Method for identifying and characterizing surface defects of objects and cracks on fatigue-tested brake discs using artificial intelligence

Abstract

A method for identifying and characterizing surface defects on an object includes: The method comprises the steps of acquiring at least one digital image of an object or part of an object for which surface defects need to be identified, providing the acquired at least one digital image to a training algorithm by means of artificial intelligence and / or machine learning techniques, identifying by the training algorithm one or more surface defects present in the acquired at least one digital image, and generating digital information related to each identified surface defect. The method also determines, for each identified surface defect, at least one respective dimensional parameter representative of at least one dimension of the surface defect and at least one respective location parameter representative of the location of the surface defect relative to a reference point or line present in the image or a two-dimensional spatial coordinate system associated with said reference point or line. The determining step is performed by further processing of the digital information by electronic processing means.

Description

[Technical field]

[0001] The present invention relates to a method for identifying and characterizing surface defects on an object using artificial intelligence (AI).

[0002] More particularly, the present invention further relates to an artificial intelligence (AI) based method for identifying and characterizing cracks on brake discs. [Background technology]

[0003] The use of artificial intelligence (AI) and computer vision (CV) techniques for the detection of surface defects or cracks and their dimensional quantification is now well established. Such tools involve the analysis of images taken by an operator or a robot by means of appropriately designed algorithms (usually one or more neural networks), in which surface defects of different sizes and severity may be present. The main application field concerns the monitoring of industrial public infrastructure and / or products operating in harsh environments (nuclear reactors, underwater structures, etc.).

[0004] Prior to the advent of these methods, images were reviewed by human operators, but this procedure is very costly in terms of resources. Therefore, given the availability of images, it is a no-brainer to automate this process using well-established artificial intelligence and computer vision techniques applicable to the detection of objects in images.

[0005] However, in order to apply such techniques on a large scale, several problems remain unsolved, i.e., several needs are in any case not fully met.

[0006] Firstly, a need is felt to identify the location of surface defects not only in an absolute spatial coordinate system but also in a spatial coordinate system that is linked to the relevant parts of the object being inspected and that are present in the image being analysed.

[0007] Secondly, in many applications, for example in mechanical parts that operate dynamically or are subjected to dynamic or fatigue testing, there is also a need to not only identify but also monitor the appearance and development of surface defects under dynamic conditions.

[0008] The above requirements are not met by known solutions.

[0009] An important and typical application is the need to identify and monitor cracks in brake discs.

[0010] The prior art has not attempted to harness the potential of artificial intelligence or machine learning (ML) techniques and algorithms to identify cracks in brake discs.

[0011] According to currently used procedures, to determine the resistance of a braking system to thermomechanical stresses, the braking system is tested on a dynamic test bench, in which a predefined braking sequence is applied in terms of operating parameters (rotational speed, brake pressure / torque, temperature). The test protocol provides for the bench to be stopped at predefined time intervals and for the stationary discs to be visually inspected by an operator.

[0012] If cracks are identified on both sides of the braking surface of the disc, the length of the longest crack on each side is measured with a vernier caliper and recorded. If, during a stop, the operator detects a crack whose length exceeds a certain threshold, expressed as a percentage of the radial extension of the braking surface, or which is too close to the outer or inner edge of the braking surface, the test is immediately aborted.

[0013] Fatigue tests carried out in this way are also very expensive in terms of resources due to the long test periods (especially the longer ones, which can last several weeks). Indeed, in addition to the long machine runs, they require the constant presence of an operator to carry out manual measurements of the cracks during the downtimes prescribed by the protocol. Moreover, the regular downtimes cause further costs. Indeed, it is necessary to periodically stop the bench, wait for the disks to cool down and allow the operator to access them.

[0014] Furthermore, measurements made by an operator are not always reliable and accurate, introducing an additional source of error into the analysis of the behavior of the test part.

[0015] Finally, this method of conducting fatigue tests does not allow extracting all available information from the experiment. Indeed, it would be interesting to know the length, radial and angular position of all cracks identified on the disc, in addition to cyclic information on the length of the longest crack present on the braking surface. Furthermore, it would be interesting to collect data on the evolution of these quantities over time and on the number of cracks present. If these data are available, the behavior of the tested product can be investigated in more detail. Summary of the Invention

[0016] The object of the present invention is to provide a method for identifying and characterizing surface defects on an object by using artificial intelligence, thereby at least partially avoiding the drawbacks mentioned above with reference to the prior art and meeting the aforementioned needs, which are particularly felt in the technical field considered.

[0017] This object is achieved by the method according to claim 1.

[0018] Further embodiments of such a method are defined in claims 2-13 and 27.

[0019] A further object of the present invention is to provide a method for identifying and characterizing cracks on brake discs by use of artificial intelligence. Such object is achieved by a method as claimed in claim 14.

[0020] Further embodiments of the method are defined in claims 15-26.

[0021] Related to this objective, another objective of the present invention is to exploit the potential of AI combined with classical CV techniques to automate fatigue testing of brake discs. More specifically, to automate the identification and quantification of cracks that develop in brake discs during testing in order to make the experiment more efficient in terms of the resources used and to maximize the amount of information extracted. Moreover, the automation of the process meets the need to make the results obtained more reliable, reproducible and objective.

[0022] Another object of the present invention is to provide a method for performing fatigue testing of mechanical components, in particular brake discs, employing the above-mentioned method for identifying and characterizing surface defects and cracks. Such object is achieved by the methods according to claims 28 and 29, respectively. [Brief description of the drawings]

[0023] Further characteristics and advantages of the method according to the invention will become apparent from the following description of preferred exemplary embodiments, given in a non-limiting manner with reference to the attached drawings, in which:

[0024] [Figure 1] FIG. 1 is a block diagram illustrating one embodiment of the method according to the present invention.

[0025] [Diagram 2] FIG. 2 shows an experimental setup for carrying out fatigue tests on brake discs with which the implementation of the method according to the invention is associated.

[0026] [Diagram 3]FIG. 3 is a simplified block diagram illustrating some of the steps involved in one embodiment of the method.

[0027] [Figure 4] FIG. 4 shows a brake disc with known cracks labelled according to steps of a method according to an embodiment of the invention.

[0028] [Diagram 5] FIG. 5 shows example images provided as input to the machine learning algorithm during the learning step according to an embodiment of the method of the present invention.

[0029] [Figure 6] FIG. 6 shows an example of an image obtained at output from a machine learning algorithm according to an embodiment of the method of the present invention.

[0030] [Figure 7] FIG. 7 shows the geometric parameters and optical diagram of a pinhole camera employed in an embodiment of the method of the present invention.

[0031] [Figure 8] FIG. 8 shows an exemplary arrangement of a portion of a brake disc that allows a reference coordinate system to be associated with an image of the brake disc.

[0032] [Figure 9] Figure 9 shows the precision-recall plot.

[0033] [Figure 10] FIG. 10 is a simplified block diagram of a system in which the method according to the present invention can be implemented. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0034] A method for identifying and characterizing surface defects on an object is described.

[0035] The method comprises the steps of acquiring at least one digital image of an object or part of an object in which surface defects have to be identified, then providing the acquired at least one digital image to an algorithm trained by artificial intelligence and / or machine learning techniques, then identifying by the training algorithm one or more surface defects present in the acquired at least one digital image, and generating digital information associated with each identified surface defect.

[0036] The method then provides for determining, for each identified surface defect, at least one respective dimensional parameter representative of at least one dimension of the surface defect and at least one respective positional parameter representative of the position of the surface defect relative to a reference point or line present in the image or a two-dimensional spatial coordinate system associated with said reference point or line.

[0037] The aforementioned determining step is carried out by further processing of said digital information by electronic processing means.

[0038] According to one embodiment, the method is configured to identify and characterize surface defects on a mechanical component under dynamic conditions.

[0039] In this case, said acquiring step consists of acquiring in sequence a plurality of digital images of the mechanical part, acquired during the dynamic evolution of the movement of the mechanical part.

[0040] To monitor the dynamic evolution of the presence, size and location of surface defects, the aforementioned providing, identifying, generating and determining steps are performed continuously on successively acquired digital images.

[0041] According to an implementation option, said dynamic conditions consist of fatigue tests of the mechanical components.

[0042] In this case, the method further comprises a step of establishing surface defect evaluation criteria for deciding whether to continue or stop the fatigue test, and further a step of continuously comparing information related to the temporal evolution of the surface defects with the established evaluation criteria, and then continuing the fatigue test if all the surface defect evaluation criteria are met, or alternatively, stopping the fatigue test if at least one of the evaluation criteria is not met.

[0043] According to one embodiment of the method, said training algorithm is an algorithm trained by a pre-training step based on a training data set consisting of digital training images, which represent objects of the same type as those whose surface defects have to be identified and characterized, provided as input to the training algorithm, said objects having surface defects whose respective dimensional parameters and respective positional parameters are known, which are also provided as input to the training algorithm.

[0044] According to an implementation option of said embodiment, said pre-training step starts with a pre-training algorithm based on a pre-training dataset different from said training dataset by applying transfer learning techniques to arrive at a training algorithm.

[0045] In such implementation options, transfer learning (TL) techniques are used to build machine learning (ML) algorithms, i.e., a pre-trained algorithm is selected on another dataset. Transfer learning is a technique in machine learning that offers to apply knowledge gained while solving a similar problem to solve a problem.

[0046] Reusing or transferring information from previously learned tasks to learn new tasks can significantly improve the recognition performance of machine learning algorithms (especially deep learning algorithms).

[0047] According to an implementation option of the aforementioned embodiment, the pre-learning step includes tagging or labeling known surface defects present in each of the digital training images, and then calibrating parameters of the training algorithm based on the processed digital training images by tagging or labeling.

[0048] According to a possible implementation option, the aforementioned tagging or labeling step is performed by highlighting obvious surface defects on digital training images, manually and / or with facilitating software support.

[0049] According to an embodiment, the method further comprises the step of validating the predictive capability of the trained algorithm on a further dataset of digital validation images.

[0050] According to an embodiment of the method, the training algorithm is a machine learning algorithm based on a neural network.

[0051] According to various embodiments, the neural network comprises a deep neural network, or a convolutional neural network, or a region-based convolutional neural network.

[0052] According to another embodiment, the trained algorithm is a machine learning algorithm based on a deep object detector or a two-stage deep object detector.

[0053] According to one embodiment of the method, the aforementioned step of identifying one or more surface defects present in at least one acquired digital image comprises recognizing the surface defects by a training algorithm and, for each recognized surface defect, identifying the spatial coordinates of the surface defect with respect to a reference coordinate system of the acquired digital image, to which portions of the depicted object are also referenced in a known manner.

[0054] The aforementioned step of generating information relating to each surface defect comprises generating, for each identified surface defect, digital information representative of the aforementioned spatial coordinates of the surface defect, and storing such digital information so as to make it available for subsequent processing.

[0055] According to one embodiment, the determining step comprises determining, for each surface defect, respective dimensional and positional parameters based on spatial coordinates of the surface defect.

[0056] According to an embodiment, the method further comprises, prior to the acquiring step, a step of performing a calibration of the image acquisition means and, following the calibration, a step of acquiring data to correct the effects of geometric distortion in the image acquisition.

[0057] According to various possible applications, the method may be employed to detect surface defects on various possible surfaces, for example smooth, rough, spongy or other surfaces.

[0058] According to various embodiments, the method is applied to identify and characterize surface defects on wood and / or plastic and / or fabric objects.

[0059] According to another embodiment, the method is applied to identify and characterize surface defects on objects of glassy, ​​ceramic, cement, and metallic materials.

[0060] It should be noted that the method presented above, by its nature, can be applied to a wide range of objects made of different materials with respect to those described above.

[0061] It should also be noted that in some possible applications, the method is used to detect various types of surface defects, such as cracks, holes, tears, scratches, chips, stains, etc.

[0062] In fact, due to its characteristics, the method presented above can be applied to a wide range of surface defects, which can generally be defined as any non-uniformity that can be captured by an image with respect to the background, e.g., any non-uniformity that the human eye can somehow perceive with respect to a uniform background.

[0063] According to various embodiments, the step of acquiring a digital image is performed by known image acquisition means, such as a camera, video camera, or other image acquisition device in the visible spectrum.

[0064] In a preferred embodiment, the method, carried out according to any one of the embodiments presented above, is used in the field of detecting and monitoring cracks on brake discs, more particularly on braking surfaces or elements of brake discs.

[0065] Such an embodiment is described in more detail below with reference to Figures 1-11.

[0066] According to such an embodiment, the method is configured to identify and characterize cracks on a braking surface or element of a brake disc, in such a case, said object is a brake disc and said defect is a crack in the brake disc.

[0067] In such a case, the aforementioned acquisition step consists of acquiring at least one digital image of the braking surface or element of the brake disc, said set of at least one digital image being representative of the entire annulus corresponding to said braking surface or element.

[0068] The providing step comprises providing at least one acquired digital image to an algorithm trained by artificial intelligence and / or machine learning techniques.

[0069] The identifying step comprises identifying, by the training algorithm, one or more cracks present in the at least one acquired digital image and generating digital information associated with each identified crack.

[0070] The aforementioned dimensional parameter in this case consists of the crack length (i.e. the extended dimension of the crack since the crack is primarily a one-dimensional defect).

[0071] The aforementioned position parameters consist of the position of the crack relative to the edge and / or braking surface of the brake disc, so that the determination step consists of determining, by further processing as described above, for each crack identified, a respective length and a respective position parameter representative of the position of the crack relative to the edge and / or braking surface of the brake disc.

[0072] According to an implementation option of such an embodiment, the method is configured to identify and characterize cracks on a braking surface or element of a brake disc under dynamic conditions.

[0073] In this case, the acquisition step consists of sequentially acquiring a plurality of digital images of the braking surface or elements of the brake disc, acquired during the dynamic evolution of the operation of the brake disc, and the providing, identifying, generating and determining steps are performed successively on the sequentially acquired digital images in order to monitor the dynamic evolution of the presence, length and location of cracks.

[0074] According to an application example of the method in which the aforementioned dynamic conditions consist of a brake disc fatigue test, the method further comprises a step of establishing crack evaluation criteria for deciding whether to continue or stop the fatigue test, and further a step of continuously comparing information related to the temporal evolution of the cracks with the established evaluation criteria, whereafter the method consists in continuing the fatigue test if all crack evaluation criteria are met and instead stopping the fatigue test if at least one of the evaluation criteria is not met.

[0075] According to various possible implementation options of this embodiment, said evaluation criteria consist of one or more of the following criteria:

[0076] the length of each crack is less than a predefined maximum length and the continuation of the fatigue test is no longer considered acceptable; and / or

[0077] The edges of all cracks are further from the edge of the braking surface or brake disc than a predefined minimum distance which is no longer acceptable for continuing the fatigue test.

[0078] According to an implementation option of this embodiment, the learned algorithm is an algorithm that is trained by a pre-learning step based on a learning data set consisting of digital images of braking surfaces with known cracks, together with input information related to the size and location of the known cracks, which are provided as input to the training algorithm.

[0079] With regard to the machine learning or artificial intelligence algorithms used in the method for identifying and characterizing cracks on a brake disc, all of the implementation options already presented with respect to the more general method for identifying and characterizing surface defects can be used.

[0080] According to the already mentioned implementation option of the aforementioned embodiment, said pre-training step operates starting from a pre-training algorithm based on a pre-training dataset different from said training dataset by applying transfer learning techniques in order to arrive at a training algorithm.

[0081] In other words, to build a machine learning (ML) algorithm, a transfer learning (TL) technique is used, i.e. an algorithm is selected that is pre-trained on another dataset. Transfer learning, in the context of machine learning, is a technique that provides for the application of knowledge gained during the solution of a similar problem to solve a problem.

[0082] According to the specific implementation experimented with, the crack recognition algorithm was created by transfer learning, where the Mask-RCNN algorithm was trained on a huge dataset (more than 200,000 examples) of images and annotations of common objects in the relevant context (e.g. cars, people, aircraft), and then, through a more specific training process, it was taught to recognize cracks on brake discs.

[0083] By applying the aforementioned transfer learning techniques, much better algorithmic performance can be achieved with the same number of images used for training (with respect to traditional deep learning algorithms trained to recognize cracks, starting from "scratch").

[0084] According to an implementation option, the aforementioned pre-training step consists of tagging or labelling the known cracks present in each of the digital training images, and then calibrating the parameters of the training algorithm based on the processed digital training images by tagging or labelling.

[0085] According to an embodiment, the aforementioned tagging or labeling step is performed by drawing lines on the digital training images, manually and / or with the assistance of facilitating software, tracing the spatial trends of each apparent crack.

[0086] According to one embodiment, the "labelMe" tool is used.

[0087] According to a possible mode of operation, such a tool generates an “accompanying” file whose information content specifies where the cracks are located in the image, for example by reporting a list of the pixel-by-pixel coordinates of the endpoints of all cracks present in the image.

[0088] According to an implementation option, the aforementioned step of identifying one or more cracks present in at least one acquired digital image comprises recognizing the cracks by means of a learned algorithm and, for each recognized crack, identifying the spatial coordinates of the ends of the crack, approximated as a segment, with respect to a reference coordinate system of the acquired digital image, to which a depicted portion of the brake disc or brake surface is also referenced in a known manner.

[0089] In this case, the aforementioned step of generating information relating to each crack consists of generating, for each identified crack, digital information representative of the spatial coordinates of the crack and storing such digital information to make it available for subsequent processing.

[0090] According to an implementation option, the aforementioned step of generating information further comprises generating at least one processed digital image each including highlights and / or indications associated with one or more identified cracks.

[0091] According to one embodiment, the aforementioned step of determining, for each identified crack, at least one respective parameter representative of length and crack location is performed by an untrained image processing algorithm.

[0092] According to a particular implementation option, the aforementioned step of determining, for each identified crack, at least one parameter representative of the length and the crack position is performed by an untrained Computer Vision (CV) algorithm.

[0093] According to one embodiment, the aforementioned steps of determining, for each identified crack, at least one respective parameter representative of the length and crack location are performed by a further machine learning algorithm.

[0094] According to another implementation option, the aforementioned step of determining, for each identified crack, at least one respective parameter representative of the length and crack location is performed by the same trained machine learning (ML) algorithm configured to perform the aforementioned step of identifying one or more cracks.

[0095] In the latter case, a single ML algorithm performs all steps of the method, from the image to the crack length and / or location.

[0096] In this case, according to an embodiment variant included in the present invention, an end-to-end deep learning algorithm is used, which generates the length of the crack directly from the image and its position relative to a real reference system (e.g. the edge of the disk) always contained in the image, without going through the determination of image coordinates.

[0097] According to implementation options, the aforementioned step of determining, for each identified crack, at least one respective parameter representative of the length and crack location, comprises the following steps:

[0098] Calculating the length of the crack based on the coordinates of each end.

[0099] Calculating each of the at least one parameter representative of a crack position as a distance from the edge of the end of the crack closest to the edge, based on the coordinates of the end and the coordinates of the edge, with respect to the reference system.

[0100] According to another embodiment, said step of calculating a parameter representative of the crack position consists of calculating the radial and angular position of the crack on the brake disc.

[0101] Next, a method for performing a fatigue test on a mechanical component will be described.

[0102] The method comprises carrying out, during the performance of a fatigue test, a method for identifying and characterizing surface defects according to any one of the previously described embodiments.

[0103] The method thus includes continuing the fatigue test if all crack evaluation criteria of a predefined set of evaluation criteria are met, and alternatively, stopping the fatigue test if at least one of the evaluation criteria is not met.

[0104] Next, a method for carrying out a fatigue test on a brake disc will be described.

[0105] The method comprises carrying out a method for identifying and characterising cracks on a brake disc according to any one of the previously described embodiments during the performance of a fatigue test.

[0106] The method then provides for proceeding with the fatigue test if all crack evaluation criteria of a predefined set of evaluation criteria are met, and alternatively, stopping the fatigue test if at least one of the evaluation criteria is not met.

[0107] The predefined evaluation criteria include, for example:

[0108] The length of each crack is less than a predefined maximum length and the continuation of the fatigue test is no longer considered acceptable; and / or

[0109] The ends of all cracks are located at a predefined minimum distance from the edge of the braking surface or brake disc below which they are no longer considered acceptable for the continuation of the fatigue test.

[0110] Purely by way of non-limiting example, further details of a method according to a particular embodiment of the invention focusing on a method for identifying and characterizing cracks on the surface of a brake disc subjected to fatigue testing are reported below with reference to Figures 1-10.

[0111] The logical flow of this method is shown in Figure 1.

[0112] Prior to the start of the fatigue test on the dynamometric bench, an experimental setup is installed on the dynamometric bench. The experimental setup is capable of periodically acquiring still images of different portions of the braking surface over the entire duration of the test. In this example, the portions of the disc that are photographed are such that it is possible to periodically acquire information relating to the entire annular portion of the braking surface over the entire duration of the test.

[0113] According to implementation options, acquisition is performed simultaneously on both sides of the disc.

[0114] In the implementation option, the aforementioned system (experimental rig) mounted on the test bench consists of two metal supports. Each support contains a camera appropriately selected to fit the dimensions of the bench brake system and to have the widest possible operating temperature range. The arms are mounted at a pre-defined distance from the surface of the disc, allowing the frame to be focused (see Figure 2). The optical axis of each camera must reach the disc in a direction perpendicular to the disc surface.

[0115] According to the implementation options, the Dynamic Bench software manages the image acquisition system, which manages the angular position of the brake disc, the lighting, the image acquisition time, acquires and stores the acquired images. Once the image acquisition is finished, the bench goes into a paused state to wait for the results of the image processing.

[0116] The images thus captured can then be input into machine learning (ML) models or algorithms to identify possible cracks.

[0117] According to the implementation option, the transfer learning method was used to build the ML algorithm, i.e. an algorithm was selected that is pre-trained on another dataset.

[0118] In the example presented here, the Mask-RCNN model, which is based on a neural network (NN) trained on the COCO open source dataset, was chosen.

[0119] Typically, the development flow for an ML algorithm is input preparation, tagging, and model training (see Figure 3).

[0120] No input preparation step is performed since the algorithm used takes as input images directly acquired by the bench camera, which is advantageous in terms of computational load and therefore also time, which is a key factor since the algorithm is designed to run online on the bench.

[0121] The tagging process involves manually labeling cracks depicted in images taken on the bench during testing. In particular, this process consists of drawing lines on the images that trace the spatial trend of each apparent crack. In the rare cases where a crack resembles a broken line, it is tagged as a segment connecting the crack's endpoints.

[0122] Accurate tagging is a prerequisite for deep learning algorithms to work well.

[0123] According to the implementation options included in the present invention, the tools used to support the tagging activity are taken from an open source tool (labelMe). An example of an image tagged with labelMe is shown in FIG.

[0124] The tagging step is followed by a traditional training process: a subset of the tagged dataset (consisting of 101 image files) is provided as input to the AI ​​algorithm to calibrate the model's parameters and adapt it to provide predictions. An example of the tagged input used for model training is shown in Figure 5.

[0125] According to a particular implementation option, the aforementioned subset of the tagged dataset is enriched by data augmentation techniques.

[0126] Once an algorithm is trained, its predictive capabilities are validated on other data sets of the same nature.

[0127] Once the algorithm detects a crack on the input image with confidence above a pre-set threshold, the geometric coordinates of its start and end points are saved.

[0128] According to an implementation option, the coordinate system is the image coordinate system.

[0129] The cracks are considered as segments, which is a valid approximation in almost all cases. This data can be displayed in graphical form on the starting image (see Figure 6).

[0130] The next step of the method involves applying classical CV techniques to process the information related to each identified crack and reliably calculate its length without geometric distortion. In fact, each image acquired through a camera has a certain degree of distortion as a function of how the instrument was calibrated. This means that lengths of equal value on the acquired image do not necessarily correspond to equal lengths in reality.

[0131] In the present invention, the camera is calibrated once during the preparation of the experimental setup using a physical reference. This calibration allows the calculation of the camera's inherent distortion parameters. On the basis of these parameters, it is possible to correct the phenomenon using concatenation tools, such as the application of a matrix camera. After this processing, the distances measured on the image are proportional to the real distances according to a certain coefficient.

[0132] Once the distortion has been corrected, the length of each crack in the image can be determined from the coordinates of its endpoints in any unit. The formula is that of a segment in a Euclidean plane. By comparing the calculated lengths of all cracks seen in the image, the longest crack can be determined.

[0133] If this comparison is extended to all cracks present on the two braking surfaces of the disc, captured in multiple images taken within a sufficiently short time, it can be determined which crack is the longest.

[0134] Conversion of crack length values ​​from any unit to mm can be easily done by applying the pinhole camera model (model shown in Figure 7).

[0135] Referring to FIG. 7, the relationships between different image acquisition parameters are as follows:

[0136] H=(d / f)(S / R)n

[0137] where H (mm) is the length of the identified pattern (e.g., the length of a crack) represented by n pixels in the image, d (mm) is the working distance (camera-object distance), f (mm) is the focal length of the camera, S (mm) is the size of the camera sensor, and R (pixels) is the resolution of the camera sensor.

[0138] If the length of the longest crack identified above exceeds a threshold declared by the operator and made available to the algorithm prior to the start of the test, the test is automatically aborted.

[0139] The second criterion for whether the test can continue or not is the minimum safe distance between the crack and the outer edge of the braking surface: the test is therefore stopped if at least one crack appears in the outer strip of the braking surface, even if the entire strip is not cracked.

[0140] This requires knowing the position of the outer edge, described by the equation of an ellipse, for each angular position of the braking surface constructed in the camera frame.

[0141] To determine such an equation, proceed as follows: Before starting the test, on a portion of the band that is fully framed within the frame, trace three rays on the ring with a marker of a color that stands out against the background (see Figure 8). The coordinates of the three intersections of these rays with the circumference of the ring are used to calculate the desired equation. This equation is then subjected to a mathematical correction transformation in the same way as the coordinates of the crack ends.

[0142] From the transformed equation it is possible to determine the location of the periphery of the braking surface portion bordered at any desired angle value.

[0143] If the peripheral edge of at least one crack is located at a distance from the edge that is small relative to a threshold value (which can be expressed in pixels or mm), the test is interrupted.

[0144] From the start of the fatigue test, the algorithm runs periodically, inspecting all images necessary to cover both sides of the braking surface of the tested disc. If at least one of the criteria that triggers a test stop is met, the test is automatically interrupted and a notification is sent to the operator.

[0145] The method described herein allows to periodically collect a wealth of information related to the growing cracks during fatigue testing (number at a given time instant, length of each crack, location relative to the periphery) From such a data set it is possible to reconstruct the time evolution of the mechanical response of the product to the fatigue phenomenon.

[0146] In terms of performance, the accuracy of the algorithm as a function of the recall variable is shown in FIG.

[0147] The metrics depicted concern the model's performance on the test dataset, i.e. the subset of data that was not used to train the AI ​​model, with the IoU (intersection on union) parameter set equal to 0.5.

[0148] In the literature, "accuracy" means how many true positives (i.e., cracks identified by the model that are actually cracks) there are relative to the sum of true positives and false positives (cracks that are incorrectly identified as such by the model).

[0149] Instead, the recall variable quantifies the true positives over the sum of true positives + false negatives (i.e., cracks that are actually present that were not tagged as such by the model). The mean average precision (mAP) of this model was 0.85.

[0150] One embodiment of a system capable of implementing the above-described method according to the present invention is shown in FIG.

[0151] The components of the system shown in FIG. 10 are as follows:

[0152] An AI Server (using one or more electronic processors or computers). The AI ​​Server contains one or more software modules capable of implementing the AI ​​models used, or the machine learning algorithms used (the "AI Inference" block), and possibly further services.

[0153] Centralized Electronic Archive: A centralized electronic archive stores much of the archived data resulting from performance of the method, such as images of cracks, results of crack detection, summary reports of cracks present, etc.

[0154] At least one lab bench, which includes at least one electronic processor or computer (which in the example architecture shown in FIG. 10 is a slave computer) in addition to a lab bench I / O interface, capable of receiving, processing, and providing digital data such as crack images, crack detection results, and crack summary reports.

[0155] According to an implementation option, at least one electronic processor or computer present on the laboratory bench is configured to carry out (by means of one or more specific elements) a step of determining at least one dimensional parameter and at least one respective positional parameter, by implementing an algorithm (which may be untrained) developed for this purpose, for example a Computer Vision (CV) algorithm, this algorithm being executed by at least one software module loaded and executable on the computer itself.

[0156] According to the implementation options described above (and illustrated in FIG. 10), the method is implemented by the synergistic cooperation of two algorithms: an algorithm trained by AI or ML techniques (for crack recognition) and loadable / executable on a server computer, and another untrained computer vision algorithm (for characterizing the dimensions and location of identified cracks) and loadable / executable on a bench computer.

[0157] Apparently, the two computers are operatively connected to each other.

[0158] According to another implementation option, both crack recognition and characterization are performed by a single computer, e.g. the control computer of the laboratory bench (embedded solution), on which software modules implementing both ML and CV algorithms are present and executable.

[0159] According to another implementation option, the functions of the method are performed by a system implemented in a cloud and / or serverless architecture.

[0160] Thus, the object of the invention as set out above is fully achieved by the method described above, thanks to the features disclosed in detail above. The advantages and the technical problems solved by the method according to the invention have already been described above with reference to different features and aspects of the method.

[0161] Those skilled in the art can make modifications and adaptations to the above-described method embodiments or replace them with other functionally equivalent elements to meet the unexpected needs without departing from the scope of protection of the appended claims. All features described above as belonging to one possible embodiment can be implemented independently of the other described embodiments.

Claims

1. 1. A method for identifying and characterizing surface defects on an object, comprising: acquiring at least one digital image of the object or a portion of the object in which the surface defects are identified; providing the acquired at least one digital image to a training algorithm using artificial intelligence and / or machine learning techniques; identifying, by the training algorithm, one or more surface defects present in the at least one acquired digital image; generating digital information associated with each of the identified surface defects; determining, for each of the identified surface defects, at least one dimensional parameter representative of at least one dimension of the surface defect and at least one positional parameter representative of a position of the surface defect relative to a reference point or line present in the digital image or a two-dimensional spatial coordinate system associated with the reference point or line; said determining step being carried out by further processing of said digital information by electronic processing means; the training algorithm is an algorithm that is trained by a pre-training step on a training dataset that includes digital training images that are provided as input to the training algorithm; the training data set represents objects of the same type as the object in which the surface defects are to be identified and characterized; The method, wherein the object has surface defects with known dimensional and location parameters, the surface defects also being provided as input to the training algorithm.

2. 2. The method of claim 1, wherein the pre-training step starts with a pre-training algorithm based on a pre-training dataset different from the training dataset by applying transfer learning techniques to arrive at the training algorithm.

3. configured to identify and characterize the surface defects on the mechanical component under dynamic conditions; the acquiring step includes sequentially acquiring a plurality of the digital images of the mechanical part acquired during a dynamic evolution of a motion of the mechanical part; 2. The method of claim 1, wherein the providing, identifying, generating, and determining steps are performed continuously on the continuously acquired digital images to monitor the dynamic evolution of the presence, size, and location of the surface defects.

4. the dynamic conditions include a fatigue test of the mechanical component; The method further comprises: establishing criteria for evaluating the surface defects configured to determine whether to continue or discontinue the fatigue test; continuously comparing information relating to the temporal evolution of the surface defects with the evaluation criteria; continuing the fatigue test if all of the surface defect criteria are met; The method of claim 3 , comprising the step of aborting the fatigue test if at least one of the evaluation criteria is not met.

5. The pre-training step includes: performing tagging or labeling of the known surface defects present in the digital training images; The method of claim 4 , further comprising the step of calibrating parameters of the training algorithm based on the digital training images processed by the tagging or labeling.

6. 6. The method of claim 5, wherein the step of tagging or labeling is performed on the digital training images manually and / or with software assistance by highlighting the apparent surface defects.

7. 5. The method of claim 4, further comprising validating the predictive ability of the training algorithm on a further dataset of digital validation images.

8. The method of claim 4 , wherein the training algorithm is a machine learning algorithm based on a neural network.

9. The method of claim 8 , wherein the neural network comprises a deep neural network, or a convolutional neural network, or a region-based convolutional neural network.

10. The method of claim 4 , wherein the training algorithm is a machine learning algorithm based on a deep object detector or a two-stage deep object detector.

11. the identifying step of identifying the one or more surface defects present in at least one acquired digital image includes recognizing the surface defects by the training algorithm; and for the recognized surface defects, identifying spatial coordinates of the surface defects with respect to a reference coordinate system of the acquired digital image; The reference coordinate system also references the depicted portion of the object in a known manner; 2. The method of claim 1, wherein the generating step of generating information related to the surface defects comprises: generating, for the identified surface defects, the digital information representative of the spatial coordinates of the surface defects; and storing the digital information so that it is available for subsequent processing.

12. The method of claim 11 , wherein the determining step comprises determining, for each of the surface defects, the dimensional parameters and the positional parameters based on the spatial coordinates of the surface defects.

13. before said acquiring step, performing a calibration of the means for acquiring said digital images; 2. The method of claim 1, further comprising, following said calibration, acquiring data to correct for the effects of geometric distortion in said digital image.

14. 1. A method configured to identify and characterize cracks on a braking surface or element of a brake disc under dynamic conditions, comprising: the object is a brake disc, the surface defect is a crack in the brake disc; the acquiring step includes acquiring at least one digital image of the braking surface or element of the brake disc; the set of at least one digital image represents an entire annulus corresponding to the braking surface or element; the providing step includes providing the acquired at least one digital image to the training algorithm by artificial intelligence and / or machine learning techniques; the identifying step includes identifying, by the training algorithm, one or more cracks present in the at least one acquired digital image and generating digital information associated with each of the identified cracks; The dimensional parameters include a length of the crack; the location parameters include the location of the crack relative to an edge of the brake disc and / or braking surface; A method according to any one of claims 1 to 13, wherein the determining step comprises determining, by further processing, for each of the identified cracks, the length and position parameters representative of the position of the crack relative to the edge of the brake disc and / or braking surface.

15. 1. A method configured to identify and characterize said cracks on said braking surface or element of said brake disc under dynamic conditions, comprising: the acquiring step comprises sequentially acquiring a plurality of said digital images of said braking surface or element of said brake disc acquired during a dynamic evolution of the operation of said brake disc; 15. The method of claim 14, wherein the providing, identifying, generating, and determining steps are performed continuously on continuously acquired digital images to monitor the dynamic evolution of the presence, length, and location of the crack.

16. the dynamic conditions include a fatigue test of the brake disc; establishing criteria for evaluating the crack to determine whether to continue or discontinue the fatigue test; continuously comparing information relating to the crack evolution over time with the evaluation criteria; if all of the crack evaluation criteria are met, continuing the fatigue test; 16. The method of claim 15, comprising the step of aborting the fatigue test if at least one of the evaluation criteria is not met.

17. The evaluation criteria are: The crack length is less than a predefined maximum length that is no longer considered acceptable for the continuation of the fatigue test; and / or 17. The method of claim 16, wherein the ends of all the cracks are more than a predefined minimum distance away from the edge of the braking surface or brake disc that is no longer acceptable for continuing the fatigue test.

18. 18. The method of claim 17, wherein the training algorithm is an algorithm that is trained by a pre-training step based on a training data set consisting of digital images of the braking surface having known cracks provided as input to the training algorithm along with input information related to the size and location of the known cracks.

19. The pre-training step includes: performing tagging or labeling of known cracks present in each of said digital training images; 20. The method of claim 18, comprising calibrating parameters of the training algorithm based on the digital training images processed by the tagging or labeling.

20. 20. The method of claim 19, wherein the tagging or labeling step is performed by manually and / or with software assistance drawing lines on the digital training images that trace spatial trends of apparent cracks.

21. The step of identifying one or more cracks present in the at least one acquired digital image comprises: recognizing the crack by the training algorithm; and for the recognized crack, identifying the spatial coordinates of the ends of the crack approximated as a segment with respect to a reference coordinate system of the acquired digital image; The reference frame also references a depicted portion of the brake disc or braking surface in a known manner, 15. The method of claim 14, wherein the step of generating information associated with each of the cracks includes generating, for each identified crack, digital information representative of the spatial coordinates of the crack, and storing the digital information to make it available for subsequent processing.

22. 22. The method of claim 21, wherein the step of generating information further comprises generating at least one respective processed digital image including highlights and / or indications associated with the one or more identified cracks.

23. 14. The method of claim 1, wherein for each of the identified cracks, the step of determining the length of the crack and at least one respective parameter representative of the position of the crack is performed by an untrained computer vision algorithm.

24. the step of determining, for each of the identified cracks, the length of the crack and at least one parameter representative of the crack location is performed by a further trained machine learning algorithm; or The method of claim 14 , wherein the method is performed using the same trained machine learning algorithm configured to perform the step of identifying one or more cracks.

25. The step of determining, for each of the identified cracks, a length of the crack and at least one parameter representative of the position of the crack comprises: calculating a length of the crack based on the coordinates of the ends of the crack; 22. The method of claim 21, comprising calculating the at least one respective parameter representative of the position of the crack as a distance to the crack edge closest to the crack edge based on the coordinates of the crack end and the coordinates of the crack edge relative to the reference coordinate system.

26. 15. The method of claim 14, wherein the step of calculating the parameter representative of the position of the crack comprises calculating the radial and / or angular position of the crack on the brake disc.

27. 14. A method according to any one of claims 1 to 13, operating on objects of wood and / or plastic and / or textile and / or vitreous and / or ceramic and / or cementitious and / or metallic material.

28. 1. A method for performing a fatigue test on a mechanical component, comprising: - during a fatigue test, carrying out a method for identifying and characterizing surface defects according to any one of claims 1 to 13; continuing the fatigue test if all crack criteria of a predefined set of criteria are met; Aborting the fatigue test if at least one of the evaluation criteria is not met.

29. 1. A method for conducting fatigue testing of a brake disc, comprising: During the fatigue test, performing the method for identifying and characterizing cracks according to claim 14; continuing the fatigue test if all crack criteria of a predefined set of criteria are met; aborting the fatigue test if at least one of the evaluation criteria is not met; The predefined evaluation criteria are: the length of the crack is less than a predefined maximum length and the continuation of the fatigue test is no longer considered acceptable; and / or The ends of all the cracks are at least a predefined minimum distance from the edge of the braking surface or brake disc at which point the cracks are deemed no longer acceptable for the continuation of the fatigue test.

Images