achieve high accuracy in detecting adhesive coating anomalies with a small number of samples using support vector machines. Simultaneously, by selecting fewer important audio features, the efficiency of model training and operation can be increased. By automating the entire detection process without human intervention, efficiency can be improved, manpower can be saved, and detection accuracy can be increased.

[0068] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A glue application detection system, characterized in that, This adhesive coating detection system is used to detect adhesive layers applied to a test object and includes: A carrying device for unfolding and retracting the object under test; The detection device includes: The sound receiving module is used to collect sound when the object under test is unfolded, and generate the first audio data; Memory is used to store the classification model; and A processor, coupled to the radio module and the memory, is used for: The first audio data is preprocessed to generate the second audio data; Feature extraction is performed on the second audio data to generate feature data; The feature data is input into the classification model, which then determines whether the feature data belongs to a normal or abnormal glue application distribution. In response to the determination that the adhesive coating is abnormally distributed, an abnormal warning is output.

2. The adhesive coating detection system as described in claim 1, characterized in that, When the processor performs this preprocessing, it is further used for: The first audio data is converted from the time domain to the frequency domain to obtain the first frequency domain data; Adjusting the gain of the first frequency domain data for each of the multiple frequency bands yields the second frequency domain data; and The second frequency domain data is converted to the time domain to obtain the second audio data.

3. The adhesive coating detection system as described in claim 2, characterized in that, When this processor performs this feature extraction, it is more used for: Calculate multiple audio features of the second audio data, including the zero-crossing rate, energy, energy entropy, spectral centroid, or spectral slip of the second audio data, to generate the feature data; as well as Store the feature data in this memory.

4. The adhesive coating detection system as described in claim 3, characterized in that, These multiple audio features further include the spectral entropy, spectral transition, spectral dispersion, Mel frequency cepstral coefficients, chromaticity, or chromaticity difference of the second audio data.

5. The adhesive coating detection system as described in claim 4, characterized in that, The classification model was trained by a computing device before being stored in the memory. The training included: Multiple normal audio samples and multiple abnormal audio samples were obtained; The computing device inputs the multiple normal audio samples and the multiple abnormal audio samples as training datasets into the classification model. For the training dataset, several parameters of the classification model are tuned to optimize the objective function, where the objective function depends on the several parameters of the classification model and the training dataset; as well as In response to the optimization of the objective function, the multiple parameters of the classification model are stored in the memory.

6. The adhesive coating detection system as described in claim 1, characterized in that, The test object includes a face material to be bonded and a backing paper, and the face material to be bonded is coated with an adhesive layer to adhere to the backing paper.

7. The adhesive coating detection system as described in claim 1, characterized in that, The carrier is a roll-to-roll device, comprising: The first roll is used to unfold the object to be tested; as well as The second roll, located at a first distance from the first roll, is used to collect the object to be tested; The second distance between the sound receiving module and the unfolded part of the object under test is less than the first distance.

8. The adhesive coating detection system as described in claim 7, characterized in that, The first roll further includes a first clip and a second clip, and the object to be tested is located between the first clip and the second clip.

9. A method for detecting adhesive coating, characterized in that, Applicable to the test object, this method includes the following steps: The object to be tested is unfolded by the support device; When the object under test is unfolded, the sound is picked up by the sound receiving device, generating the first audio data; The processor preprocesses the first audio data to generate the second audio data; The processor extracts features from the second audio data to generate feature data; The processor inputs the feature data into a classification model, which then determines whether the feature data belongs to a normal or abnormal glue application distribution. In response to the determination that the adhesive application is abnormally distributed, the processor outputs an abnormal warning.

10. The adhesive coating detection method as described in claim 9, characterized in that, The step of preprocessing the first audio data by the processor to generate the second audio data includes the following steps: The processor converts the first audio data from the time domain to the frequency domain to obtain the first frequency domain data; The processor adjusts the gain of the first frequency domain data in each of the multiple frequency bands to obtain the second frequency domain data. as well as The processor converts the second frequency domain data to the time domain to obtain the second audio data.

11. The adhesive coating detection method as described in claim 10, characterized in that, The classification model was trained before being stored in memory, and the training included the following steps: Multiple normal audio samples and multiple abnormal audio samples were obtained; The multiple normal audio samples and the multiple abnormal audio samples are used as training datasets and input into the classification model. For the training dataset, several parameters of the classification model are tuned to optimize the objective function, where the objective function depends on the several parameters of the classification model and the training dataset; as well as In response to the optimization of the objective function, the multiple parameters of the classification model are stored in the memory.

Images