Ai-based method and system for quality control on pieces using ultrasonic signals

Abstract

The present invention relates to an artificial intelligence-based method and system for detecting anomalies in objects, in particular brake discs, by means of ultrasounds. The suggested method uses artificial intelligence tools to classify the anomalies by transferring the ultrasonic signal, which contains information about the discs, to the images and obtaining a fast result with a low level of computation.

Description

[0001] Al-based method and system for quality control on pieces using ultrasonic signals

[0002] To: Brembo S.p.A.

[0003] Inventors: Mattia Lorenzetti, Luca Loria, Elena

[0004] Mazzoleni, Micael Rescati, Enea Zappella, Junqian Zhang.

[0005] The present invention relates to an Al-based method and system for quality control on pieces using ultrasonic signals, in particular on brake discs.

[0006] More in particular, the present invention relates to an artificial intelligence-based method and system for detecting rejects on brake discs, by means of ultrasounds. The suggested method employs artificial intelligence tools to identify the presence of anomalies, in particular cracks, in the material by transferring the ultrasonic signal, which contains the information about the discs, to the images and rapidly obtaining a good / reject piece classification result, respecting the cycle times.

[0007] Background art

[0008] During the manufacturing of brake discs, it is necessary to inspect products to identify cracks and isolate the defective discs from the good ones. This activity is particularly challenging because the time devoted to controls must respect tight manufacturing schedules, but finding cracks can be complex, especially because some are not superficially visible.

[0009] A possible solution is to use an X-ray machine to identi fy defects not visible superficially, but the disadvantage is that it is expensive and requires a lot of maintenance .

[0010] In an earlier attempt to identi fy cracks , in patent document CN112082742A, the inventors built a library of cracks and expanded it with finite element simulations . In order to identi fy whether the new disc has a crack and where it is located, they compared the disc with each crack in the library and calculated the similarity . However, collecting all possible cracks is expensive and simultaneously requires more time and more computational resources to compare each tested disc with all the crack types in the library .

[0011] With the development of laser technology, laser- produced ultrasounds can be a good tool to obtain information about the quality of the disc . There are more limitations and di f ficulties when the ultrasounds are applied to disc manufacturing . One of the greatest di f ficulties in creating an automated system for classi fying a good or rej ect disc based on ultrasonic technology is that of knowing the types of defects which will occur and having defective specimens available . A defect in the discs is usually associated with a particular manipulation of the pieces , but the frequency is low and of the order of ppm . The lack of availability or the absence of defective pieces prevents the possibility of applying the most common defect recognition algorithm . The cracks are found on the surface of the pieces , but they are not always visible to the naked eye . The depiction of the disc ring and an example of a reject disc are shown in Figures 1 and 2, respectively, in which the crack observation zone is dashed in Fig. 1, an optical image is shown in Fig. 2 (a) , and an X-ray image is shown in Fig. 2 (b) . For this reason, surface inspection is not sufficient, but an alternative, again noninvasive, system is needed.

[0012] A possible solution is to find a crack by converting the signals from the ultrasounds into images and identifying those exhibiting anomalies. CN111598881B suggested a method of detecting anomalies from images, but the detected anomalies were real objects in the images. US11688415B2 employs a neural network to monitor the machine operation condition starting from sound signals from sound acquisition devices. The features are extracted from spectrograms and by classifying the operation into normal or abnormal. This solution is used to identify the malfunction of machinery, at some points in the process, considering audible sound, but not to classify the features of physical products. A technical problem is that the recorded sound can exhibit noise due to the surrounding environment, equipment, and workers .

[0013] Patent document US 11,378,551 describes a method for determining the presence of one or more discontinuities in the object based on emitted and then detected ultrasonic signals. Such a solution classifies every single detail of an object to detect anomalies. Patent document US11688415B2 provides a system and methods for classifying machine anomalies and behavior. An audio acquisition device can be connected to a mechanical apparatus comprising a first component and a second component. The first component and the second component can generate audible noise separately. The audio acquisition device can generate an audio response signal caused by the vibration diagram. The system can receive the signal generated by the microphone of the audio acquisition device. The system can determine, based on a machine learning model and on the signal, an abnormal event associated with the first component, a second component, or a combination thereof. Alternatively or additionally, the system can classify the machine operation according to the second machine learning model .

[0014] The article by HUA CHENQUAN ET AL: "Defect Detection Method of Carbon Fiber Sucker Rod Based on Multi-Sensor Information Fusion and DBN Model", SENSORS, vol. 22, no. 14, 11 July 2022 (2022-07-11) , page 5189, XP093179913, ISSN: 1424-8220, DOI: 10.3390 / x22145189,

[0015] (URL : https : / / www.mdpi . com / 1424- 8220 / 22 / 14 / 5189 / pdf [retrieved on 2024-06-27) uses both ultrasound and image recognition for crack detection in mechanical pieces. However, the authors use ultrasound to generate the images, but not use the images directly to detect the cracks. Indeed, the authors clean the images by pre-processing and extract features of their texture to detect cracks in mechanical pieces, which requires high effort and long time.

[0016] The article by CASTELLINI P ET AL: "Laser Doppler Vibrometry: Development of advanced solutions answering to technology’s needs", MECHANICAL SYSTEMS AND SIGNAL PROCESSING, ELSEVIER, AMSTERDAM, NL, vol. 20, no. 6, 1 August 2006 (2006-08-01) , pages 1265-1285, XP024930378, ISSN: 0888-3270, DOI: 10.1016 / J. YMSSP, describes applications of Laser Doppler Vibrometry (LDV) . One of the most significant applications covered in the paper is the analysis brake disc vibrations, in particular to address the phenomenon of "brake squeal". This application brings together several complex challenges for the measurement systems, including the ability to operate on rotating structures with low roughness surfaces, extract detailed information on 3D vibrations, and make measurements over a wide frequency range with high accuracy. Signal processing plays a key role in improving LDV performance by addressing noise problems and improving data quality through advanced techniques, such as filtering, windowing, and speckle noise reduction. Autofocus and automatic parameter optimization systems simplify the set-up, while innovations, such as multipoint vibrometry, allow faster data acquisition and analysis of transient events. Despite its versatility, LDV has challenges such as sensitivity to environmental factors, signal-to-noise ratio problems on poorly diffusive surfaces, and limitations in spatial resolution and operating distance for large structures.

[0017] Document CN114414660 A describes a method of identifying cracks in a wheel set of a railway vehicle. The method addresses interference caused by uneven illumination, reflections, scratches, stains, and rust in images (photographs) acquired for the wheel set, improving the image quality through a method of enhancing the axis number images . Furthermore , an improved method for recogni zing cracks based on principal component analysis ( PCA) is used to identi fy and report cracks in the acquired images for the detection of defects in the wheel set . This technique has the drawback of having to work on a plurality of photographs in the case of complex pieces , such as brake discs , making the analysis timeconsuming .

[0018] A need is felt for a solution based on signal and ultrasounds in a closed environment , not subj ect to external noise and aimed at identi fying cracks in the material of a brake disc by searching for anomal ies in ultrasonic signals , with an advantageous cycle time achieved by utili zing image- speci fic deep learning algorithms which allow faster prediction .

[0019] Object and subject-matter of the invention

[0020] It is the obj ect of the present invention to provide an Al-based method and system for quality control on pieces using ultrasonic signals , which overcome the drawbacks and solve the problems of the prior art .

[0021] In particular, it is an obj ect of this invention to automate the detection of anomalies , in particular cracks , in discs for all discs manufactured with ultrasonic technology, without having a database of defective pieces available .

[0022] The present invention relates to a method and system according to the appended claims . Detailed description of embodiments of the invention

[0023] List of drawings

[0024] The invention will now be described by way of a nonlimiting illustration, with particular reference to the figures in the accompanying drawings, in which:

[0025] - figure 1 shows an inspected brake disc ring, according to the prior art;

[0026] - figure 2 shows an example of a crack in a disc tooth, optical image on the left, X-ray image on the right, according to the prior art;

[0027] - figure 3 shows an example of inspection by means of ultrasounds, according to an embodiment of the invention;

[0028] - figure 4 shows a flow chart of an embodiment of the method of the invention;

[0029] - figure 5 shows an example of input to the Al model for a disc in the absence of anomalies. It is an image depicting a disc with 18 teeth, in which each individual spectrogram is obtained from a wavelet transform for each signal, i.e., each tooth of the disc;

[0030] - figure 6 shows a Patch Distribution Modeling (PaDiM) pattern to detect anomalies in an image;

[0031] - figure 7 shows an example of the output of the Al model for a good disc (in the absence of anomalies) including the probability (77%) of not belonging to a normal class of brake disc, i.e., presence of re j ect ; - figure 8 a shows an example of the output of the Al model for a reject disc (with presence of anomalies) including the probability (100%) of not belonging to a normal class of brake disc, i.e., presence of re ect .

[0032] It is here specified that elements of different embodiments can be combined to provide further embodiments, without restrictions, by respecting the technical concept of the invention, as those skilled in the art will effortlessly understand from the description .

[0033] The present description also makes reference to the prior art for the implementation thereof in relation to the detail features not described, e.g., elements of minor importance usually used in the prior art in solutions of the same type.

[0034] When an element is introduced, it is always understood that there can be "at least one" or "one or more".

[0035] When a list of elements or features is given in this description, it is understood that the finding according to the invention "comprises" or alternatively "consists of" such elements.

[0036] When listing features in the same sentence or bullet list, one or more of the single features can be included in the invention without connection with the other features on the list.

[0037] Two or more of the parts (elements, devices, systems) described above can be freely associated and considered as part kits according to the invention. Embodiments

[0038] As mentioned above , the inspection of cracks ( anomalies in general ) must be carried out with nondestructive technology . The approach of the invention is to use ultrasonic inspection which does not require any coupling fluid, therefore the application is dry . The ultrasonic inspection covers the area between two consecutive teeth of the disc, then the overall inspection of the disc is completed after checking all teeth . The presence of a crack in the inspected area causes changes in the signal emitted by the material . The piece is excited with a laser ( creation of the sound wave within the material ) . The acoustic emission of the piece is detected with an optical microphone . Fig . 3 shows a picture of the operation . The disc is indicated by 100 , the teeth by 110 , the inspected area by 150 , the excitation laser by 200 , the beam by 250 , and the optical microphone by 300 .

[0039] It is worth speci fying here that the sound perceptible by human hearing is detected in the aforesaid patent US 11688415B2 , while inaudible signals (<20 Db ) due to excitation of the material by the laser are detected in the system of the invention, by virtue of the optical microphone .

[0040] The target cycle time for completion of a machine cycle can be 0 . 80 minutes per piece , evaluated as the time between consecutive unloading of 2 pieces , performed with a robot .

[0041] This invention is a complete solution which contains several components , including the ultrasonic bench described above with laser and microphone, an SW consisting of several modules which include the Artificial Intelligence steps described below.

[0042] The automated steps of the Al-based method 500 are as follows, after the start 510, as shown in Fig. 4:

[0043] 1. Ultrasonic signal collection 520: for each step, the control considers laser excitation and acoustic signal acquisition, and an amplitude signal is acquired for each tooth in a disc.

[0044] 2. Signal processing 530: each raw signal is cleaned and standardized in the signal processing step. The length of the single amplitude signal is represented by thousands of points. The signal portion which contains the most information is considered. According to a generally known technique, the signal is normalized by subtracting the signal mean and then dividing it by its standard deviation to reduce the raw signal redundancy.

[0045] 3. Spectrogram generation 540: in the spectrogram generation process, the signal of the processed single tooth is transferred to an image, the spectrogram, by applying a wavelet transform which maintains the scale and position of the signal. This transformation compared with the Short-Term Fourier Transform (STFT) of the prior art known in the field allows obtaining a better resolution in both the time and natural frequency dimensions, precisely because the anomaly is localized in the signal. The disc has n teeth, where n > 10, and the n spectrograms are merged on top of one another to generate a single image which is the input of the next step. An example of a spectrogram representing the entire disc is shown in Figure 5: it is an example of input to the Al model for a disc in the absence of anomalies. It is an image depicting a disc with 18 teeth, where each individual spectrogram is obtained from a wavelet transform for each signal, i.e., each tooth of the disc. 4. Al classification 550: the Al model classifies the disc image, n merged spectrograms, by calculating the probability of anomalies. The chosen model belongs to the Patch Distribution Modeling (PaDiM) framework for detecting anomalies in an image. It uses a pre-trained Convolutional Neural Network (CNN) of patch embedding and multivariate Gaussian distributions to obtain a probabilistic representation of the normal class (piece class without anomalies) . The operation process of PaDiM is shown in Figure 6.

[0046] Anomaly detection is a binary classification between the normal and abnormal classes. It is not possible to train a fully supervised model for this task because abnormal examples are often missing. The selected model is an excellent candidate for industrial applications because the anomalies are represented by a rare event and thus are often not known in advance. Examples of maps of produced by the method described herein and applied to our use case are shown for a good disc and a reject disc in Figure 7 and 8, respectively. Indeed, figure 7 shows an example of the output of the Al model for a good disc (in the absence of anomalies) including the probability (77%) of not belonging to a normal brake disc class, i.e., presence of reject, while figure 8 a shows an example of the output of the Al model for a reject disc (with presence of anomalies) including the probability (100%) of not belonging to a normal class of brake disc, i.e., the presence of reject.

[0047] 5. Output 560. If the probability of not belonging to the normal class is above a predetermined threshold, the output is NOK, otherwise it is considered OK. If the output is OK, the robot unloads the piece into an appropriate container to continue along the line; instead, if the output is NOK, the robot places the disc in the reject section. The process ends in 570 and starts again with a new disc.

[0048] A practical experiment of this invention was made with an initial set of 50 discs without rejects. Since the frequency of breakage is of the order of ppm, anomaly detection algorithms were used in the construction of the input dataset to recognize anomalies never seen before and not known in advance. Synthetic data enrichment was strongly utilized to triple the input data set and achieve good performance. The techniques used were disc rotation, noise addition, and synthetic image generation.

[0049] The machine learning algorithm obtained (and installed on an electronic processing unit or computer) learns the situation of normality of the images, i.e., the absence of defects, and is capable of classifying new images as abnormal or not. It is a very powerful tool because it is resilient to system configuration changes and is also capable of isolating defects never seen or on new types of discs in the absence of historical data availability. This step provided an overall accuracy of 98% and only one false positive out of 50 images, a very good result considering the low number of discs available. These numbers were calculated on a test set consisting of 7 good discs and 3 reject discs exhibiting cracks (or anomalies in general) , repeated 5 times, i.e., 50 collections of disc. The cracks were artificially generated because the natural cracks were not available at the beginning of the manufacturing of a new item.

[0050] According to an aspect of the system of the invention, there is no need for an interface to communicate with the user, because the model output returns the value to the machine which directs the action of the robot in placing the disc in the reject section or continuing with the manufacturing steps. The raw and processed data generated (dental signals, spectrograms, probability score) are saved in a database to be available for continuous monitoring by the quality department. This allows identifying model drifts and requalifying the anomaly detection algorithm.

[0051] Although the invention has been disclosed for cracks in brake discs, it is applicable to the inspection of any mechanical piece made of any material in view of any (mechanical) anomaly, such as materials including one or more of steel, metals, alloys, concrete, wood, composite materials, and other types of materials.

[0052] Preferred embodiments have been described above and variations of the present invention have been suggested, but it should be understood that those skilled in the art may make modi fications and changes without departing from the related scope of protection, as defined by the appended claims .

Claims

CLAIMS1 . An Al-based and computer-implemented method for the quality control of a piece comprising a plurality of portions , comprising the following steps :A. Acquiring ( 520 ) an ultrasonic response signal from a piece subj ected to laser emission for each portion of the piece ;B . Processing ( 530 ) the ultrasonic signal for detecting anomalies for each portion of the piece ;C . Generating ( 540 ) a plurality of spectrograms with a spectrogram for each of said plurality of portions , from the signal processed in step B, and superimposing the plurality of spectrograms into a single spectrogram image ;D . Using ( 530 ) an arti ficial intelligence algorithm, providing said single spectrogram image as input and obtaining a probability of not belonging to a normal piece class as output , the arti ficial intelligence algorithm being trained on a plurality of single spectrogram images related to pieces in said normal piece class ;E . Determining the presence of anomalies i f said probability of not belonging to the normal piece class is higher than a predetermined threshold .2 . A method according to claim 1 , wherein in step B the ultrasonic signal is normalized by subtracting the average of the ultrasonic signal and dividing the result by the standard deviation of the ultrasonic signal .

3. A method according to claim 1 or 2 , wherein in step C the generation of spectrograms is carried out by means of a Wavelet trans form which maintains the scale and position of the ultrasonic signal .4 . A method according to one of claims 1 to 3 , wherein said piece is a brake disc, said plurality of portions are the teeth of the brake disc .5 . A method according to claim 4 , wherein said anomalies are cracks on the brake disc .

6. A method according to one of claims 1 to 5 , wherein said arti ficial intelligence algorithm i s a Convolutional Neural Network, CNN .7 . A method according to one of claims 1 to 6 , wherein the determination of said probability of not belonging to the normal class is obtained by a Patch Distribution Modeling ( PaDiM) .8 . A computer program for the quality control of a piece comprising a plurality of portions , comprising code-based means configured so that , when executed on an electronic processing unit , they perform the steps from B to E of the method according to one of claims 1 to 7 .

9. An apparatus for the quality control of a piece comprising a plurality of portions , comprising anultrasonic emitter, an optical microphone for detecting the acoustic response of a piece on which the ultrasounds are directed, as well as an electronic processing unit , the apparatus being characteri zed in that the computer program according to claim 8 is installed on said electronic unit .

Images