Method for identifying and characterizing noises emanating from vehicle braking systems using artificial intelligence

Abstract

A method for identifying and characterizing noises generated from a vehicle's braking system comprises detecting noises generated by the vehicle's braking system under dynamic operating conditions and generating digital audio data representative of the detected noise. The method then comprises analyzing the digital audio data with a noise analyzer to identify potential squirt events and respective likely squirt frequencies, and generating squirt frequency information indicative of the squirt frequencies of the identified potential squirt events. The method also comprises filtering the digital audio data with high-pass filtering to remove spectral components with frequencies below a filtering frequency to generate filtered digital audio data, and generating respective spectrograms based on the filtered digital audio data, the spectrograms representing in graphical form information present in the filtered digital audio data, consisting of sound signal intensity, as a function of time and frequency. The method then comprises providing said spectrograms and said squirt frequency information to a trained algorithm, the algorithm being trained using artificial intelligence and / or machine learning techniques. The method also includes identifying, by a trained algorithm, noise events based on the spectrogram and squeak frequency information, classifying the identified noise events, and finally providing information about the identified noise events, each characterized by a respective category. The classifying step includes classification according to a plurality of categories, including at least a first category consisting of noises to be detected caused by characteristic dynamic operation of the braking system, and a second category consisting of anomalous noises caused by operation or test anomalies.

Description

[Technical field]

[0001] The present invention relates to a method for identifying and characterizing noise generated by a vehicle braking system using artificial intelligence (AI).

[0002] The invention also relates to a method based on artificial intelligence (AI) for analyzing noise generated during dynamic testing of a vehicle braking system. [Background technology]

[0003] Today, ride comfort is one of the important factors when evaluating a vehicle.

[0004] In particular, in the automotive industry, customer complaints about the so-called NVH (noise, vibration and harshness) issues associated with brake systems have become a serious problem.

[0005] For this reason, it is essential to understand the phenomena related to the generation of undesirable vibrations and noise and to prevent them from occurring from the design stage of the brake system.

[0006] The physical phenomenon of friction-induced vibration (FIV) is highly complex and characterized by nonlinear and multi-scale phenomena, and therefore no analytical mathematical theory exists that can comprehensively predict the behavior of the above-mentioned systems.

[0007] As a result, much of the research and analysis in this field is carried out experimentally, in both industrial and non-industrial environments: large amounts of data are collected by testing braking systems under controlled conditions and recording their responses as input parameters are varied.

[0008] To measure noise generation during the development phase, the braking system is tested on a dynamometric (or roller stand) test bench, where pre-determined braking sequences are applied.

[0009] During testing, noise is typically measured with a noise analyzer installed on the test bench. Standard noise analyzers identify noise events based on amplitude criteria and known Fourier transform-based spectral processing methods. This means that the noise analyzer is designed to recognize "tonal noise", i.e. events characterized by a very narrow frequency band. An example of tonal noise is "squirrel", a type of noise characterized by a spectrum with a narrow frequency band and a very high amplitude relative to the background (typically with an intensity of 50 dB or more above the background).

[0010] A noise analyzer will typically produce a result file containing information about the identified squirts (braking event number, presence or absence of squirt, time marker, frequency, sound pressure, duration, etc.).

[0011] However, because several different FIV phenomena may exist simultaneously, other types of noise may also occur during testing. For example,

[0012] Collision noise (also called "clan noise"), i.e. instantaneous pulses covering a wide frequency range, caused by collision phenomena between parts;

[0013] Chirp / wire brush, i.e. noise of many short pulses at different frequencies;

[0014] Artifacts, i.e., noise events with a wide wavenumber band and high intensity.

[0015] Known noise analyzers have difficulty, i.e. cannot reliably, recognize these types of noise (whose spectral characteristics are significantly different from squirrels) and may erroneously identify other types of noise as squirrels.

[0016] Furthermore, the detection and classification effectiveness that a noise analyzer can achieve is severely limited by the noise inherent in the test environment, which leads to both false positives (FPs) where squirts are detected when in fact they are not, and false negatives (FNs) where squirts are present but not detected.

[0017] It is well known that the output of a noise analyzer is only partially reliable, but on the other hand, it is also known that an engineer knowledgeable in noise and vibration can classify different types of noise based on the spectrograms generated from acoustic measurements.

[0018] A spectrogram is a conversion of noise data (measured frequency, time of occurrence, and sound intensity) into an image.

[0019] Therefore, it is known that it may be possible to convert an audio track into a spectrogram with an equivalent amount of information.

[0020] Based on this, a method is proposed to interpret the information in the spectrogram using machine learning (ML) algorithms.

[0021] These techniques employ learning techniques that start from known data available in the form of spectrograms, where known noise events are highlighted and associated with information about each known type of noise.

[0022] Given the technical complexity of the problem, especially the difficulty of detecting, recognizing, and distinguishing between the vast number of noises that may occur in practice due to many different sources, despite progress in the field, there remain many open questions and unmet needs, such as the application of ML algorithms for the interpretation of spectrograms.

[0023] First, it should be noted that the distinction between different noise classes is not always clear even for domain experts. For example, identifying a chirp from a very short, small squirt can be difficult even when applying ML algorithms to a spectrogram. Squirts of a certain frequency do not necessarily have a constant intensity, and as a result change color on the spectrogram. If the squirt intensity changes constantly, it can be difficult to visually understand whether it is a single event or several different events following each other in rapid succession.

[0024] In the aforementioned known solutions, a single frequency squeak is always classified as a single event, regardless of its intensity modulation, which may lead to an erroneous assessment.

[0025] Furthermore, on a spectrogram, the fundamental frequency of a squirt may appear at multiple fundamental frequencies, with higher harmonics (i.e. noise whose spectrogram shape resembles the original squirt), but with decreasing intensity. In known noise analyzers, these higher harmonics are often erroneously classified as if they were separate squirts, even though they should be separated and excluded from the analysis, and ideally recognized and classified as a separate category.

[0026] Moreover, another issue not addressed by known methods, but important during testing, is the treatment of low-frequency (<500Hz) noise events. Indeed, it is common for spectrograms to show low-frequency noise (often very loud) due to the test environment.

[0027] Because the color scale of the spectrogram is normalized to the maximum and minimum intensity values ​​detected in the analyzed time interval, i.e., it is a relative scale, low-frequency noise events will distort the color scale and hinder the visualization of events of interest at higher frequencies, reducing the effectiveness of noise recognition and classification algorithms.

[0028] Finally, another important aspect not addressed by known solutions (even those employing ML algorithms) concerns the recognition of noise categories indicative of anomalies. In particular, anomalies may be related to an improper experimental configuration (e.g., incorrect installation of some element on the dynamo bench) or a poorly designed brake system (e.g., pads hitting the brake caliper during braking). On the other hand, it is highly desirable to be able to recognize and classify noises associated with anomalies in order to generate appropriate anomaly warning signals, for example, during tests.

[0029] Thus, in the field of recognition and classification of noises generated by braking systems, the solutions known to date do not provide a fully satisfactory solution and there remain many unmet needs. Summary of the Invention

[0030] The present invention therefore relates to a method for improving the quality of the analysis noise generated during dynamic testing of an automotive braking system using artificial intelligence (AI).

[0031] In particular, it is an object of the present invention to provide a method for identifying and characterizing the noise generated by a braking system of a motor vehicle by using artificial intelligence, which makes it possible to at least partially overcome the drawbacks mentioned above with reference to the prior art and to meet the aforementioned needs, which are particularly felt in the technical field considered. Said object is achieved by a method according to claim 1.

[0032] Further embodiments of such a method are defined in claims 2-24. [Brief description of the drawings]

[0033] 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:

[0034] [Figure 1]FIG. 1 shows six examples of spectrograms identifying different types or categories of noise that are recognizable by the method according to the invention.

[0035] [Diagram 2] FIG. 2 shows further examples of spectrograms identifying additional noise categories recognizable by the method according to the invention.

[0036] [Diagram 3] FIG. 3 is a simplified block diagram illustrating some of the steps involved in one embodiment of the method.

[0037] [Figure 4] FIG. 4 is a simplified block diagram illustrating some of the steps involved in one embodiment of the method.

[0038] [Diagram 5] FIG. 5 shows an example of a spectrogram used in an embodiment of the method.

[0039] [Figure 6] FIG. 6 shows an example of a spectrogram used in an embodiment of the method.

[0040] [Figure 7] FIG. 7 shows an example of a spectrogram used in an embodiment of the method.

[0041] [Figure 8] FIG. 8 shows a spectrogram on which the tagging operation has been performed.

[0042] [Figure 9] FIG. 9 is a block chart illustrating one embodiment of a method according to the present invention.

[0043] [Figure 10] FIG. 10 shows four accuracy recovery diagrams associated with four different types of noise that can be recognized by the method.

[0044] [Figure 11] FIG. 11 is a simplified block diagram of a system capable of implementing the method according to the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0045] A method for identifying and characterizing noise generated by a vehicle's braking system is described.

[0046] The method includes the steps of first detecting noise generated by a vehicle braking system under dynamic operating conditions and generating digital audio data representative of the detected noise.

[0047] The method then comprises analyzing the digital audio data with a noise analyzer to identify potential squirt events and their respective likely squirt frequencies, and generating squirt frequency information indicative of the squirt frequencies of the identified potential squirt events.

[0048] The method then comprises the steps of filtering said digital audio data by high-pass filtering to remove spectral components with frequencies below a filtering frequency to generate filtered digital audio data, and generating a respective spectrogram based on the filtered digital audio data, the spectrogram representing in graphical form the information present in the filtered digital audio data, with sound signal intensity, as a function of time and frequency.

[0049] The method then includes providing the spectrogram and the squirrel frequency information to a trained algorithm, the algorithm being trained using artificial intelligence and / or machine learning techniques.

[0050] The method also includes identifying, by a trained algorithm, noise events based on the spectrogram and squirrel frequency information, classifying the identified noise events, and finally providing information about the identified noise events, each characterized by a respective category.

[0051] The aforementioned classification step comprises classification according to at least the following categories: a first category consisting of noises to be detected which are generated by characteristic dynamic behavior of the braking system, and a second category consisting of abnormal noises which are generated by operating or test anomalies.

[0052] According to one embodiment of the method, the aforementioned categories into which the noise is classified further include a third category that includes higher order harmonics that do not originate from physically generated noise.

[0053] According to an embodiment, the aforementioned first category of noise comprises squirt noise and / or chirp / wire brush noise and / or artifacts, ie noise with a wide frequency band and high intensity.

[0054] According to an implementation option of the aforementioned embodiment, the step of classifying the identified noise event further includes recognizing and further classifying the first category of noise as belonging to one of several subcategories (squeal noise, chirp / wire brush noise, artifacts).

[0055] According to one embodiment of the method, the aforementioned step of classifying the identified noise events further comprises recognizing and further classifying the second category of noise as belonging to one of several subcategories: abnormal noise due to imperfections in the test bench ("bench noise"), or noise due to collisions between braking system components ("clan noise").

[0056] According to an embodiment of the method, the dynamic operating conditions of the braking system that are the source of the noise are test conditions, and a predefined sequence of test braking events characterized by predefined parameters is applied to the braking system, said predefined parameters comprising at least a predefined rotational speed and / or a predefined brake pressure.

[0057] In this case, the method steps set forth above are performed at each test braking event.

[0058] According to one embodiment of the method, said trained algorithm is an algorithm trained by a preliminary training step on the basis of a training data set having spectrograms characterized according to said classification of noise into a number of categories and / or sub-categories corresponding to known conditions and desired as a result of the analysis, said spectrograms of the training data set and further information on the known classification of each noise event are provided as input to the trained algorithm.

[0059] According to an implementation option, the pre-training step comprises the steps of tagging or labeling known noise events present in each of the training spectrograms and calibrating parameters of the training algorithm based on the tagged or labelled processed training spectrograms.

[0060] According to an example implementation, the "tagging" process does not modify the spectrogram, but instead generates additional "ancillary" data.

[0061] According to an implementation option, the aforementioned tagging or labeling step is performed manually by drawing a rectangle over the patterns in the training spectrograms that have been identified as noise events and associating the rectangle with a label indicating the category and / or subcategory of the noise event, with reference to the aforementioned classification.

[0062] According to another implementation option, the aforementioned steps of tagging or labeling are performed with the support of enabling software (such as "labellmg").

[0063] According to another implementation option, the aforementioned step of tagging or labelling is supported by listening to an audio file representative of the detected noise.

[0064] According to one embodiment, the method includes the further step of validating the predictive ability of the trained algorithm on an additional dataset of tagged validation spectrograms.

[0065] According to one embodiment of the method, said trained algorithm is a machine learning algorithm based on a neural network.

[0066] According to possible implementation options, said neural network comprises a deep neural network, or a convolutional neural network, or a zonal convolutional neural network, or a region-based convolutional neural network.

[0067] According to another embodiment of the method, the trained algorithm is a machine learning algorithm based on a deep object detector or a two-stage deep object detector.

[0068] According to one embodiment, in addition to generating a spectrogram, the method further comprises generating a segmented spectrogram, in which points are graphically highlighted according to the intensity band they belong to within a number of intensity bands delimited by respective predefined thresholds.

[0069] In this case, the spectrogram is provided as additional input to the training algorithm, in addition to the unsegmented spectrogram and information about the estimated squirm frequency.

[0070] According to implementation options, the aforementioned intensity bands in which points are highlighted in the respective manner consist of a high intensity band bounded at the lower end by a first threshold, a medium intensity band below the first threshold and a low intensity band below the second threshold, between the first and second thresholds.

[0071] According to a particular embodiment, the first threshold is set to 50 dB and the second threshold is set to 30 dB.

[0072] According to an embodiment of the method, the aforementioned step of generating digital audio data representative of the detected noise includes generating files and / or audio tracks captured while performing tests on the braking system.

[0073] According to a particular embodiment, the audio tracks are in ".wav" or ".mpeg" format.

[0074] According to one embodiment of the method, said step of analysing the digital audio data comprises identifying noise events, and therein squirt events, based on intensity or amplitude related criteria and / or based on frequency related criteria by spectral methods such as Fourier transform, and generating by the noise analyzer a tabular first data file recording all potential squirt events identified by the noise analyzer, and for each squirt event, the time instant at which it occurred, its duration, its respective squirt frequency, maximum and / or average sound pressure and / or amplitude and / or sound intensity at that time instant.

[0075] In this case, the above-mentioned squirrel frequency information is obtained in tabular form by the above-mentioned first data file.

[0076] According to an implementation option, said filtering frequency in the step of filtering the digital audio data by high-pass filtering is 500 Hz.

[0077] According to one embodiment of the method, the aforementioned step of providing information about the identified squeak events comprises generating a second data file in tabular form based on the results of processing by the trained algorithm, in which are recorded all the squeak events identified by the trained algorithm and, for each squeak event, the time instant at which it occurred, its duration, the respective squeak frequency, the maximum and / or average sound pressure and / or amplitude and / or sound intensity at that time instant.

[0078] According to one option of embodiment of the method, the second data file in tabular form is an improvement of the first data file in tabular form, in which all false positives resulting from events recognized as belonging to a third category consisting of higher harmonics and / or all events recognized as belonging to a second category consisting of noise resulting from anomalies are removed.

[0079] According to an embodiment of the method, if a noise event belonging to a third category is identified, the method further comprises generating a warning and / or alarm signal associated with the identified noise event of the third category.

[0080] According to an implementation option, if the identified noise event belongs to the abnormal noise due to test bench imperfections subcategory, the current test is stopped and a further step of validating the text bench is taken.

[0081] Further details of the method are described below with reference to FIGS. 1-11 for illustrative and non-limiting purposes only, according to certain embodiments of the present invention.

[0082] In this example, the noise generation is measured during the development step of the braking system by testing on a dynamometric test bench, during which predefined braking sequences are applied with respect to operating parameters such as rotational speed and brake pressure.

[0083] The noise generated during the test is measured by a known noise analyzer mounted on the bench itself.

[0084] For this purpose, a standard noise analyzer can be employed that is configured to identify noise events based on amplitude criteria and / or spectral methods including Fourier transform. The output provided by the noise analyzer is a tabular data file that lists, for each braking event tested, the relevant data specific to the detected squeak event, i.e., time instant, frequency, maximum and average sound pressure, duration, etc.

[0085] Data files may contain errors in the form of false positives.

[0086] To improve the quality of the measurement and / or detection, it is necessary to remove frequency values ​​associated with false positives.

[0087] In this regard, the present invention, particularly in the example embodiments described herein, employs artificial intelligence (AI) techniques to check / verify all noise events recorded by the noise analyzer during testing and correct the original file (e.g., by removing false positives) to obtain more reliable, reproducible and objective test results.

[0088] For example, to build machine learning algorithms ML, i.e. algorithms that are pre-trained on, for example, another dataset, "transfer learning" training techniques are employed.

[0089] As a further example, with regard to the algorithms already mentioned above, it is worth noting that ML algorithms known per se can be used, including, for example, the "Mask-RCNN" model based on a neural network (NN) trained using the open source COCO dataset as a training dataset.

[0090] A typical flowchart of an ML algorithm includes the steps of input preparation, tagging, and model training, as shown in Figure 3.

[0091] In the present invention, these steps are implemented with special features, especially regarding input preparation and tagging, to achieve improvements over the prior art and yield an enriched / improved model of AI-based noise detection.

[0092] The input preparation step includes all operations aimed at processing the numerical information contained in the output table file (provided by the noise analyzer) in order to obtain digital information suitable to be more effectively provided as input to an ML algorithm, e.g. a "deep learning" model.

[0093] In the embodiment described herein, this is achieved by a series of steps, such as those shown in FIG. 4, in particular by suitably converting the sound data into a digital image suitable for use as input to an ML algorithm.

[0094] First, a high-pass filter is applied to each detected audio track to remove low-frequency (e.g., <500 Hz) background noise that can introduce biases or shifts in the color scale of the resulting digital image (also called a "spectrogram") and have detrimental effects in subsequent processing steps.

[0095] From the audio data manipulated as described above, a spectrogram, i.e. a graph of frequency as a function of time, colored according to the local intensity of the sound signal, is generated for each braking event, for each time-frequency point on the graph.

[0096] A spectrogram is shown, for example, in FIG.

[0097] According to an implementation option, in order to make the image more readable, additional operations are performed on the spectrogram to highlight high and medium intensity points and make them stand out from the background.

[0098] All points whose intensity exceeds 50 dB (high intensity) are colored in a bright first color (e.g., pink).

[0099] All points with an intensity between 30 dB and 50 dB (medium intensity) are colored in a second vibrant color (e.g. dark red).

[0100] This is shown in Figure 6, which depicts the spectrogram before (left) and after (right) segmentation, where pink and dark red are rendered in lighter grey tones, respectively, compared to the background.

[0101] By setting the two thresholds above, two ranges of strength are defined and displayed, which represent additional levels of information that the user considers important (to the ML algorithm).

[0102] According to another implementation option, multiple thresholds are defined and then multiple (and thus any number greater than two) intensity division intervals are defined by setting multiple thresholds.

[0103] Concerning the "tagging" or labeling step, implementation options include manual tagging of the spectrogram. This "tagging" step involves identifying predefined categories of sounds in the image (spectrogram). Accurate tagging of training images is an important prerequisite for achieving effective operation of the algorithms to be trained (e.g. "deep learning").

[0104] In this example, the noise categories considered are (already mentioned and illustrated above):

[0105] Squeal, chirp / wire brush, artifacts; Higher harmonics; bench noise (i.e., anomalous noise indicating poor experimental arrangement); A clanging noise indicates a problem with the brake system design.

[0106] How these categories appear in spectrograms is shown in Figure 1. The top left spectrogram is squirt, the top center spectrogram is wire brush, the top right spectrogram is squirt with higher harmonics, the bottom left spectrogram is noise artifacts, the bottom center spectrogram is clang noise, and the bottom right spectrogram is test bench noise.

[0107] Figure 2 shows a single-frequency squirt with varying intensity.

[0108] Figure 5 shows the squirrel in a spectrogram with advantageous highlighting.

[0109] The "tagging" operation is performed, for example, by drawing rectangles in the image and associating with each rectangle a label indicating one of the aforementioned noise categories.

[0110] Depending on the specific implementation option, the "tagging" step is supported by software that adds some functionality, such as the aforementioned open source tool "labellmg".

[0111] (i) First, the user can switch between regular color spectrograms and segmented spectrograms depending on the type of coloring scheme that makes it easier to recognize a given pattern on the image. Moreover, color segmentation facilitates the definition of shared rules for tagging noise events in the image, thus making the process more objective and reproducible. For example, in the case of single-frequency squeaks with varying intensity, the two thresholds defined for the segmentation of the spectrograms make it possible to establish shared and unambiguous rules for determining whether the analyzed event consists of a single squeak or of several squeaks that are close to each other but different (see, for example, Figure 7, which shows the situation of the same single-frequency squeaks on a spectrogram without segmentation (left) and on a spectrogram with segmentation (right)).

[0112] (ii) Furthermore, as a further option for the implementation of this method, the frequency at which the noise analyzer recorded each squirt could be displayed on the spectrogram. In the spectrogram coordinates, these frequency values ​​correspond to horizontal lines drawn at different heights.

[0113] (iii) Additionally, a further implementation option of the method provides the ability to listen to an audio track that corresponds to the displayed spectrogram, thus supporting visual pattern recognition.

[0114] The three aforementioned options, especially if all of them are adopted, will help to improve the quality of the tagging and thus both the training and the complete performance of the algorithm.

[0115] Reference is now made to Figure 8, which illustrates a Graphical User Interface (GUI) of the "labellmg" tagging tool enhanced with features (i), (ii), and (iii) above.

[0116] The tagging step is followed by a learning process, which in the example considered here is carried out in the following way: A "tagged" subset of the dataset (consisting of 1017 audio files) is provided as input to the AI ​​algorithm to calibrate the model parameters and make them suitable for making predictions. For each braking event, the tagged data provided to the algorithm includes three interrelated entities (shown in Figure 4): the spectrogram, the segmented spectrogram, and the set of frequencies at which the noise analyzer identified the squeak (the latter information can be extracted directly from the noise analyzer and included in its tabular output file).

[0117] In the embodiment described here, after the algorithm is trained, its predictive ability is tested on another data set of the same nature. If the algorithm identifies noise other than squirts (or no noise) with a confidence level above some pre-set threshold, then a false positive will appear, i.e. an event that is recognized as a squirt by the noise analyzer, but in fact is not.

[0118] At this point, the algorithm transfers the identified spectrogram information into the time, frequency and intensity domains and removes the events of interest from the initial file.

[0119] Once this process is complete, the output file will have the same header and structure as the original table file, but will have any squirrel-related metrics identified by the artificial intelligence (AI) as false positives removed.

[0120] An overall flow chart of the above embodiment is shown in FIG.

[0121] If a noise event is recognized during testing as clang noise or test bench noise, according to implementation options, the method not only removes the corresponding frequency value from the tabular file, but also issues a warning that there is noise that may be associated with an anomaly that may adversely affect the entire test.

[0122] If bench noise is identified and the test in which the noise was identified is still in progress, the operator may decide to stop or pause the test so that they can verify that the bench is properly set up.

[0123] If the ML algorithm is run after the completion of a test in which at least one bench noise or crash noise was reported, this information may prove to be particularly useful, as it alerts the operator to unexpected or atypical results obtained during the analyzed test session.

[0124] In the examples described here, testing was performed on subsets of the dataset that were tagged complementary to the subset employed during training.

[0125] The test dataset consists of 270 audio files containing a total of 683 squeak events, 29 chirp / wire brush events, 20 clang noise events, and 42 noise artifact events.

[0126] The precision-recall curve plots for the object detection task are shown in Figure 10 along with the corresponding mean average precision (mAP) and precision (AP) across all values ​​of the recall curve. The four plots in Figure 10 are squirrels (top left), chirp / wire brush (top right), noise artifacts (bottom left), and clang noise (bottom right). The Intersection on Union (IoU), used to determine the agreement between reality and prediction, was set to 0.5.

[0127] These performance levels, together with the addition of certain post-processing logic (e.g., removal of squirrel harmonics at multiple frequencies according to the integer value of the lowest frequency detected in the same time interval), show that introducing 4% false removals from the total number of proposed removals (i.e., increasing the false negative rate from 0% to 2%) can reduce the occurrence of false squirrels by 30% (from 35% to 5%).

[0128] The removal of noise other than squirrel noise is carried out in parallel during dynamic testing.

[0129] During testing, when one or more occurrences of squirts are recorded by the base system, an audio file will be recorded and stored.

[0130] According to the implementation option, the audio files and a portion of the squirrel detection report are sent to a centralized server for analysis. A squirrel detection report is created in which false occurrences of squirrels identified by the AI ​​system are filtered. This report is made available in the centralized repository along with alerts on anomalies (i.e., the presence of bench noise and clan noise).

[0131] One embodiment of a system capable of implementing the above method according to the present invention is shown in FIG.

[0132] The components of the system shown in Figure 11 are N dynamo (or roller stand) test benches (each shown with an audio file, squirt report, basic detector and file transmission block), all connected to a centralized artificial intelligence (AI) server consisting of a "noise detector" block running machine learning algorithms, which stores audio files, warnings / alerts and corrected reports on the occurrence of squirts.

[0133] Thus, the above-mentioned objects of the invention are fully achieved by the above-mentioned method, thanks to the features disclosed in detail above. The advantages and technical problems solved by the method according to the invention have been described above with reference to various features and aspects of the method.

[0134] To meet foreseeable needs, those skilled in the art can make modifications and adaptations to the above-described method embodiments or replace them with other functionally equivalent elements without departing from the scope of the following claims. All features described above as belonging to possible embodiments can be implemented without regard to the other embodiments described above.

Claims

1. 1. A method for identifying and characterizing noise generated by a vehicle braking system, comprising: detecting noise generated by the vehicle braking system under dynamic operating conditions; generating digital audio data representative of the detected noise; analyzing the digital audio data with a noise analyzer to identify potential squirt events and their respective likely squirt frequencies, and generating squirt frequency information indicative of the squirt frequencies of the identified potential squirt events; filtering the digital audio data by high-pass filtering to remove spectral components with frequencies below a filtering frequency to generate filtered digital audio data; generating, based on the filtered digital audio data, respective spectrograms that graphically represent sound signal strength information present in the filtered digital audio data as a function of time and frequency; providing the spectrogram and the squeak frequency information to a trained algorithm, the trained algorithm being trained using artificial intelligence and / or machine learning techniques; identifying noise events based on the spectrogram and the squeak frequency information using the trained algorithm, and classifying the identified noise events according to a plurality of categories, including at least a first category consisting of noises to be detected that are generated by characteristic dynamic behavior of the vehicle braking system, and a second category consisting of abnormal sounds that are generated by operational or testing abnormalities; providing information about the identified noise events.

2. The method of claim 1 , wherein the plurality of categories into which noise is classified further includes a third category that includes higher order harmonics that are not derived from physically generated noise.

3. The first category of noise includes broadband, high-intensity noise. Squealing noises, and / or Chirp / wire brush noise, and / or The method of claim 1 , including artifacts.

4. The step of classifying the identified noise events comprises:

4. The method of claim 3, further comprising the step of recognizing and classifying the first category of noise as belonging to one of the subcategories: squirrel noise, chirp / wire brush noise, or artifact.

5. The step of classifying the identified noise events further comprises:

2. The method of claim 1, further comprising: recognizing and classifying the second category of noise as belonging to one of the subcategories of abnormal noise due to imperfections in a test bench or noise due to a collision between components of the vehicle braking system.

6. the dynamic operating conditions of the vehicle brake system that produce the noise are test conditions; 5. The method of claim 1, wherein a sequence of predefined test braking events characterized by predefined parameters is applied to the vehicle braking system, the predefined parameters comprising at least a predefined rotational speed and / or a predefined brake pressure.

7. said trained algorithm being trained by a pre-training step on the basis of a training data set comprising spectrograms corresponding to known conditions and characterized according to said classification of noise into said categories and / or sub-categories desired as a result of the analysis, The method of claim 1 , wherein the spectrograms of the training data set and information about known classifications of the noise events are provided as inputs to the trained algorithm.

8. The pre-training step comprises: tagging or labeling known noise events present in each of the spectrograms; The method of claim 7 , comprising the step of calibrating parameters of the trained algorithm based on the spectrogram processed by the tagging or labeling.

9. 9. The method of claim 8, wherein the tagging or labeling step is performed manually by drawing a rectangle over a pattern in the spectrogram identified as a noise event and associating with the rectangle a label indicating the category and / or subcategory of the noise event referenced in the classification.

10. The method of claim 8 , wherein the tagging or labeling step is performed with the support of enabling software.

11. The method of claim 8 , wherein the tagging or labeling step is supported by listening to an audio file representative of the detected noise.

12. 8. The method of claim 7, further comprising validating the predictive ability of the trained algorithm on an additional data set of tagged validation spectrograms.

13. The method of any one of claims 7 to 12, wherein the algorithm to be trained is a neural network-based machine learning algorithm.

14. 14. The method of claim 13, wherein the neural network comprises a deep neural network, or a convolutional neural network, or a zonal convolutional neural network, or a region-based convolutional neural network.

15. 13. The method of any one of claims 7 to 12, wherein the trained algorithm is a machine learning algorithm based on a deep object detector or a two-stage deep object detector.

16. In addition to generating the spectrogram, generating a segmented spectrogram; in said segmented spectrogram, points are graphically highlighted depending on the intensity band to which they belong within a plurality of intensity bands delimited by respective predetermined thresholds, 5. The method of claim 1, wherein the segmented spectrogram is provided as an additional input to the training algorithm in addition to the unsegmented spectrogram and estimated squeak frequency information.

17. 17. The method of claim 16, wherein the intensity bands within which the points are highlighted include a high intensity band bounded below by a first threshold, a medium intensity band between the first threshold and a second threshold below the first threshold, and a low intensity band below the second threshold.

18. 5. The method of claim 1, wherein the step of generating digital audio data representative of the detected noise comprises generating files and / or audio tracks captured while performing tests on the vehicle braking system.

19. analyzing the digital audio data, - identifying said noise events, especially squeak events, based on intensity or amplitude related criteria and / or based on frequency related criteria by spectral methods such as Fourier transform; generating, by said noise analyzer, a first data file in tabular form recording all potential squirt events identified by said noise analyzer and, for each squirt event, the time instant at which it occurred, its duration, its respective squirt frequency, and the maximum and / or average sound pressure and / or amplitude and / or sound intensity at that time instant; the squirrel frequency information is obtained in tabular form from the first data file; The method of claim 1.

20. 5. The method according to claim 1, wherein in the step of filtering the digital audio data by high-pass filtering, the filtering frequency is 500 Hz.

21. 2. The method of claim 1, wherein the step of providing information about the identified squeak events comprises generating, based on the results of processing by the trained algorithm, a second data file in tabular form that records all squeak events identified by the trained algorithm and, for each squeak event, the time instant at which it occurred, its duration, its respective squeak frequency, and the maximum and / or average sound pressure and / or amplitude and / or sound intensity at that time instant.

22. the second data file in tabular form is an improvement of the first data file in tabular form; 22. The method of claim 1, 19, or 21, wherein all false detections resulting from events recognized as belonging to the third category containing high-order harmonics and / or all events recognized as belonging to the second category containing noise derived from the anomaly are removed.

23. 6. The method according to claim 1, wherein when a noise event belonging to the second category is identified, the method further comprises generating a warning and / or alarm signal associated with the identified noise event of the second category.

24. 24. The method of claim 23, further comprising: if the identified noise event of the second category belongs to anomalous noise caused by imperfections in the test bench, stopping a current test and verifying the test bench.

Images