Method and system for real-time detection of anomality of object surface to durability test by using

Abstract

A method (100) for real-time detection of anomalies of an object undergoing a durability test, the method comprising the steps of: a) acquiring (101), by a digital image acquisition device operably connected to the test bench, a digital image of the object or of a portion of the object from which the anomalies are to be detected; b) providing (102), by the image acquisition means, the acquired digital image to a first data processing unit operatively connected to the digital image acquisition means and adapted to execute an anomaly detection algorithm trained by artificial intelligence and / or machine learning techniques; c) assigning (103), by the first data processing unit, a value to each pixel of the acquired digital image by executing a trained anomaly detection algorithm, the value representing a level of matching between the pixel and the same pixel of the at least one reference digital image, the at least one reference digital image represents a normal state of the object obtained after training the anomaly detection algorithm; d) comparing (104), by the first data processing unit, the value assigned to each pixel of the acquired digital image with a set first threshold value by executing a trained algorithm, assigning a normal state to the pixel if the assigned value is below the set first threshold value, and assigning a normal state to the pixel if the assigned value is above the set first threshold value; if so, assigning an abnormal state to the pixel; e) assigning (105), by the first data processing unit, a normal state or an abnormal state to the acquired digital image on the basis of the state assigned to each pixel of the acquired digital image by executing a trained algorithm, and if the number of pixels assigned with an abnormal state is above a set second threshold value, if the number of pixels assigned with an abnormal state is not above the set second threshold value, assigning (105) the normal state or the abnormal state to the acquired digital image. If the number of the pixels distributed with the abnormal state is lower than a set second threshold value or the surface density of the pixels distributed with the abnormal state is lower than the set third threshold value, distributing the abnormal state to the acquired digital image, and if the number of the pixels distributed with the abnormal state is lower than the set second threshold value or the surface density of the pixels distributed with the abnormal state is lower than the set third threshold value, distributing the abnormal state to the acquired digital image. If so, a normal state is assigned to the acquired digital image.

Description

Technical Field

[0001] The present invention relates to a method and related system for real-time detection of anomalies in an object undergoing durability testing through the use of artificial intelligence (AI). Background Art

[0002] Nowadays, the use of artificial intelligence (AI) technology to identify abnormal images relative to images representing a normality condition is being consolidated.

[0003] Such tools analyze, by means of appropriately conceived algorithms (usually one or more neural networks), manually or automatically acquired images in which abnormalities of different types and severity may be present.

[0004] The algorithm is suitably trained to recognize images representing a normal state and to identify deviations from such a reference as anomalies.

[0005] When the diversity of abnormal patterns makes it difficult to collect sufficient abnormal data to perform training on abnormal images, so-called anomaly detection algorithms are preferred over deep learning-based object detection algorithms.

[0006] Examples of application areas for these technologies are searching for defects in industrial products and identifying anomalies in images of traffic or crowds.

[0007] Additionally, anomaly detection algorithms are used when the defect being searched for is known but very rare, so few images with that defect are available.

[0008] There has been no prior art attempt to exploit the potential of artificial intelligence (AI) to automate durability testing of assembled components, such as vibration durability testing.

[0009] In this regard, durability testing is applicable to products that may be subjected to harmonic vibration, broadband vibration or mechanical shock during transportation or operation.

[0010] Durability testing aims to define the dynamic behavior of the sample and detect any mechanical weaknesses or degradation of specified properties.

[0011] This type of testing is common in the automotive, rail, military, aviation, aerospace, nuclear, and telecommunications sectors for testing complex components.

[0012] Depending on the method used, durability testing (in the specific case of vibration durability testing) involves applying a vibration profile (random, sinusoidal, etc.) to the component and monitoring the condition of the component being tested over time by inspecting or measuring its specific properties.

[0013] The duration of a general test of this type is of the order of tens of hours (24 to 48), ending when the test reaches its natural end or as soon as any defect is detected on the component (ie a change in position of a subcomponent, its separation from the body, etc.).

[0014] The vibration tests thus performed are very expensive in terms of resources, also due to their lengthy duration.

[0015] In fact, in addition to the long-term occupation of the machine, the presence of an operator is required to monitor the condition of the components.

[0016] Furthermore, this method of conducting the tests does not allow for the extraction of all available information from the experiments.

[0017] In fact, up to now, the condition of the components has been checked at regular intervals in search of anomalies, but it would be interesting to know the exact moment when the first anomaly occurs for two reasons.

[0018] First, the test can be stopped immediately and the machine freed up for the next test.

[0019] Furthermore, the availability of more accurate data on the moment the anomaly occurred, combined with data on the state of the console during the test, allows for a better study of the behavior of the article being tested.

[0020] The increased available data (including, for example: number of cycles where anomalies occurred, load cycles) will also allow intervention in the most critical areas of the tested component during the design step.

[0021] The described use case requires that the algorithm produce as few false positives as possible when evaluating whether an image subjected to the algorithm is anomalous or non-anomalous.

[0022] However, existing anomaly detection algorithms can only achieve the required performance level with lengthy training and a large number of images.

[0023] In view of the above, there is a strong need today to have methods and systems for real-time detection of anomalies in objects undergoing durability testing by using artificial intelligence (AI), which methods and systems are able to optimize the resources used in terms of training duration and the number of images acquired while maximizing the amount of information available. Summary of the Invention

[0024] The object of the present invention is to design and provide a method for real-time detection of anomalies in objects subjected to durability tests by using artificial intelligence (AI), which method allows to at least partially eliminate the drawbacks complained of above with reference to the prior art and which method is particularly capable of optimizing the resources used in terms of training duration and number of acquired images while maximizing the amount of information available.

[0025] Such an object is achieved by a method for real-time detection of anomalies of an object subjected to a durability test by using artificial intelligence (AI) according to claim 1 .

[0026] A system for detecting anomalies of objects subjected to durability tests in real time by using artificial intelligence (AI), suitable for implementing the above-mentioned method, is also an object of the present invention.

[0027] Further advantageous embodiments of the method and system are the subject matter of the respective dependent claims. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Further features and advantages of the method and the associated system according to the invention will become apparent from the following description of a preferred embodiment given by way of non-limiting indication with reference to the accompanying drawings, in which:

[0029] - Figure 1 A system for real-time detection of abnormalities of an object subjected to durability testing by using artificial intelligence (AI) according to an embodiment of the present invention is shown by a block diagram;

[0030] - Figure 2 A system for real-time detection of abnormalities of an object subjected to durability testing by using artificial intelligence (AI) according to another embodiment of the present invention is shown by a block diagram;

[0031] - Figure 3 A system for real-time detection of abnormalities of an object subjected to durability testing by using artificial intelligence (AI) according to another embodiment of the present invention is shown by a block diagram;

[0032] - Figure 4 A system for real-time detection of abnormalities of an object subjected to durability testing by using artificial intelligence (AI) according to another embodiment of the present invention is shown by a block diagram;

[0033] - Figure 5 shows an example of a digital image provided as input to a machine learning algorithm during a training step according to an embodiment of the method of the present invention;

[0034] - Figure 6a and Figure 6b respectively show examples of digital images obtained as output from a process using a data enhancement technique according to an embodiment of the method of the present invention;

[0035] - Figure 7a shows an example of a raw digital image overlaid with a color scale indicating the level of anomalies detected in each pixel of the image;

[0036] - Figure 7b Shown Figure 7a Magnified image of;

[0037] - Figure 8 The operation of a method for real-time detection of abnormalities of an object subjected to durability testing by using artificial intelligence (AI) according to the present invention is shown from a logical perspective through a functional block diagram;

[0038] - Figure 9 A block diagram shows the training steps of the method for real-time detection of anomalies of an object subjected to durability testing by using artificial intelligence (AI) according to the present invention;

[0039] - Figure 10 A method for real-time detection of abnormalities of an object undergoing a test of the object by using artificial intelligence (AI) according to the present invention is shown by a block diagram.

[0040] It should be noted that in the drawings, identical or similar elements will be indicated by the same numerical or alphanumeric references. DETAILED DESCRIPTION

[0041] refer to Figures 1 to 4 , reference numeral 10 generally designates a system for real-time detection of abnormalities of an object subjected to a durability test by using artificial intelligence (AI) according to the present invention, hereinafter also referred to as a detection system or simply a system.

[0042] For the purpose of this description, Figures 1 to 4 An “object”, which is shown schematically only and indicated by reference numeral 1 , is any assembled component that can be used for a braking system of a vehicle, for example a brake caliper assembled on a brake disc, such as, for example Figure 5 、 Figure 6a 、 Figure 6b 、 Figure 7a and Figure 7b The objects shown in .

[0043] For the purposes of this description, an "abnormality" of an object refers to any undesirable change (whether in shape, appearance, or function) that is visible upon inspection of the object being analyzed.

[0044] In the specific case of assembled components that may be used in a braking system of a vehicle, the anomaly is related to a functional interruption, such as, for example, a broken spring, a broken pin, a loose bolt, a pin slipping off.

[0045] Other examples of anomalies could be paint corrosion, unclear markings, or other types of surface damage.

[0046] For the purposes of the present description, a "durability test" instead refers to a resistance test of an object, such as, for example, a vibration resistance test or a fatigue resistance test.

[0047] In both types of endurance testing, load cycles are applied to the object, which are characterized by a set sequence in terms of amplitude, application frequency and total duration, which in turn can also be defined in terms of the total number of load cycles.

[0048] In more detail, for example:

[0049] - In the case of vibration resistance testing, the load cycles applied to the object include dynamic testing, high application frequency, fast dynamics and short duration (approximately one to two days);

[0050] - In the case of fatigue resistance testing, the load cycles applied to the object include slow (almost static) testing, low application frequency, negligible dynamics and lengthy duration (approximately 30 days).

[0051] According to the invention, the system 10 comprises a test bench 20 configured to subject the object 1 to a durability test.

[0052] The test station 20 may be controlled by a corresponding electronic controller (not shown) operatively connected to the test station 20 .

[0053] In particular, the test station 20 can be controlled by command signals originating from a corresponding electronic controller, the command signals allowing stopping or not stopping the test, the duration of the test, the type of test, etc.

[0054] The system 10 also includes a digital image acquisition device 30 (eg, a digital camera or still camera) operably connected to the test station 20 .

[0055] The digital image acquisition device 30 is configured to acquire a digital image of the object 1 or a digital image of a portion of the object 1 where abnormality should be detected. The portion of the object 1 may coincide with the entire object 1.

[0056] The digital image acquisition device 30 is configured to acquire a still digital image (ie, unblurred) of the object 1 or a portion of the object 1 .

[0057] The digital image acquisition device 30 is controlled by an electronic computer (according to different embodiments described below), programmed with corresponding firmware, and for its use, includes setting corresponding digital image acquisition parameters such as, for example, exposure level and / or time, number of pictures per second, etc.

[0058] The system 10 further comprises a first data processing unit 40 , such as a microprocessor or microcontroller of a first electronic computer 41 , which is operatively connected to the digital image acquisition device 30 .

[0059] The first electronic computer 41 is local to the test bench 20 and is preferably an edge-type electronic computer.

[0060] The image acquisition device 30 is configured to provide the acquired digital image to a first data processing unit 40 operatively connected to the digital image acquisition device 30 .

[0061] The first data processing unit 40 is configured to execute an anomaly detection algorithm AD trained by artificial intelligence and / or machine learning techniques.

[0062] The trained anomaly detection algorithm AD is an inference algorithm, such as, for example, the PaDiM (Patch Distribution Modeling) algorithm.

[0063] Other examples of inference algorithms may be PatchCore and FastFlow.

[0064] The trained anomaly detection algorithm AD is implemented by the first data processing unit 40 through a set convolutional neural network (CNN).

[0065] Returning to the present invention, the first data processing unit 40 is configured to execute the steps of a method for real-time detection of anomalies of an object subjected to a durability test as described below.

[0066] The digital image acquisition device 30 is configured to acquire a digital image of the object 1 or a digital image of a portion of the object 1 where abnormality should be detected.

[0067] The digital image acquisition device 30 is configured to provide 102 the acquired digital image to the first data processing unit 40 .

[0068] The first data processing unit 40 is configured to assign a value to each pixel of the acquired digital image by executing the trained anomaly detection algorithm AD, the value representing the matching level (abnormality level) between the pixel and the same pixel of at least one reference digital image, the at least one reference digital image representing the normal state of the object 1 obtained after training of the anomaly detection algorithm AD.

[0069] The value indicating the matching level (abnormality level) is a value between 0 and 1.

[0070] In more detail, during analysis (inference), the trained anomaly detection algorithm AD extracts multiple features from the reference digital image of the object 1 through a set convolutional neural network.

[0071] For the purposes of this invention, "features" refer to mathematical objects (matrices and vectors) that describe certain characteristics of a digital image.

[0072] For example, there are "high level" features, "medium level" features, and "low level" features in a digital image.

[0073] In the case of a convolutional neural network designed to recognize faces in digital images, "low-level" features might identify lines and edges. These are then joined to form "mid-level" features (e.g., by combining edges to obtain shapes such as ellipses or circles), which are then joined again to obtain "high-level" features such as eyes and mouths (which are fundamental to recognizing faces).

[0074] For the purpose of applying the present invention, “low-level” features are used since only the first level of the set convolutional neural network is considered.

[0075] Once multiple features are extracted from the reference digital image of the object 1, the trained anomaly detection algorithm AD will determine a series of parameters that represent the distribution of multiple features present in the reference digital image, which represents the normal state of the object 1 obtained after the training of the said anomaly detection algorithm AD.

[0076] Such a series of parameters is Gaussian distributed, with each parameter being represented by its own mean and variance for the purposes of its calculations, etc.

[0077] During analysis (inference), the trained anomaly detection algorithm AD assigns a value representing the abnormality level to each pixel of the digital image of the object 1, and the assigned value is obtained by extracting multiple features from the collected digital image of the object 1 through a set convolutional neural network and using a series of determined parameters. The determined parameters represent the distribution of multiple features present in the reference digital image representing the normal state of the object 1 obtained after the training of the anomaly detection algorithm AD.

[0078] In more detail, the trained anomaly detection algorithm AD extracts multiple features of the acquired digital image of the object 1 and compares them with the above-mentioned series of parameters (conventional Gaussian distribution) and calculates the distance to normality using an appropriate mathematical formula (e.g., percentile of the distribution).

[0079] Since there are many different features among the multiple features, and considering that there are many Gaussian distributions, the trained anomaly detection algorithm AD calculates the average of the calculated distances and performs additional normalization to obtain a value between 0 and 1 representing the matching level (anomaly level).

[0080] The value representing the level of matching between the pixel and the same pixel of at least one reference digital image may be a color on a color scale ranging from lighter colors, such as yellow, to darker colors, such as blue.

[0081] In this example, lighter colors represent low match levels, while darker colors represent high match levels.

[0082] Figure 7a and Figure 7b The corresponding enlarged view shown shows an example of a captured digital image of the object 1 , which is superimposed with a color scale of the type described above.

[0083] Figure 7a and Figure 7b The brighter portion of object 1 is the portion where more abnormalities were detected during the durability test.

[0084] The first data processing unit 40 is further configured to compare the value assigned to each pixel of the acquired digital image with a set first threshold by executing the trained anomaly detection algorithm AD.

[0085] If the assigned value is below a set first threshold, the first data processing unit 40 is configured to assign a normal state to the pixel.

[0086] If the assigned value is above a set first threshold, the first data processing unit 40 is configured to assign an abnormal status to the pixel.

[0087] The first data processing unit 40 is configured to assign a normal state or an abnormal state to the acquired digital image based on a state assigned to each pixel of the acquired digital image by executing the trained abnormality detection algorithm AD.

[0088] If the number of pixels assigned the abnormal state is higher than a set second threshold, and the surface density of pixels assigned the abnormal state is higher than a set third threshold, the first data processing unit 40 is configured to assign an abnormal state to the acquired digital image by executing the trained anomaly detection algorithm AD.

[0089] If the number of pixels assigned to the abnormal state is lower than the set second threshold, or the surface density of pixels assigned to the abnormal state is lower than the set third threshold, the first data processing unit 40 is configured to assign a normal state to the acquired digital image by executing the trained anomaly detection algorithm AD.

[0090] The set first threshold, the set second threshold, and the set third threshold depend on specific usage situations, and thus can be appropriately set according to specific situations.

[0091] The set second threshold and the set third threshold can replace each other.

[0092] In more detail:

[0093] A first threshold value is set which represents the level of matching that is considered acceptable and is therefore a value between 0 and 1, for example equal to 0.7;

[0094] - the second threshold is set to the (absolute) minimum number of pixels, depending on the size of the anomaly to be detected, so it is typically a number between 100 and 1000;

[0095] The third threshold value is set to be a number of pixels that depends on the total number of pixels of the digital image, such as 1000 / (1024×1024). The third threshold value is set to be a value that is agnostic to the resolution of the digital image.

[0096] According to an embodiment, the first data processing unit 40 is configured to continue the durability test on the object 1 by executing the trained anomaly detection algorithm AD, when a normal state is assigned to the acquired digital image, by executing step a) to acquire the next digital image of the object (1), and performing steps b) to e) on the acquired next digital image.

[0097] According to an embodiment, in combination with the previous embodiment, the first data processing unit 40 is configured to interrupt the durability test on the object 1 (stop the test stand) if an abnormal state is assigned to the acquired digital image.

[0098] In an embodiment, in combination with the previous embodiment, the first data processing unit 40 is configured to send a corresponding message (eg, via e-mail) to the operator of the test bench 20 when the durability test is interrupted.

[0099] In an embodiment, in conjunction with the aforementioned embodiment including the interruption of the durability test, the first data processing unit 40 is configured to store first information indicating the interrupted durability test in the first storage unit 50 (in Figures 1 to 4 In the figure (shown schematically in FIG), the first storage unit is operatively connected to the first data processing unit 40.

[0100] Such first information indicating that the durability test has been interrupted includes:

[0101] - data related to the setting of the test bench 20;

[0102] - the time trend of the operating conditions generated during the test, e.g. in the case of vibration tests, the vibration frequency, vibration amplitude, number of cycles during the test, etc.;

[0103] - Detecting anomalies and controlling the number of test load cycles that have produced a positive result (e.g., 80,000 out of 100,000 (inclusive));

[0104] - the moment when the first anomaly is detected;

[0105] - The location of the anomaly in the digital image;

[0106] - Digital image of the control (stopping the test bench) for which the durability test of object 1 was carried out was interrupted.

[0107] According to an embodiment, in combination with any of the aforementioned embodiments, the first data processing unit 40 is configured to end the durability test on the object 1 when a test duration setting value set during the setup of the test station 20 (for example, set by an operator) is reached in the absence of a captured digital image assigned an abnormal state.

[0108] For example, the test duration setting value may be between 24 hours and 28 hours.

[0109] According to an embodiment, in combination with the previous embodiment, the first data processing unit 40 is configured to store second information indicating that the durability test has ended in a first storage unit 50 operatively connected to the data processing unit 40 .

[0110] The second information indicating the endurance test is completed includes:

[0111] - data related to the setup environment (settings) of the test bench 20;

[0112] - The time trend of the operating conditions generated during the test, e.g. in the case of a vibration test, the vibration frequency, vibration amplitude, number of cycles during the test, etc.

[0113] Reference again Figures 1 to 4The system 10 further comprises a second data processing unit 60 (e.g., a microprocessor or microcontroller of a second electronic computer 61), which is configured to execute the steps of the method for real-time detection of anomalies of an object subjected to durability testing according to the present invention described below.

[0114] The second electronic computer 61 is remote from the test bench 20 .

[0115] The second data processing unit 60 is further configured to train the anomaly detection algorithm AD by means of a corresponding training algorithm TR within a set initial time interval of the durability test to which the object 1 to be inspected is subjected.

[0116] The training algorithm T-DR is implemented by the second data processing unit 60 through a set convolutional neural network.

[0117] The training that can be performed by the training algorithm TR allows modifying a series of parameters of the trained anomaly detection algorithm AD, these parameters being used to assign an anomaly level to each pixel of the digital image of the object 1 during analysis (inference).

[0118] The initial time interval (training time) is set to be approximately several minutes to a maximum of 15 minutes, such as 10 minutes. A maximum of 5 minutes is used to acquire the digital image of the object 1, and a maximum of 5 minutes is used to train the anomaly detection algorithm AD. The training of the anomaly detection algorithm AD can also be completed in a shorter time, such as 2 minutes.

[0119] In more detail, in this training step, the second data processing unit 60 is configured to acquire a plurality of digital images of the object 1 by means of the digital image acquisition device 30 operatively connected to the second data processing unit 60 .

[0120] Figure 5 An example of a digital image of an object 1 acquired during training of an anomaly detection algorithm AD is shown.

[0121] The second data processing unit 60 is configured to process the acquired multiple digital images of the object 1 .

[0122] In an embodiment, the second data processing unit 60 is configured to perform a first processing on the acquired multiple digital images of the object 1 using a data enhancement technique.

[0123] In particular, the first processing consists in applying rules for varying brightness and contrast to the acquired digital image so that it is compatible with the brightness variations that have been experimented with during the durability tests, for example variations in the intensity and / or color of natural light, in addition to a possible switch from natural light to artificial light.

[0124] In an embodiment, in combination with the previous embodiment, the second data processing unit 60 is configured to perform the second processing on the plurality of acquired digital images of the object 1 that have been subjected to the first processing.

[0125] For example, the second processing includes identifying, in the acquired digital image that has undergone the first processing, portions of the object 1 to be inspected during the durability test (i.e., portions of the object 1 that should be detected for abnormalities), and excluding other portions of the object 1 that are not to be inspected during the durability test (i.e., portions of the object 1 that do not need to be detected for abnormalities).

[0126] In other words, the identification obtainable by the second processing includes isolating or surrounding the portion of the object 1 to be detected for abnormality in the acquired digital image, rather than excluding other portions of the object 1 that are not of interest for detecting abnormalities.

[0127] For example, in a digital image comprising a brake caliper and a background, the second processing allows only the portion corresponding to the brake caliper to be identified as part of the digital image of the object 1 to be subjected to the test.

[0128] Generally, processing the acquired plurality of digital images of the object 1 (a first processing, and possibly a second processing) comprises using a set convolutional neural network, for example a network trained on a public generic digital image dataset that is labeled and for which weights are assigned to the network nodes that are provided in a freely available online repository.

[0129] Returning to the present invention in general, the second data processing unit 60 is configured to provide the processed plurality of digital images of the object 1 to the anomaly detection algorithm AD for training.

[0130] It should be noted that if the anomaly detection algorithm AD is a PaDiM algorithm, the choice of this type of algorithm, in addition to being implemented in an edge computer as described below, advantageously allows ensuring a satisfactory time performance for training.

[0131] For example, if it is a PaDiM algorithm, the anomaly detection algorithm AD is capable of processing three pictures per second (fps), while the duration of training is at most two minutes.

[0132] The set convolutional neural network based on the anomaly detection algorithm AD is configured, during training, to learn to identify a plurality of features that can be extracted from the set convolutional neural network for each portion or block of the digital image.

[0133] In more detail, for each digital image, iteration (i.e., inference) is performed on a set convolutional neural network based on anomaly detection algorithm AD, which affects the above-mentioned multiple features with appropriate weights.

[0134] The set convolutional neural network determines a series of parameters that represent the distribution of multiple features present in the reference digital image, which represents the normal state of the object 1 acquired during the training of the anomaly detection algorithm AD.

[0135] At the end of the training, the above series of parameters for each feature are obtained for each portion or block of the acquired and processed digital image.

[0136] This type of training is also defined as “online” training, since the data required for training becomes available sequentially and at each pass (ie for each acquired digital image), the parameters of the anomaly detection algorithm AD are updated.

[0137] It should also be noted that incorporating the digital images subjected to the first and second processing described above according to the embodiment into the reference dataset plays a decisive role in improving the judgment performance of the trained anomaly detection algorithm AD.

[0138] The processed plurality of digital images of the object 1 represent a plurality of reference digital images that may be used by the first data processing unit 40 to detect anomalies of the object subjected to the durability test by means of the trained anomaly detection algorithm AD.

[0139] Figure 6a and Figure 6b An example of a processed digital image of an object 1 is shown.

[0140] From the perspective of architecture, back to the system 10, according to the embodiment, any one of the embodiments described above and the Figures 1 to 4 As shown by the dashed line, the system 10 further includes a uniform background panel PS, which is arranged on the test bench 20 at a position opposite to the digital image acquisition device 30 .

[0141] The uniform background panel PS may be a neutral color panel (eg, white).

[0142] The uniform background panel PS advantageously allows obtaining a frame of the object 1 and a sharper, cleaner acquired digital image of the object 1 .

[0143] Figure 5 An example of a digital image of the object 1 acquired in the above-described state is shown again.

[0144] According to an embodiment, in combination with any one of the above embodiments and Figure 1 、 Figure 3 and Figure 4As shown, the system 10 includes a connection device 70, such as a USB relay type connection device, which is placed between the test station 20 and a data communication network (not shown in the figure) to which the test station 20 is connected.

[0145] For example, the connection device 70 is placed between the test bench 20 and the first data processing unit 40 (ie, the first electronic computer 41 ).

[0146] The connection device 70 is configured for controlling electrical control signals into the input of the test bench 20 .

[0147] If the electrical control signal is lower than or equal to a set limit threshold, the system 10 is configured to stop the operation of the test station 20 .

[0148] For example, the electrical control signal at the input to the test station 20 is equal to 5V, while the set limit threshold is equal to 1V.

[0149] In an embodiment, in combination with any of the above embodiments and Figure 1 As shown, the data processing unit 40 (i.e., the first electronic computer 41, such as an edge-type electronic computer) is configured to execute the trained anomaly detection algorithm AD (as already described above), control the digital image acquisition device 30, and control the electrical control signal provided by the connection device 70 to the input of the test bench 20.

[0150] According to an embodiment, in combination with any of the above embodiments in which the connection device 70 is present and Figure 3 and Figure 4 As shown, the system 10 includes an electronic stage computer 80 that is directly connected to the digital image acquisition device 30 and is operatively connected to the first data processing unit 40 .

[0151] In this embodiment, the electronic stage computer 80 is configured to store the digital images acquired from the digital image acquisition device 30 .

[0152] In an embodiment, the first electronic computer 41 is a remote electronic computer relative to the test bed 20 , such as a company network server or a cloud, and includes a first data processing unit 40 representing a virtual machine.

[0153] In an embodiment, Figure 3 As shown, the first data processing unit 41 is configured to control the storage of new digital images acquired from the digital image acquisition device 30 in the electronic stage computer 80 .

[0154] In this embodiment, an electronic bench computer 80 is directly connected to the connection device 70 and is configured to control the electrical control signals into the input of the test bench 20 .

[0155] In an embodiment, as an alternative to the aforementioned embodiment and as Figure 4 As shown, the first electronic computer 41 is directly connected to the connection device 70 and the first data processing unit 40 is configured to control the electrical control signals into the input of the test bench 20 .

[0156] According to an embodiment, as any of the above embodiments (such as Figure 2 ), the system 10 comprises an on-board electronic platform on which the first data processing unit 40 is mounted.

[0157] Therefore, the on-board electronic platform can be considered equivalent to the first electronic computer 41 defined above.

[0158] The onboard electronic stage is directly connected to the test stage 20 and directly connected to the digital image acquisition device 30 .

[0159] In this embodiment, the first data processing unit 40 is configured to control the image acquisition device 30 .

[0160] In this embodiment, the onboard electronic board includes a plurality of pins (i.e., copper wires with voltages set by the onboard electronic board), which are controlled by the software of the onboard electronic board and are configured to set or acquire digital electronic signals, for example, equal to 3.3V or 5V.

[0161] With reference to the aforementioned drawings and Figures 8 to 10 , describes a method 100 for real-time detection of anomalies of an object 1 subjected to a durability test according to the present invention, hereinafter also referred to as detection method or simply method.

[0162] It should be noted that the components and information mentioned below in the context of describing the method have already been described above with reference to the system 10 and therefore will not be repeated for the sake of brevity.

[0163] Method 100 includes the symbolic step of initiating the STR.

[0164] The method 100 comprises the steps of: a) capturing 101 the object 1 or the portion of the object 1 to be detected for abnormality by the digital image capturing device 30; Figure 8 : n=1), the digital image acquisition device 30 is connected to the test bench 20 in an operative manner.

[0165] The method 100 comprises a step b) of providing 102 the acquired digital image by the image acquisition device 30 to a first data processing unit 40, which may be operably connected to the digital image acquisition device 30 and adapted to execute an anomaly detection algorithm AD trained by artificial intelligence and / or machine learning techniques.

[0166] The method 100 comprises a step c) of assigning 103 a value to each pixel of the acquired digital image by the first data processing unit 40 by executing a trained anomaly detection algorithm AD, the value representing a level of matching (anomaly level) between the pixel and the same pixel of at least one reference digital image, the at least one reference digital image representing a normal state of the object 1 obtained after training of the anomaly detection algorithm AD.

[0167] The method 100 further comprises step d): the first data processing unit 40 compares 104 the value assigned to each pixel of the acquired digital image with a set first threshold by executing the trained anomaly detection algorithm AD.

[0168] If the assigned value is below a set first threshold, the normal state is assigned to the pixel.

[0169] If the assigned value is above a set first threshold, an abnormal state is assigned to the pixel.

[0170] The method 100 further comprises a step e) of assigning 105 a normal state or an abnormal state to the acquired digital image based on the state assigned to each pixel of the acquired digital image by executing the trained anomaly detection algorithm AD by the first data processing unit 40 .

[0171] If the number of pixels assigned the abnormal state is higher than a set second threshold, and the surface density of pixels assigned the abnormal state is higher than a set third threshold, the acquired digital image is assigned the abnormal state CA.

[0172] If the number of pixels assigned the abnormal state is lower than a set second threshold, or the surface density of pixels assigned the abnormal state is lower than a set third threshold, the normal state CN is assigned to the acquired digital image.

[0173] The method 100 then comprises a symbolic step ED of ending.

[0174] according to Figure 10In the embodiment shown by the dashed line, the method 100 comprises the following steps: in the case where the normal state is assigned to the acquired digital image, the first data processing unit 40 acquires the next digital image of the object (1) by executing step a), and executes steps b) to e) on the acquired next digital image to continue 106 the durability test on the object 1 ( Figure 8 :n=n+1).

[0175] According to the embodiment, in combination with any of the above embodiments and Figure 10 As shown in dashed lines, the method 100 comprises the step of interrupting 107 the durability test of the object 1 by the first data processing unit 40 in case an abnormal state is assigned to the acquired digital image.

[0176] According to the embodiment, in combination with the above embodiment and Figure 10 As shown by the dashed line in , the step of interrupting 107 the durability test of the object comprises the following steps: the first data processing unit 40 sends 108 a corresponding message (eg, an e-mail) to the operator of the test station 20 .

[0177] According to an embodiment, in combination with any of the above embodiments including interrupting the durability test and Figure 10 As shown by the dotted line in , the step of interrupting 107 the durability test of the object includes the following steps: the first data processing unit 40 stores 109 first information representing the interrupted durability test in the first storage unit 50, and the first storage unit 50 is operably connected to the first data processing unit 40.

[0178] According to the embodiment, in combination with any of the above embodiments and Figure 10 As shown by the dashed line in FIG, the method 100 comprises the step of ending 110 the durability test of the object 1 by the first data processing unit 40 when a test duration setting value set during the test bench setup is reached in the absence of acquired digital images assigned an abnormal state.

[0179] According to the embodiment, in combination with the above embodiment and Figure 10 As shown by the dotted line in , the step of ending 110 the durability test on the object 1 includes the following steps: the first data processing unit 40 stores 111 second information indicating that the durability test has ended in the first storage unit 50 operably connected to the first data processing unit 40.

[0180] According to the embodiment, in combination with any of the above embodiments and Figure 10As shown in dashed lines, the method 100 comprises a step f) of training 112 the anomaly detection algorithm AD by the second data processing unit 60 by means of a corresponding training algorithm TR within a set initial time interval of the durability test to which the object to be inspected is subjected.

[0181] According to the embodiment, in combination with the above embodiment and Figure 10 As shown by the dotted line in , step g) of training 112 includes step f1): the second data processing unit 60 acquires 113 multiple digital images of the object through the digital image acquisition device 30, and the digital image acquisition device 30 is operably connected to the second data processing unit 60.

[0182] In addition, Figure 10 In the embodiment shown by the middle dashed line, step f) of training 112 includes step f2): the second data processing unit 60 processes 114 the plurality of acquired digital images of the object 1 .

[0183] Further details of the process are provided above, depending on the implementation.

[0184] In an embodiment, the step f2) of performing the processing 114 comprises a step of performing 115 a first processing by the second data processing unit 60 on the plurality of acquired digital images of the object 1 using a data enhancement technique.

[0185] As already mentioned above, the first processing consists in applying rules for varying the brightness and contrast to the acquired digital image so that it is compatible with the brightness variations that have been experimented with during the durability tests, for example variations in the intensity and / or color of natural light, in addition to a possible switch from natural light to artificial light.

[0186] In an embodiment, in combination with the previous embodiment, the step f2) of performing the processing 114 comprises a step of performing 116 a second processing by the second data processing unit 60 on the acquired plurality of digital images of the object 1 that have been subjected to the first processing.

[0187] For example, as already mentioned above, the second processing includes identifying (i.e., isolating or outlining) portions of the object 1 to be subjected to inspection during the durability test (i.e., portions of the object 1 for which abnormalities should be detected) in the acquired digital image that has undergone the first processing, and excluding other portions of the object 1 not to be inspected during the durability test (i.e., portions of the object 1 for which abnormalities do not need to be detected).

[0188] In addition, Figure 10In the embodiment shown by the dashed line, step g) of training 112 includes step g3): the second data processing unit 60 provides 117 a plurality of processed digital images of the object 1 to the anomaly detection algorithm AD to be trained.

[0189] The processed plurality of digital images of the object 1 represents a plurality of reference digital images that can be used by the first data processing unit 40 by means of the trained anomaly detection algorithm AD to detect anomalies of the object subjected to the durability test.

[0190] Referring now to the accompanying drawings, an embodiment of a method for real-time detection of an abnormality of an object subjected to a durability test according to the present invention will be described.

[0191] The object 1 is placed on a test bench 20 to perform a durability test such as a vibration resistance test.

[0192] The object 1 is arranged between the uniform background panel PS and the digital image acquisition device 30 .

[0193] Before the test begins, the operator sets the load profile to which the object 1 is to be subjected during the vibration resistance test, ie the operator sets the duration, vibration frequency, vibration type, etc. of the test.

[0194] At this time, within a set initial time interval (eg, 10 minutes) of the anti-vibration test, the anomaly detection algorithm AD is trained, and the algorithm will be subsequently implemented by the first data processing unit 40 during the anti-vibration test.

[0195] In more detail, with respect to the trained test bench 20 , the second data processing unit 60 of the remote electronic computer 61 trains the anomaly detection algorithm AD by means of a corresponding training algorithm TR.

[0196] The training algorithm T-DR is implemented by the second data processing unit 60 by means of a set convolutional neural network (e.g., a network trained on a labeled public general digital image dataset and for which weights are assigned to network nodes provided in a freely available online repository).

[0197] During training, the training algorithm TR modifies a series of parameters of the anomaly detection algorithm AD.

[0198] In more detail, in this training step, the second data processing unit 60 acquires a plurality of digital images of the object 1 by means of the digital image acquisition device 30 operatively connected to the second data processing unit 60 .

[0199] The second data processing unit 60 processes the acquired multiple digital images of the object 1 by training the algorithm TR and using a data enhancement technique implemented on a set convolutional neural network.

[0200] In particular, the second data processing unit 60 performs a first processing on the plurality of acquired digital images of the object 1, wherein the illumination of the acquired digital images is varied so as to be compatible with the brightness variations that have been experimented with during the durability test (e.g. variations in the intensity and / or color of natural light in addition to a possible switch from natural light to artificial light).

[0201] In addition, the second data processing unit 60 performs a second processing on the multiple digital images collected of the object 1 that have undergone the first processing, highlighting the parts of the object 1 that are to be inspected during the durability test (i.e., the parts of the object 1 where abnormalities should be detected) in the digital images collected that have undergone the first processing, and excluding other parts of the object 1 that are not to be inspected during the durability test (i.e., the parts of the object 1 that do not need to be detected for abnormalities).

[0202] The second data processing unit 60 provides the processed plurality of digital images of the object 1 to the anomaly detection algorithm AD.

[0203] Once the training is completed, real-time detection of anomalies of the object 1 undergoing the vibration resistance test begins.

[0204] The digital image acquisition device 30 acquires a digital image of the object 1 or a portion of the object 1 where abnormalities are to be detected. The digital image acquisition device 30 is operatively connected to the test bench 20 .

[0205] The image acquisition device 30 provides the acquired digital images to a first data processing unit 40 operatively connected to the digital image acquisition device 30 and adapted to execute an anomaly detection algorithm AD trained by artificial intelligence and / or machine learning techniques.

[0206] The first data processing unit 40 compares the acquired digital image with at least one reference digital image representing a normal state of the object 1 obtained after training of the anomaly detection algorithm AD by executing the trained anomaly detection algorithm AD.

[0207] The first data processing unit 40 assigns a value to each pixel of the acquired digital image by executing the trained anomaly detection algorithm AD, the value representing the matching level between the pixel and the same pixel of at least one reference digital image.

[0208] The first data processing unit 40 compares the value assigned to each pixel of the acquired digital image with a set first threshold value by executing the trained anomaly detection algorithm AD.

[0209] If the assigned value is lower than a set first threshold, the normal state CN is assigned to the pixel.

[0210] If the assigned value is higher than a set first threshold, the abnormal state CA is assigned to the pixel.

[0211] The first data processing unit 40 assigns a normal state CN or an abnormal state CA to the acquired digital image based on the state assigned to each pixel of the acquired digital image by executing the trained abnormality detection algorithm AD.

[0212] In more detail:

[0213] If the number of pixels assigned the abnormal state CA is higher than a set second threshold, and the surface density of pixels assigned the abnormal state CA is higher than a set third threshold, the first data processing unit 40 assigns the abnormal state CA to the acquired digital image;

[0214] If the number of pixels assigned the abnormal state CA is lower than a set second threshold, or the surface density of pixels assigned the abnormal state CA is lower than a set third threshold, the first data processing unit 40 assigns the normal state CN to the acquired digital image.

[0215] In a case where the normal state CN is assigned to the acquired digital image, the first data processing unit 40 continues to perform the durability test on the object 1 , acquires the next digital image of the object 1 , and performs the above-described operations on the next acquired digital image.

[0216] In case an abnormal state CA is assigned to the acquired digital image, the first data processing unit 40 interrupts the durability test of the object 1 by sending a corresponding alarm message to the operator of the test bench 20 and storing first information indicating the interrupted durability test in the first storage unit 50, which is operatively connected to the first data processing unit 40.

[0217] It can be understood that the purpose of the present invention has been fully achieved.

[0218] By exploiting the potential of artificial intelligence, and in particular anomaly detection algorithms trained through artificial intelligence techniques, the method and system of the present invention allow for automating testing, thereby making experiments more efficient in terms of resources used and maximizing the amount of information extracted.

[0219] Furthermore, using an anomaly detection algorithm to perform fatigue resistance testing on a test bench advantageously enables avoiding the step of labeling digital images of objects for training the algorithm, relative to an object detection algorithm.

[0220] Those skilled in the art may make changes and adaptations to the embodiments of the methods and related systems described above, and may replace elements with other functionally equivalent elements to meet the needs of the circumstances, without departing from the scope of the following claims. Each feature described above as belonging to possible embodiments can be implemented without regard to the other embodiments described.

Claims

1. A method (100) for detecting an anomaly of an object (1) undergoing a durability test in real time, the method comprising the following steps when performing a durability test on a test bench (20): a) acquiring (101) a digital image of the object (1) or a portion of the object (1) to be detected for abnormality by a digital image acquisition device (30), the digital image acquisition device being operatively connected to the test bench (20); b) providing (102) the acquired digital image by the image acquisition device (30) to a first data processing unit (40), the first data processing unit being operatively connected to the digital image acquisition device (30) and being adapted to execute an anomaly detection algorithm (AD) trained by artificial intelligence and / or "machine learning" techniques; c) assigning (103) by the first data processing unit (40) a value to each pixel of the acquired digital image by executing the trained anomaly detection algorithm (AD), the value representing the level of matching between the pixel and the same pixel of at least one reference digital image, the at least one reference digital image representing a normal state of the object (1) obtained when training the anomaly detection algorithm (AD); d) comparing (104) the value assigned to each pixel of the acquired digital image with a set first threshold by executing the trained anomaly detection algorithm (AD) by the first data processing unit (40), If the assigned value is below a set first threshold, a normal state (CN) is assigned to the pixel, assigning an abnormal state (CA) to the pixel if the assigned value is above a set first threshold; e) assigning (105) a normal state (CN) or an abnormal state (CA) to the acquired digital image based on the state assigned to each pixel of the acquired digital image by executing the trained abnormality detection algorithm (AD), and assigning the abnormal state (CA) to the acquired digital image if the number of pixels assigned the abnormal state (CA) is higher than a set second threshold value and the surface density of pixels assigned the abnormal state (CA) is higher than a set third threshold value, If the number of pixels assigned the abnormal state (CA) is lower than a set second threshold, or the surface density of pixels assigned the abnormal state (CA) is lower than a set third threshold, the normal state (CN) is assigned to the acquired digital image.

2. The method (100) according to claim 1, comprising the steps of: in a case where the normal state (CN) is assigned to the acquired digital image, the first data processing unit (40) acquires a next digital image of the object (1) by executing step a), and executing steps b) to e) on the acquired next digital image to continue (106) performing the durability test on the object (1).

3. The method (100) according to claim 1 or 2, comprising the step of interrupting (107) the durability test of the object (1) by the first data processing unit (40) in the event that the abnormal state (CA) is assigned to the acquired digital image.

4. The method (100) according to claim 3, wherein: The step of interrupting (107) the durability test of the object comprises a step of sending (108) a corresponding message by the first data processing unit (40) to an operator of the test station (20).

5. The method (100) according to claim 3 or 4, wherein: The step of interrupting (107) the durability test on the object comprises the steps of storing (109) by the first data processing unit (40) first information indicating that the durability test has been interrupted in a first storage unit (0), the first storage unit being operatively connected to the first data processing unit (40).

6. The method (100) according to any one of the preceding claims, comprising the step of ending (110) the durability test of the object (1) by the first data processing unit (40) when a test duration setting value set during the test bench setup is reached in the absence of a captured digital image to which the abnormal state (CA) is assigned.

7. The method (100) according to claim 6, wherein: The step of ending (110) the durability test on the object (1) comprises the following steps: the first data processing unit (40) stores (111) second information indicating that the durability test has ended in the first storage unit (50) operatively connected to the first data processing unit (40).

8. The method (100) according to any one of the preceding claims, comprising the following step f): the anomaly detection algorithm (AD) is trained (112) by a second data processing unit (60) using a corresponding training algorithm (TR) within a set initial time interval of the durability test to which the object (1) to be inspected is subjected.

9. The method (100) according to claim 8, wherein: The step f) of performing training (112) comprises the following steps: f1) acquiring (113) a plurality of digital images of the object (1) by the second data processing unit (60) via the digital image acquisition device (30), the digital image acquisition device being operatively connected to the second data processing unit (60); f2) processing (114) the plurality of digital images of the object (1) acquired by the second data processing unit (60); f3) The second data processing unit (60) provides (117) the processed multiple digital images of the object (1) to the anomaly detection algorithm (AD) to be trained, wherein the preliminary processed multiple digital images of the object (1) represent a plurality of reference digital images, and the plurality of reference digital images can be used by the first data processing unit (40) with the aid of the trained anomaly detection algorithm (AD) to detect anomalies of the object undergoing durability testing.

10. A system (10) for detecting anomalies of an object (1) undergoing a durability test on a test bench (20) in real time, comprising: a test bench (20) configured to subject the object (1) to a durability test; a digital image acquisition device (30) operatively connected to the test bench (20), the digital image acquisition device being configured to acquire a digital image of the object (1) or a portion of the object (1) to be detected for anomalies; - a first data processing unit (40) operatively connected to the digital image acquisition device (30) and configured to execute an anomaly detection algorithm (AD) trained by artificial intelligence and / or "machine learning" techniques; The first data processing unit (40) is configured to execute the method for real-time detection of anomalies in an object (1) subjected to a durability test according to any one of the preceding claims 1 to 6.

11. The system (200) according to claim 10, further comprising a second data processing unit (60), which is configured to perform the steps of the method for real-time detection of anomalies of an object (1) subjected to durability testing according to any of the preceding claims 8 to 9.

Images