System for predicting tyre tread profiles of abstract length using ai

Abstract

The prediction of the profile of a tyre product exiting an extruder, depending on parameters of the extruder (EXTRD), comprises: - obtaining (221) a contextual profile (PL, PT); - dividing (223) the contextual profile into iterations of a sliding window (FGi) spanning the contextual profile over a set of iterations (FGtot), each iteration of the window (FGi) being centred on a point of interest (Pi) and including at least one point (Pi-1) behind the point of interest and at least one point (Pi+1) in front of the point of interest; - using a learning model (225) trained on the prediction of a point of interest (Pi) of a real profile (PR), depending on a window (FGi) of a contextual profile (PL, PT) centred on the point of interest (Pi), recursively (RCR) on the set of iterations so as to obtain (226) a prediction of the full real profile.

Description

[0001] DESCRIPTION

[0002] TITLE: System for predicting tread profiles of abstract lengths using AI

[0003] technical field

[0004] The invention relates to a method for training an automatic recognition model of material variations in tire product profiles (including the production of products intended to be incorporated into tires).

[0005] Previous techniques

[0006] In the tire industry, tires are required to meet various performance requirements (for example, low rolling resistance, improved wear resistance, comparable grip in wet and dry conditions, sufficient mileage, etc.). A tire is an object with a known geometry, generally comprising several superimposed layers of rubber (called "ply"), as well as a metal or textile fiber structure that forms a carcass reinforcing the tire's structure. The type of rubber and the type of reinforcement are chosen according to the desired final characteristics. A tire also includes a tread, which is added to the outer surface of the tire.

[0007] Referring to Figure 1, a representative tire 10 comprises a tread 12 intended to make contact with a ground via a tread surface 12a. The tire 10 further comprises a crown reinforcement including a working reinforcement 14 and a reinforcement reinforcement 16, the working reinforcement 14 having two working layers 14a and 14b. The tire 10 also includes two sidewalls (a sidewall 18 being shown in Figure 1) and two ribs 20 reinforced with a bead 22. A radial carcass layer 24 extends from one rib to the other, surrounding the bead in a known manner. The tread 12 has reinforcements consisting, for example, of superimposed layers having known reinforcing wires. In embodiments of the tire 10, the tire may include a rubber 26 which dissipates the static electricity produced during rolling.

[0008] The tread, 12, is typically formed by extrusion (carried out by a machine called an extruder), that is, by shaping the tread material, before curing, using an extrusion die. A cross-section of a tire tread reveals a contour called the tread profile. The tread profile, as it exits the extruder, determines, in particular, certain performance characteristics of the tire.

[0009] Therefore, it is important to master the industrial formation of tire treads, from the stage of new product design in relation to specifications, to the stage of quality control of the tread at a given stage of the production chain.

[0010] Predicting tread profiles at the extruder outlet is an open problem in the field of extrusion. Through experience, industry rules are established, allowing for the iterative design of extrusion blades and the determination of adjustment parameters to form the desired profile.

[0011] These approaches have led to improvements in the design time of extrusion blades.

[0012] However, conventional techniques do not allow for real-time prediction, for a given blade and a set of machine parameters, of a precise profile of the extruded tread.

[0013] Indeed, the vast majority of classical techniques operate using finite element simulation methods, which can prove to be correct and reliable but generate very high computational and time costs.

[0014] On the other hand, there are as many profiles as there are treads, so the method must have a generalized scope of application that allows, in particular, the prediction of profiles of arbitrary length. Furthermore, the integration of physical and material data characteristic of the extrusion process (temperatures, pressures, speeds, etc.) contributes to the considerable complexity of techniques aimed at predicting tread profiles at the extruder outlet.

[0015] However, the possibility of real-time prediction of a tread pattern at the extruder outlet, taking into account the aforementioned constraints, would not only facilitate the design of new, even more efficient tire products, but would also allow for real-time regulation of any deviation out of tolerance from the theoretical profile in existing manufacturing lines.

[0016] Description of the invention

[0017] In this regard, a set of algorithms and technical choices is proposed, according to the implementation methods defined below, to satisfy all the needs expressed above, in particular by means of a recursive time horizon prediction algorithm combined with artificial intelligence (i.e. a machine learning model) trained to make profile predictions, usable in real time.

[0018] In some implementations, the proposed technique uses a fixed-window recurrent prediction system adapted to tread profiles that do not have a single length or identical landmarks. The technique can thus process any profile and revert to a flexible, fixed-length representation. In this respect, an iterative sliding window is used, where each window contains the point of interest to be predicted and contextual data for the theoretical profile and the tread blade.

[0019] Thus, according to one aspect of the invention, a computer-implemented method is proposed for developing a machine learning model for predicting a tire product profile at the exit of an extruder as a function of contextual parameters of the extruder, comprising:

[0020] - obtaining training profiles, each comprising a real tire product profile measured at the outlet of an extruder and at least one contextual profile of the respective extruder; - a decomposition of the training profiles into iterations of a sliding window traversing each training profile over a set of iterations, each iteration of the window being centered on a point of interest and including at least one point behind the point of interest and at least one point in front of the point of interest, in the real profile and in said at least one contextual profile;

[0021] - training a self-supervised machine learning model to predict the point of interest in the real profile, taking as input to this model the iterations of the sliding window on the training profiles.

[0022] For example, said at least one point behind the point of interest and said at least one point in front of the point of interest are the points located immediately behind the point of interest and immediately in front of the point of interest.

[0023] According to one implementation method, each training profile includes the following contextual profiles of the extruder: an extruder blade profile and a theoretical profile of the pneumatic product.

[0024] According to one implementation method, each contextual profile also includes an identification of the corresponding positions of the points of interest in the blade profile and in the theoretical profile.

[0025] Depending on the implementation method, each training profile also includes machine parameters of the extruder.

[0026] According to one implementation method, each training profile further includes an identification of the corresponding positions of the points of interest in the actual profile and in said at least one contextual profile of the extruder.

[0027] According to one implementation method, the process further includes:

[0028] - a transformation of the training profiles, initially in a point cloud format representing a discrete trace of the profiles, into respective images representing a continuous trace of the respective profiles; and in which:

[0029] - said self-supervised machine learning model is a computer vision algorithm for image prediction, and the training of this model takes as input the sets of iterations of the sliding window on the images of the training profiles.

[0030] According to another aspect of the invention, a computer-implemented method is also proposed for predicting a tire product profile at the exit of an extruder based on contextual parameters of the extruder, comprising:

[0031] - obtaining contextual parameters of the extruder including at least one contextual profile;

[0032] - a decomposition of said at least one contextual profile into iterations of a sliding window traversing said at least one contextual profile over a set of iterations, each iteration of the window being centered on a point of interest and including at least one point behind the point of interest and at least one point in front of the point of interest;

[0033] - the use of a learning model trained to predict a point of interest of a real profile as a function of an iteration window of at least one contextual profile centered on the point of interest, recursively on said iteration set so as to obtain a succession of predictions of point of interest forming a prediction of the entire real profile.

[0034] Advantageously, the learning model trained to predict a real profile interest point based on an iteration window of at least one contextual profile centered on the interest point, is developed by the model development process defined above.

[0035] According to one implementation method, said obtaining of said at least one contextual profile includes obtaining the following contextual profiles: an extruder blade profile and a theoretical profile of the pneumatic product.

[0036] According to one implementation method, said contextual parameters further include an identification of the corresponding positions of the points of interest in the blade profile and in the theoretical profile.

[0037] According to one implementation method, said contextual parameters also include machine parameters of the extruder.

[0038] According to one implementation method, the process further includes:

[0039] - a transformation of said at least one contextual profile, initially in a point cloud format representing a discrete trace of said at least one profile, into respectively at least one image representing a continuous trace of said at least one profile; and wherein:

[0040] - said learning model is a computer vision algorithm for image prediction, trained to predict a continuous portion of a real profile centered on a point of interest within an iteration window.

[0041] According to another aspect, a method for real-time adjustment of extruder settings is also proposed in the event of a defect detected in a tire product profile exiting the extruder, comprising a series of implementations of the tire product profile prediction process at the extruder exit as defined above, by varying at least one of the contextual parameters of the extruder and an identification in the series of a prediction of a suitable integer real profile, the adjustment of the extruder settings being made according to the contextual parameters of the prediction of the suitable integer real profile.

[0042] Brief description of the drawings

[0043] Other advantages and features of the invention will become apparent upon examination of the detailed description of implementation methods and the accompanying drawings, in which the figures are shown:

[0044] [Fig. l] previously described, illustrates a schematic cross-sectional view of an example of a known tire;

[0045] [Fig.2A] illustrates an example of a method for developing a machine learning model for predicting a tire product profile at the output of an extruder;

[0046] [Fig.2B] illustrates an example of a method for predicting a tire product profile at the output of an extruder;

[0047] [Fig. 2C] illustrates an example of a real-time adjustment method for an extruder;

[0048] [Fig.3] illustrates an example of profile decomposition into sliding window iterations;

[0049] [Fig.4] illustrates an example of identifying corresponding positions in a blade profile and in a theoretical profile; [Fig.5] illustrates examples of predictions of real profiles.

[0050] Detailed description

[0051] Figure 2A illustrates an example of method 210 for developing a machine learning model for predicting a tire product profile at the output of an extruder, for example a tread, as a function of contextual parameters of the extruder.

[0052] The process 210 is implemented by computer, and can for example be materialized by a computer program product comprising instructions which, when the program is executed by a computer, lead the latter to implement the process 210 described below; or by a computer-readable recording medium comprising instructions which, when executed by a computer, lead the latter to implement the process 210 as described below.

[0053] The process 210 includes a step 211 of obtaining PE drive profiles (see figure 3) each comprising a real profile PR of tire product measured at the outlet of an extruder and at least one contextual profile PL, PT of the extruder respectively for each real profile.

[0054] For example, each PE drive profile includes the following two contextual extruder profiles: an extruder blade profile PL and a theoretical profile PT of the pneumatic product obtained by this extrusion blade. Advantageously, each PE drive profile can also include extruder machine parameters PrmMch.

[0055] Training profiles, for example, are derived from a DAT database acquired through practical experience in factories and belonging to a know-how specific to each industrial production line.

[0056] The machine parameters of the PrmMch extruder include, for example, the physicochemical operating conditions of the extruder, such as temperature, pressure, and extrusion flow rate. These PrmMch machine parameters can be measured by sensors, relative to each drive profile (actual profile PR, blade profile PL, and theoretical profile PT) in the DAT database. The selection of physical parameters used in implementing processes 210 and 220 (see Figure 2B) will not be detailed here, and may, for example, be based on an exploratory analysis of data previously collected on the extruder.

[0057] The process 210 includes a step 213 of decomposing the training profiles into iterations of a sliding window with a fixed horizon.

[0058] In this regard, we refer to figure 3.

[0059] Figure 3 illustrates an example of decomposition of a PE drive profile, comprising a blade profile PL, a theoretical profile PT associated with this blade, and the actual profile PR of an extrusion implemented for example according to given machine parameters PrmMch.

[0060] Successive iterations of the sliding window FGi, 0 <i<N, parcourent le profil d’ entrainement, entièrement parcouru après un ensemble d’itération respectif FGtot, dépendant de la taille du profil, c’ est-à-dire par exemple de la largeur de la bande de roulement.

[0061] Each iteration of the FGi window is centered on an interest point Pi and includes at least one point behind Pi-1 the interest point (or for example a number k of points behind the interest point Pi-k... Pi-1), and at least one point in front of Pi+1 the interest point (or for example a number k' of points in front of the interest point Pi+1... Pi+k'), in the actual profile PR and in the contextual blade profiles PE and theoretical PT.

[0062] Figure 3 shows that variations which are difficult to predict a priori, for example the extent and amplitude of a lateral bead in the actual PR profile in relation to an oblique part of the PE extrusion blade (on the left in the figure), can have a significant impact on the actual PR profile compared to the theoretical PT profile, and generate a lateral shift in the positions of grooves (uneven hollow parts) and ribs (protruding parts).

[0063] Furthermore, the same type of variations, the identification of which is not trivial because, for example, they are located at different abscissa positions, can exist between the points of the blade profile PL and the points of the theoretical profile PT.

[0064] Thus, in the PE training data, it will be advantageous to provide an identification of the corresponding positions of the points in the blade profiles PL and in the respective theoretical profiles PT.

[0065] In this regard, we refer to figure 4.

[0066] Figure 4 illustrates an example of additional information, or "metadata" R(PR,PT) enabling the identification of corresponding positions of points of interest Pi in a blade profile PL and in a respective theoretical profile PT.

[0067] Thus, for each point of interest Pi- 1 , Pi, Pi+ 1 of the blade profile PL, a respective relation Ri- 1 , Ri, Ri+ 1 allows us to identify the corresponding position of the point of interest Pi- 1 , Pi, Pi+ 1 in the theoretical profile PT.

[0068] Conversely, for each point of interest Pi- 1 , Pi, Pi+ 1 of the theoretical profile PT, the respective relation Ri- 1 , Ri, Ri+ 1 allows us to identify the corresponding position of the point of interest Pi- 1 , Pi, Pi+ 1 in the blade profile PL.

[0069] We refer again to figure 2A.

[0070] The process 210 further includes a step 215 of training a self-supervised machine learning model to predict the interest points Pi of the real profile PR, taking as input to this model the iterations of the sliding window LGi of the training profile PE. For example, the training can be done recursively so as to traverse the set of iterations LGtot (positive result 'y' of the test 0 <i<N ; incrémentant i+1 ) de chaque profil d’ entrainement PE (résultat négatif « n » du test 0<i<N).

[0071] In other words, training step 215 can be implemented recursively RCR for each LGi window of each set of LGtot iterations, where the target value is given by the Pi interest point of the real profile in the current LGi window.

[0072] The training is configured to learn to predict the target value, i.e. the position of the interest point Pi in the actual profile PR, for example: as a function of the interest point Pi in the contextual blade profiles PL and theoretical PT; as a function of the k points (at least one, k>l) behind the interest point Pi-k...Pi-l in the contextual blade profiles PL and theoretical PT; as a function of the k' points (at least one, k'>l) ahead of the interest point Pi+1...Pi+k' in the contextual blade profiles PL and theoretical PT; and possibly as a function of k” points (at least one, k”>l) behind the interest point Pi-k”...Pi-l in the actual profile PR; and also advantageously as a function of the machine parameters PrmMch associated with each training profile PE.

[0073] This implementation example corresponds to a point-by-point training in predicting the real profile, the whole of which allows obtaining the prediction of the real profile by a discrete point-to-point plot, in a graphic data format of the point cloud type.

[0074] In an advantageous implementation mode, the self-supervised machine learning model is a computer vision algorithm for image prediction, and is trained to generate an image with a continuous trace of the predicted real profile.

[0075] In this regard, process 210 further includes a step 212 of transformation of the PE training profiles, initially in a graphic data format of the point cloud type of the PL, PT, PR profiles, into respective images representing a continuous trace of the respective profiles.

[0076] And so, the computer vision model training 215 for image prediction takes as input, recursively RCR, the sets of iterations of the sliding window on the images of the PE training profiles.

[0077] Transforming the problem of predicting graphical data in a point cloud into a computer vision problem facilitates the processing of the solution by training powerful commercially available tools, such as self-supervised machine learning models for computer vision image prediction. We now refer to Figure 2B.

[0078] Figure 2B illustrates an example of method 220 for predicting a tire product profile at the output of an EXTRD extruder as a function of contextual parameters of the EXTRD extruder, for example by means of the machine learning model developed by method 210 described previously in relation to Figure 2A.

[0079] The 220 process is also implemented by computer, and can for example be materialized by a computer program product containing instructions which, when the program is executed by a computer, lead the latter to implement the 220 process described below; or by a computer-readable recording medium containing instructions which, when executed by a computer, lead the latter to implement the 220 process as described below.

[0080] The process 220 includes a step 221 of obtaining contextual parameters of the extruder EXTRD, i.e. at least one contextual profile of the extruder, for example the blade profile PL equipping the extruder, and the theoretical profile PT of the pneumatic product formed by the extrusion.

[0081] Contextual parameters can also include machine parameters of the PrmMch extruder.

[0082] Obtaining the contextual blade profiles PL and the theoretical profile PT can be part of the usual business knowledge and know-how related to a given industrial production line.

[0083] In addition, the contextual parameters may include an identification R(PL,PT) of the corresponding positions Ri of the points of interest Pi in the blade profile PL and in the theoretical profile PT, as described previously in relation to Figure 4.

[0084] The process 220 includes a step 223 of decomposing said at least one contextual profile, according to the same principle as in the process 210 of model development, and for which reference can be made to the description made previously in relation to figure 3.

[0085] Thus, the decomposition of the profiles is done recursively (RCR) in iterations of a fixed-size sliding window EGi. Successive iterations traverse the entire contextual profiles over a set of iterations FGtot. Each iteration of the window FGi is centered on an interest point Pi and includes at least one point behind the interest point Pi-k...Pi-1 and at least one point before the interest point Pi+1...Pi+k'.

[0086] The process 220 includes, for example recursively for each iteration window FGi of the set, the use 225 of a learning model trained to predict an interest point Pi of the real profile PR, based on the contextual profiles PL, PT in the window FGi. Over the set of iteration windows, one can thus obtain a succession of predictions of interest point Pi forming a prediction of the entire real profile 226.

[0087] Here again, the learning model trained to predict the real PR profile can advantageously be a computer vision algorithm adapted for image prediction, and the method includes in this respect a transformation of the contextual blade profiles PL and theoretical PT, initially in a point cloud format to a respective image format, each representing a continuous trace of the respective contextual profiles.

[0088] In this case of a computer vision model for image prediction, each iteration i thus includes a decomposition 223 of the images of the contextual profiles PL, PT, into respective image windows FGi.

[0089] Each image window is centered at a position of interest Pi (or point of interest) on the continuous trace, and includes a portion of the trace located behind the point of interest, and a portion of the trace located in front of the point of interest. The image windows of the contextual blade profile PL and the theoretical contextual profile PT have a fixed and constant size across all iterations.

[0090] The prediction step 225 of the actual profile is configured to generate a portion of a continuous trace of the actual profile PR in a corresponding image window of fixed size and centered at the point of interest Pi, for each iteration, for example: as a function of the point of interest Pi in the contextual slice profiles PL and theoretical slice profiles PT; as a function of the portion of the continuous trace located behind the point of interest in the image windows of the contextual slice profiles PL and theoretical slice profiles PT (e.g., the portions of the continuous traces corresponding to the k points Pi-k...Pi-l in Figure 3); as a function of the portion of the continuous trace located in front of the point of interest in the image windows of the contextual slice profiles PL and theoretical slice profiles PT (e.g., the portions of the continuous traces corresponding to the k' points Pi+l...Pi+k' of figure 3); possibly as a function of a portion of the continuous trace located behind the point of interest in the image window of the real profile predicted during the previous iterations, i.e. for example an average, a summation or a superposition of the image windows of the real profile PR of the previous iterations; and also possibly in addition as a function of the machine parameters PrmMch associated with the implementation of the process 220.

[0091] We now refer to figure 5.

[0092] Figure 5 illustrates examples of image windows of predictions of portions of real profiles PPa, PPb, PPc, for example each obtained during an iteration of step 225, from the image window of the contextual blade profile PLA, PLb, PLc respective, and with respect to the real profiles PRa, PRb, PRc respective measured at the extruder outlet.

[0093] We can thus see that the results obtained PPa, PPb, PPc in each image window are extremely close to the real case measured PRa, PRb, PRc.

[0094] We refer again to figure 2B.

[0095] When all the image windows of the actual profile are generated (completion "end" of the "while" loop) 0 <i<N), on obtient la prédiction du profil réel entier 226.

[0096] For example, the entire profile can be "assembled" by juxtaposing the image windows predicted at each iteration, and / or by overlapping the predicted image windows on the common parts of the profile, possibly with the implementation of weighting or averaging for any inequalities in the tracings. In summary, we have described the development 210 and the use of a process 220 that makes it possible to predict the shape of the actual profile of a tire product, for example a tread, as it exits an extruder, using the theoretical target profile, the shape of the machined blade, and the machine parameters imposed on the extruder.

[0097] The prediction process relies in particular on a system for capturing physical data (thermal, pressure, speed sensors, etc.) from an industrial production line of the tire product; on a database of machined profiles and theoretical profiles (know-how); and on a self-training model with elements from the sensors and the database in order to predict the entire real profile.

[0098] The prediction process makes it possible in particular to avoid problems related to variable profile lengths, due to the use of a fixed-size window to treat each profile as an ordered succession of small profiles of fixed size.

[0099] Furthermore, the processing of profile prediction by implementing a computer vision model allows for ease of operation and development thanks to powerful self-supervised learning tools.

[0100] We now refer to figure 2C.

[0101] Figure 2C illustrates an example of the application of process 220 in a method 230 of fine corrective adjustment, in real time, of the operation of an EXTRD extruder.

[0102] Indeed, the real-time adjustment method 230 of the EXTRD extruder can be implemented in the industrial production line 101-103, for example as part of a quality control 231 of the tire product 102 coming out of the EXTRD extruder.

[0103] The industrial production line is schematically represented by a section 101 upstream of the extruder EXTRD, the tire product 102 (for example, a tread) exiting the extruder, the quality control stage 231, and a section 103 downstream of the extruder where the tire product will be assembled with other components and, for example, undergo curing. The quality control stage 231 is typically implemented occasionally and not systematically in the production line, and the downstream stages 103 are implemented if no defect is detected (negative output "n" of the defect detection test 231).

[0104] In the event of a fault detection (positive output "y" of the fault detection test 231) in the profile of the tire product 102, the prediction process 220 described above in relation to Figure 2B is used to obtain a new machine setting 235 capable of correcting the detected fault, very quickly and without disrupting the production line 101-103.

[0105] Indeed, method 230 includes a series 2260 of implementations of the tire product profile prediction process, allowing to obtain a series, for example a number M, of predictions of whole profiles 226_1 , 226_2, ... , 226_M.

[0106] Each implementation of the process 226_1, 226_2, ..., 226_M is done by varying at least one of the contextual parameters of the extruder PrmMch l, PrmMch_2, ..., PrmMch M; advantageously the machine parameters PrmMch, possibly the blade profile PL.

[0107] Thus, each prediction of whole profiles PP I , PP_2, ... , PP M shows the effect on the pneumatic product 102 of the respective setting of the contextual parameter PrmMch l , PrmMch_2, ... , PrmMch M.

[0108] Method 230 then includes an identification step 233 of the most suitable profile PP_j in the series of predictions of the real integer profiles PP I , PP_2, ... , PP M; that is to say the profile PP_j which corrects the defect detected in step 231 (output "ok" of the identification test 233).

[0109] This identification 233 can be done for example by means of classical techniques of quantification of errors or difference between two images, with respect to a target profile, such as for example the theoretical PT profile.

[0110] The PP_j profile was generated based on the respective PrmMch_j contextual parameters. The EXTRD extruder setting 235 can be made using the PrmMch_j contextual parameters employed in predicting the appropriate integer real-world PP_j profile.

[0111] Thus, it is possible to adjust in real time, that is to say without constraining delay from the point of view of the industrial production chain (which is the case of classical computer simulation techniques by finite elements).

[0112] This provides the advantage of being able to increase the frequency of quality controls and refine the thresholds for triggering defect corrections, and thus increase the overall quality of products coming off the line.

[0113] From another point of view, this also provides the advantage of greatly reducing production line interruption times related to quality controls and corrective adjustments, and thus increasing the overall quantity of products coming off the line.

Claims

DEMANDS 1. A computer-implemented method (210) for developing a machine learning model for predicting a tire product profile exiting an extruder as a function of contextual extruder parameters, comprising: - obtaining (21 1 ) training profiles (PE) each comprising a real profile (PR) of tire product measured at the outlet of an extruder and at least one contextual profile (PL, PT) of the respective extruder; - a decomposition of the training profiles (213) into iterations of a sliding window (FGi) traversing each training profile over a set of iterations (FGtot), each iteration of the window being centered on a point of interest (Pi) and including at least one point behind (Pi- 1 ) the point of interest and at least one point in front of (Pi+ 1 ) the point of interest, in the real profile (PR) and in said at least one contextual profile (PL, PT); - a training (215) of a self-supervised machine learning model to a prediction of the point of interest (Pi) in the real profile (PR), taking as input to this model the iterations of the sliding window (FGi) on the training profiles.

2. Method according to claim 1, wherein each drive profile (PE) comprises the following contextual profiles of the extruder: an extruder blade profile (PL) and a theoretical profile (PT) of the pneumatic product.

3. Method according to claim 2, wherein each contextual profile further includes an identification of the corresponding positions (R(PL,PT)) of the points of interest (Pi) in the blade profile (PL) and in the theoretical profile (PT).

4. A method according to any one of claims 1 to 3, wherein each drive profile further comprises extruder machine parameters (PrmMch).

5. A method according to any one of claims 1 to 4, further comprising: - a transformation (212) of the training profiles (TP), initially in a point cloud format representing a discrete trace of the profiles, into respective images representing a continuous trace of the respective profiles; and in which: - said self-supervised machine learning model is a computer vision algorithm for image prediction, and the training (215) of this model takes as input the iteration sets (RCR) of the sliding window on the images of the training profiles.

6. A computer-implemented method (220) for predicting a tire product profile at the exit of an extruder as a function of contextual extruder parameters (EXTRD), comprising: - obtaining (221) contextual parameters of the extruder (EXTRD) including at least one contextual profile (PL, PT); - a decomposition of said at least one contextual profile (223) into iterations of a sliding window (FGi) traversing said at least one contextual profile over a set of iterations (FGtot), each iteration of the window (FGi) being centered on an interest point (Pi) and including at least one point behind the interest point (Pi- 1 ) and at least one point in front of the interest point (Pi+ 1 ); - the use of a learning model (225) trained to predict a point of interest (Pi) of a real profile (PR) as a function of an iteration window (FGi) of at least one contextual profile (PL, PT) centered on the point of interest (Pi), recursively (RCR) on said iteration set so as to obtain a succession of predictions of point of interest forming a prediction of the entire real profile (226).

7. Method according to claim 6, wherein said obtaining of said at least one contextual profile (221) comprises obtaining the following contextual profiles: a blade profile (PL) of the extruder and a theoretical profile (PT) of the pneumatic product.

8. A method according to claim 7, wherein the contextual parameters further comprise position identification corresponding (Ri) points of interest (Pi) in the blade profile (PL) and in the theoretical profile (PT).

9. A method according to any one of claims 6 to 8, wherein said context parameters further comprise extruder machine parameters (PrmMch).

10. A method according to any one of claims 6 to 9, further comprising: - a transformation (222) of said at least one contextual profile (PL, PT), initially in a point cloud format representing a discrete trace of said at least one profile, into respectively at least one image representing a continuous trace of said at least one profile; and wherein: - said learning model (225) is a computer vision algorithm for image prediction, trained to predict a continuous portion of a real profile centered on a point of interest in an iteration window. 1 1. Method for real-time adjustment (230) of an extruder (EXTRD) in case of a detected defect (231) in a tire product profile (102) exiting the extruder, comprising a series of implementations of the method for predicting the tire product profile exiting the extruder (2260) according to any one of claims 6 to 10, by varying at least one of the contextual parameters of the extruder (PrmMch l , PrmMch_2, ... , PrmMch M) and an identification in the series of a prediction of a suitable integer real profile (233), the adjustment of the extruder settings being made (235) according to the contextual parameters (PrmMch_j) of the prediction of the suitable integer real profile.

Images