Quality control of a physical element

Abstract

Examples include a data processing method, a computer program product containing instructions for implementing the method, a data processing device configured for implementing the method, and a data processing system comprising such a data processing device. Abstract figure: [Fig. 3]

Description

Title of the invention: Quality control of a physical element technical field

[0001] This disclosure relates to the field of quality control. Previous technique

[0002] Many industrial sectors assess the quality of the parts they produce before placing them on the market, particularly to ensure that these parts perform their function correctly. However, quality control processes for parts can be costly, especially since they regularly require the intervention of a technician.

[0003] The present disclosure improves this situation. Summary

[0004] In this regard, a computer-implemented method is proposed for processing data relating to a physical element, in particular a manufactured product, the method comprising: obtaining at least one piece of data characterizing the physical element; a determination of a score representative of a state of the physical element based on the processing, by a first machine learning model, of at least one piece of data obtained; and a classification of the physical element in a classification representative of a state of the physical element from the determined score; in which the representative classification of the state of the physical element comprises three classes: a first class in which the physical element is defined as conforming, a second class in which the physical element is defined as conforming subject to performing at least one recompliance operation on the physical element belonging to a determined list of recompliance operations, and a third class in which the physical element is defined as non-conforming.

[0005] Optionally, prior to obtaining at least one data characterizing the physical element, the method further includes: receiving at least one measurement signal of the physical element; a preprocessing of a measurement signal in order to obtain a data point, at least one data point characterizing the physical element.

[0006] Optionally, at least one data characterizing the physical element includes data relating to a contour of the physical element on an image of the physical element; at least one measurement signal includes an image of the physical element; and the preprocessing of a measurement signal in order to obtain data of at least one data characterizing the physical element includes a detection of the contours of the physical element in the image.

[0007] Optionally, at least one data characterizing the physical element includes data relating to a dimension of the physical element; at least one measurement signal includes an image or point cloud of the physical element; and the preprocessing of a measurement signal in order to obtain a data of at least one data characterizing the physical element includes a determination of a dimension of the physical element from the image or point cloud.

[0008] Optionally, the method further comprises: a determination of a heat map by the first machine learning model, from at least one data characterizing the physical element; the heat map associating a plurality of coordinates of the physical element being treated with a respective intensity value representing a probability of defect.

[0009] Optionally, the method further comprises, when the physical element is classified in the second class: a classification of at least one defect in the physical element by a second machine learning model, the second machine learning model receiving as input the heat map determined by the first machine learning model; and a determination of a physical element compliance score by the second machine learning model.

[0010] Optionally, the second machine learning model also receives as input a binary compliance signal; and in which the compliance score is equal to the binary compliance signal.

[0011] Optionally, the classification of at least one defect of the physical element may include a classification of each defect of the physical element into an acceptable defect class or into a critical defect class; the acceptable defect class grouping defects belonging to a plurality of defect types, in which each defect type is associated with a correction operation belonging to the determined list of remediation operations; and the critical defect class corresponding to a class grouping the other defects.

[0012] Optionally, the method may further include determining a correction factor using a third machine learning model, the third machine learning model receiving as input at least one piece of data characterizing the physical element, the classification of at least one defect in the physical element, and the remediation score; and a modification of the first machine learning model based on the correction factor.

[0013] The application also relates to a data processing device configured for the implementation of any of the processes presented in this disclosure.

[0014] The application also relates to a data processing system relating to a physical element, in particular a manufactured product, comprising: at least one sensor configured to acquire at least one measurement signal of the physical element; a data processing device configured to implement any of the processes described in this disclosure; in which at least one data characterizing the physical element is obtained from at least one measurement signal of the physical element

[0015] The application further relates to a computer program product comprising instructions for implementing any of the processes presented in this disclosure when that program is executed by a processor.

[0016] Finally, the application relates to a non-transient computer-readable recording medium on which is recorded a program for the implementation of any of the processes presented in this disclosure when this program is executed by a processor. Brief description of the drawings

[0017] Other features, details and advantages will become apparent upon reading the detailed description below, and upon analysis of the accompanying drawings, on which:

[0018] [Fig.1] schematically represents an example of a data processing device.

[0019] [Fig.2] schematically represents an example of a processing system data.

[0020] [Fig.3] schematically represents an example of a treatment process for data.

[0021] [Fig.4] schematically represents an example of a software architecture of a first machine learning model.

[0022] [Fig.5] schematically represents an example of the classification of an element physical.

[0023] [Fig.6] schematically represents another example of a treatment process data.

[0024] [Fig.7] schematically represents yet another example of a process of data processing.

[0025] [Fig.8] schematically represents yet another example of a process of data processing.

[0026] [Fig.9] schematically represents yet another example of a process of data processing.

[0027] [Fig. 10] schematically represents yet another example of a process of data processing.

[0028] [Fig. 11] schematically represents another example of a software architecture of the first machine learning model.

[0029] [Fig. 12] schematically represents yet another example of a process of data processing.

[0030] [Fig. 13] schematically represents yet another example of a process of data processing.

[0031] [Fig.14] schematically represents an example of the classification of a defect of a physical element.

[0032] [Fig. 15] schematically represents yet another example of a process of data processing.

[0033] [Fig. 16] schematically represents an example of a software architecture of the first machine learning model and a second machine learning model.

[0034] [Fig. 17] schematically represents another example of an architecture software of the first machine learning model and the second machine learning model.

[0035] [Fig. 18] schematically represents yet another example of a process of data processing.

[0036] [Fig. 19] schematically represents an example of a software architecture of the first machine learning model, second machine learning model and third machine learning model. Description of the implementation methods

[0037] The inventors observed that partially automated quality control processes required, for each part whose quality was determined by the automated part of the process to be non-conforming, a subsequent check by a quality control operator to determine whether that part could be brought into conformity or should be definitively considered non-conforming. Therefore, although an initial sorting of the parts is carried out automatically, a second sorting by a quality control operator is necessary to determine which Parts initially deemed non-compliant by the automated process can be brought into compliance.

[0038] This disclosure proposes a solution that at least reduces, or even eliminates, the intervention of the quality control operator in determining which parts, and more broadly which physical elements to be analyzed, can be brought into conformity. In particular, this disclosure proposes a solution for automatically classifying the elements to be analyzed as conforming, the elements to be analyzed as non-conforming, and the elements to be analyzed as being able to be brought into conformity based on a score representative of a part condition determined by a machine learning model.Furthermore, since this classification is performed automatically, the variability in the quality of products within the same class is reduced, thereby reducing, or even eliminating, the subjectivity of the control process compared to quality control carried out by inspectors, where quality control differs from one inspector to another. In other words, the proposed solution offers a much more consistent and reliable classification than one performed by human intervention.

[0039] A physical item can be defined in this disclosure as any tangible item whose quality can be assessed for quality control purposes. A tangible item is distinct from digital data, which is intangible. Therefore, a physical item as defined in this disclosure does not correspond to a set of data stored in memory.

[0040] A physical element according to this disclosure may, in particular, correspond to a raw material or a product. The product may, in particular, be an intermediate product or a finished product.

[0041] A raw material can be defined as a natural or unprocessed material or substance intended for artisanal or industrial transformation. A raw material can, for example, be a metal, a metal alloy, petroleum, wood, sulfur, salt, coal, a gas, an agricultural raw material, etc.

[0042] An intermediate product can be defined as a product or material derived from raw materials after one or more transformations, and for which at least one further transformation is necessary before a finished product is obtained. In other words, an intermediate product is a product unsuitable for sale to the end consumer. An intermediate product could, for example, be granulated plastic, which is a petroleum derivative. An intermediate product could also be steel (derived from iron), fabric (derived from cotton), or paper pulp (derived from wood).

[0043] A finished product can be defined as a product derived from a raw material or an intermediate product following an industrial or artisanal process and intended for sale to the end consumer. A finished product may, in particular, correspond to a manufactured product, such as a machined part.

[0044] An example of a data processing device 1 enabling the implementation of any of the example methods 100 according to this disclosure is now described with reference to [Fig. 1].

[0045] The data processing device 1 includes a processor 11 adapted to implement any one of the example methods according to this disclosure. The processor 11 may, for example, be of the microprocessor, microcontroller, FPGA, etc. type.

[0046] The data processing device 1 also includes a memory 12 for storing the code instructions executed by the processor. The memory 12 may, for example, include ROM (Read-Only Memory), RAM (Random Access Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), or any other suitable type of storage medium. The memory 12 may, for example, include optical, electronic, or magnetic storage. In particular, the memory 12 may store code instructions enabling the processor 11 to execute any of the example methods described in this disclosure.

[0047] The processor 11 can access information in memory 12, whether it be code instructions, or other data, in particular acquired by different sensors (presented below) which could be stored in memory 12.

[0048] With reference to [Fig.2], an example of a data processing system 10 is also presented, enabling the implementation of any of the example methods 100 according to this disclosure.

[0049] The data processing system 10 may include a data processing device 1 configured for the implementation of any of the example methods 100 presented in this document.

[0050] The data processing system 10 may also include at least one sensor 2 configured to acquire at least one measurement signal of the physical element.

[0051] In examples, the data processing system 10 may include at least one sensor 2 from among the following types of sensors: a type of sensor configured to acquire an image associating two-dimensional coordinates with an intensity value; a type of sensor configured to acquire a point cloud associating three-dimensional coordinates with an intensity value; a type of sensor configured to acquire an electrical signal associating an electrical signal level with a value of a physical quantity.

[0052] The data processing system 10 may include several sensors 2 belonging to the same type.

[0053] In examples, a type of sensor configured to acquire an image may include: an image sensor configured to acquire at least one image in a predetermined color space; an ultraviolet image sensor; and an infrared image sensor.

[0054] A predetermined color space may, for example, correspond to an RGB color space (from the English for "Red Green Blue", also called RGB for "Red Green Blue") or to an HSL color space (from "Hue, Saturation, Lightness", also called HSL in English for "Hue, Saturation, Lightness"). Other color spaces can obviously be considered.

[0055] In examples, a type of sensor configured to acquire a point cloud associating three-dimensional coordinates with an intensity value may correspond to a radar sensor, or to a lidar (Light Detection And Ranging) sensor.

[0056] In examples, a type of sensor configured to acquire an electrical signal associating an electrical signal level with a value of a physical quantity may, for example, include: a pressure sensor; a temperature sensor; a conductivity sensor (thermal or electrical); a magnetic field sensor; a sound sensor (for example in the audible range or in the ultrasound range); a chemical composition sensor (e.g., a humidity sensor); and a density sensor.

[0057] It is therefore understood that the data acquired by the sensors, whether images, point clouds, or electrical signals representing physical quantities, can be used, alone or in combination with each other, to characterize the physical element.

[0058] Other examples of sensors 2 than those presented may possibly be considered depending on the physical element to be considered. It is also understood that depending on the physical element to be considered, some sensors will be more relevant than others.

[0059] The data processing device 1 of the data processing system 10 can be configured to process a measurement signal acquired by at least one sensor in order to obtain data belonging to at least one data point D characterizing the physical element. As developed later, the method examples 100 process the at least one data point D characterizing the physical element in order to, in particular, classify that physical element.

[0060] In examples, the data processing system 10 may include a screen (not shown).

[0061] An example of a data processing method 100 relating to a physical element is now presented with reference to [Fig. 3]. The method 100 can be implemented by computer. In particular, it can be implemented to perform quality control of the physical element. This method is presented for a single physical element, but it can obviously be implemented for each physical element of a plurality of similar physical elements (for example, a series of the same raw material or the same product). These physical elements can, for example, be transported successively by suitable means to a predetermined area that allows the acquisition of data used for the implementation of the method 100, in particular data acquired by one or more sensors 2. In particular, the method 100 is especially suitable for implementation to assess the quality of products coming off a production line.

[0062] It should be noted that the figures associated with the process examples 100 are merely illustrations of the process examples 100, representing, by means of blocks, the various operations that may be included in the process and described later in this document. As such, the illustrations do not convey any sequence between the operations. In other words, the operations described with reference to the figures are not necessarily carried out sequentially and may, in particular, be carried out in a different order than that shown in the figures, or be carried out in parallel, unless a given operation requires a result from another operation to be carried out. Similarly, it is not necessary for each operation to be carried out once before the same operation is repeated a second time.The frequency with which each operation is implemented is specific to it and is not necessarily linked to the implementation of other operations.

[0063] As illustrated by block 110, the method 100 comprises an operation for obtaining at least one data point D characterizing the physical element. The at least one data point D can be obtained from at least one measurement signal acquired by at least one sensor. Advantageously, the at least one data point D comprises a plurality of data determined from the measurement signals measured by each of the 2 sensors of the data processing system.

[0064] In examples, the at least one data D may include data obtained directly from a measurement signal acquired by a sensor, i.e. without processing the measurement signal from that sensor.

[0065] In examples, the at least one data D may include data obtained from a preprocessing on a measurement signal acquired by a sensor, as will be detailed later.

[0066] As illustrated by block 120, the method 100 includes an operation for determining a score Se representative of a state of the physical element. The score Se representative of the state of the physical element is determined from a processing, by a first machine learning model NNb, of at least one data point D characterizing the physical element.

[0067] The first machine learning model NNi may in particular correspond to a first neural network, which may be a convolutional neural network. In these examples, an input layer of the first neural network may comprise a plurality of input neurons, and a piece of data belonging to the at least one data D characterizing the physical element may be associated with a specific input neuron or a group of specific input neurons.

[0068] In the operation illustrated by block 120, a first machine learning model NNi is used to determine the score Se associated with a state of the physical element. Consequently, it is understood that as the first machine learning model NNi is trained, the score Se determines the state of the physical element under consideration with increasing precision. In particular, the physical element may be part of a plurality of similar elements to be analyzed (a series of the same product, for example), and process 100 is implemented on each of the elements to be analyzed to train the first machine learning model NNb.

[0069] A first schematic example of a software architecture of the first machine learning model NNi is shown in [Fig.4]. In this example, at least one data point D is entered into the first machine learning model NN1 which outputs the score Se associated with a state of the physical element.

[0070] As illustrated by block 130, the process 100 comprises a classification operation of the physical element into a classification Ce representative of a state of the physical element based on the determined score. The classification Ce representative of the state of the physical element is schematically illustrated in [Fig. 5].

[0071] This classification, which is representative of the state of the physical element, comprises (or is made up of) three classes.

[0072] The classification Ce, representative of the state of the physical element, thus includes a first class Cl in which the physical element is defined as conforming. In this case, the physical element can be classified as conforming when it meets a plurality of predetermined conformity criteria. These conformity criteria depend on the physical element in question and can therefore be determined by a person skilled in the art based on said physical element. In particular, in the case of a machined part, a conformity criterion may, for example, correspond to the absence of any hollows exceeding a predetermined width or depth in a specific area. In the case of a material, a conformity criterion may, for example, correspond to a material moisture content below a first predetermined threshold and / or above a second predetermined threshold.These examples are given as a non-exhaustive list to show that the predetermined conformity criteria are directly linked to the physical element being processed.

[0073] The classification Ce, representative of the state of the physical element, also includes a second class C2 in which the physical element is defined as compliant subject to at least one re-conforming operation being performed on the physical element. The re-conforming operation belongs to a predetermined list of re-conforming operations.

[0074] The list of corrective actions to be determined can be established by a quality control agent based on the physical component in question. In particular, the list of corrective actions can be determined based on conformity criteria. A corrective action can thus be associated with one or more conformity criteria. Conversely, some conformity criteria may not be associated with any corrective action.

[0075] The classification Ce, representative of the state of the physical element, includes a third class C3 in which the physical element is defined as non-conforming. The physical element can be classified as non-conforming when it fails to meet one criterion from the plurality of determined conformity criteria and no operation from the predetermined list of operations can modify the physical element to meet that criterion.

[0076] In examples, a physical element can be classified as belonging to the first class Cl (compliant) when the score Se associated with its state is greater than a compliance threshold.

[0077] In examples, a physical element can be classified as belonging to the third class C3 (non-conforming) when the Se score associated with its state is less than a non-conforming threshold.

[0078] In examples, a physical element can be classified as belonging to the second class C2 (compliant subject to performing at least one operation of compliance) when the Se score associated with its state is between the non-compliance threshold and the compliance threshold.

[0079] The proposed method 100 thus makes it possible to classify each physical element into one of three classes based on a score determined by a machine learning model. It is therefore no longer necessary to involve a technician to distinguish, among the non-conforming physical elements, which ones can be brought into conformity and which ones cannot.

[0080] In this respect, the process described in this disclosure offers an ingenious way to classify products within a quality control process. In particular, whereas the partially automated quality control processes identified by the inventors previously only determined whether an item was conforming or not, this disclosure proposes a new form of automated classification that also determines which items can be brought into conformity. Specifically, although detailed examples below present a learning model supervised by one (or more) quality control operator(s), once the model reaches a certain level of learning, it is no longer necessary to involve an operator to assist in the classification of physical items by process 100.Therefore, even when the model is trained to reach this predetermined level of learning with the help of a quality control operator, the operator's intervention is limited to this stage.

[0081] It should also be noted that the proposed solution, insofar as it incorporates the know-how of the quality control operator(s) who train it, makes it possible to standardize the quality level of the classified products. In particular, the proposed solution makes it possible to reduce the variability in the quality of products classified within the same class, especially in classes C1 (compliant) and C2 (compliant subject to a corrective action), by reducing, or even eliminating, the subjectivity of quality control, which differs from one quality control agent to another.

[0082] Other operations may optionally be incorporated into Process 100 and are described later in this disclosure. These operations may be incorporated into Process 100 in combination with each other unless expressly stated otherwise in this disclosure.

[0083] In examples, the non-conformity threshold can be defined dynamically, for example from a plurality of iterations of process 100 on a corresponding plurality of physical elements.

[0084] In examples, the compliance threshold can be defined dynamically, for example from a plurality of iterations of process 100 on a corresponding plurality of physical elements.

[0085] In examples, the conformity threshold and / or the non-conformity threshold can be determined from the Se scores representing the respective state of a plurality of determined physical elements. As explained previously, process 100 is intended to be implemented on several similar physical elements (several products), for example from the same production line, so that the calculation of the Se score of these elements can be used for calculating the thresholds.

[0086] In examples, the conformity threshold can be determined from a plurality of scores Se determined for physical elements having been classified in the first class Cl and in the second class C2.

[0087] In examples, the non-conformity threshold can be determined from a plurality of scores Se determined for physical elements having been classified in the third class C3 and in the second class C2.

[0088] Calculating the non-conformity and conformity thresholds dynamically, based on a plurality of iterations of the process 100, allows the gap between the two thresholds to be refined as the first machine learning model NNi learns, so as to retain in this gap only the conforming physical elements subject to carrying out a determined re-conformity operation.

[0089] In examples, the first NNi machine learning model can be supervised by a quality control operator as long as the difference between the conformity threshold and the nonconformity threshold is greater than a predetermined value. In other words, when the difference between the conformity threshold and the nonconformity threshold falls below this value, the operator's intervention is no longer necessary. Indeed, the inventors have observed that when the difference between the thresholds is sufficiently small, the self-training of the first NNi machine learning model is efficient enough for it to stabilize at classification error rates much lower than those of a quality control operator.

[0090] In examples, the process 100 may further include operations prior to the operation 110 of obtaining at least one data D characterizing the physical element.

[0091] In these examples, the method 100 may thus include an operation 101 of receiving at least one measurement signal from the physical element. This at least one measurement signal may, in particular, be acquired by at least one sensor 2.

[0092] In process examples including operation 101, process 100 may further include an operation 102 of preprocessing a measurement signal in order to obtain a data of at least one data D characterizing the physical element.

[0093] As explained previously, the at least one data D may include data derived directly from the measurement signal acquired by a sensor and / or may include data derived from a preprocessing 102 of the measurement signal from a sensor. These examples are schematically represented in [Fig. 6].

[0094] In examples, the at least one data point D characterizing the physical element may, for instance, include data relating to a contour of the physical element in an image of the physical element. The image of the physical element may be acquired by an image sensor. Having access to data relating to a contour of the physical element may allow the first machine learning model NNi to detect a defect in the shape of the physical element being processed.

[0095] In examples where at least one data point D includes data relating to a contour of the physical element, at least one measurement signal received during operation 101 may include an image of the physical element. In these examples, the preprocessing operation 102 of a measurement signal to obtain data from the at least one data point D characterizing the physical element may include a contour detection operation 1020 of the physical element's contours in the image. These examples are schematically represented in [Fig. 7]. In particular, a preprocessing neural network or a contour detection algorithm may be used to detect the contours of the physical element. The preprocessing neural network may, for example, be a convolutional neural network, which is particularly well-suited for performing this contour detection operation 1020.

[0096] In examples, the at least one data element D characterizing the physical element may, for example, include data relating to a dimension of the physical element. Data relating to a dimension may, for example, correspond to the length of a side of a surface of the physical element. Having access to data relating to a dimension of the physical element may allow the first machine learning model NNi to detect a defect in the dimensions of the physical element being processed.

[0097] In examples where at least one data point D includes data relating to a dimension of the physical element, at least one measurement signal received during operation 101 may include an image or a point cloud of the physical element. The image may, for example, be acquired by an image sensor, while the point cloud may, for example, be acquired by radar or lidar. In these examples, the preprocessing operation 102 of a measurement signal to obtain a The data, at least one data point D characterizing the physical element, may include an operation to determine a dimension of the physical element from the image or point cloud. These examples are schematically represented in [Fig. 8].

[0098] In a first option, in which the determination of a dimension of the physical element is performed from a point cloud, the dimension of the physical element can be obtained from the difference between the coordinates of two specific points. In a second option, in which the determination of a dimension of the physical element is performed from an image, the dimension of the physical element can be obtained from the difference between the coordinates associated with two pixels of the image. It is also understood that the determination of a dimension of the physical element proposed in operation 1021 can be facilitated by the prior implementation of operation 1020, which determines the contours of this element.

[0099] In examples, the at least one data point D characterizing the physical element may, for example, include data relating to an acoustic response of the physical element to a sound signal. The sound signal may, for example, be acquired by a sound sensor. In some examples, the sound signal may be acquired by the sound sensor following a test impact on the physical element being processed, for example, following a test blow. The inventors have, for example, observed that a product exhibiting mechanical deformation (in particular an internal defect, or a crack, a crevice) produces an acoustic response that differs from the acoustic response of a conforming physical element. Having access to data relating to an acoustic response of the physical element may, for example, allow the first machine learning model NNi to detect these types of defects.

[0100] In a first option, the sound signal can directly correspond to a data point relating to the acoustic response of at least one data point D characterizing the physical element. In a second option, the sound signal acquired by the sound sensor can be received during operation 101, and the preprocessing operation 102 can include a selection operation 1022 of a portion of the sound signal corresponding to a frequency range of interest in the sound signal. In which case, a data point relating to an acoustic response of the physical element can correspond to the selected sub-portion of the sound signal. These examples are schematically represented in [Fig. 9]. The processing of this type

[0101] In examples, at least one data element D may include image data. Image data may, for example, allow the detection of defects in color, texture, or shape of the physical element.

[0102] The image data of at least one data point D may, for example, correspond to a selection of a specific part of the image. In examples, the process 100 includes receiving an image during operation 101 and preprocessing the image in an operation 102 so as to subdivide the image, for example, into a plurality of parts of identical size. In examples, a subdivision of the image may correspond to image data of at least one data point D.

[0103] In examples, the method 100 may further include an operation 140 of determining a heatmap Hm by the first machine learning model. These examples are schematically represented in [Fig. 10].

[0104] The heat map Hm associates a plurality of coordinates of the physical element being processed with a respective intensity value representing a probability of defect. The probability of defect at a coordinate of the physical element can, for example, be proportional (or alternatively inversely proportional) to the intensity value associated with that coordinate. The heat map Hm can, in particular, be limited to two dimensions to facilitate its display on a screen. Indeed, a three-dimensional heat map Hm would require the implementation of a projection step of the three-dimensional coordinates of the heat map into a two-dimensional pixelated space to allow its display on a screen. A two-dimensional coordinate of the heat map Hm can correspond to a pixel on the screen.

[0105] A second schematic example of the software architecture of the first machine learning model NNi is shown in [Fig. 11]. In this example, at least one data point D is fed into the first machine learning model NNi, which outputs the score Se associated with a state of the physical element, and also outputs a heat map Hm. In particular, in examples where the first machine learning model NNi corresponds to a first neural network, the first neural network may include an output layer with a plurality of output neurons in which one output neuron determines the score Se and a group of other output neurons from among the plurality of output neurons determines the heat map Hm.

[0106] In examples where method 100 includes operation 140, method 100 may further include an operation 141 of displaying the heat map Hm on a screen (for example, on the screen of the processing system) and an operation 142 of displaying identification information on that screen for one or more pixels of the heat map Hm having an intensity value greater than a predetermined threshold. In particular, identification information may correspond to a boundary of the pixel or plurality pixels on the screen, for example by outlining these pixels on the screen. These examples are schematically illustrated in [Fig. 12]. These examples allow the identification, on a screen displaying the heat map, of probable defect areas of the physical element being processed. This notably facilitates operator assistance in training the first NNb machine learning model by providing a visual representation of what this model considers to be potentially defective on the physical element.

[0107] In examples of process 100 comprising operation 140, process 100 may further comprise, where the physical element is classified in the second class C2, several additional operations.

[0108] In particular, in these examples, the method may further include a classification operation 150 of at least one defect of the physical element by a second machine learning model NN2. The second machine learning model NN2 may, for example, correspond to a second neural network, which may be a convolutional neural network.

[0109] The second machine learning model NN2 receives as input the heat map Hm of the physical element being processed, determined by the first machine learning model NNb. From this heat map Hm, the second machine learning model NN2 performs a classification of at least one defect and, advantageously, of all the defects in the physical element being processed. Consequently, the second machine learning model NN2 also includes the identification of defects in the physical element from the heat map Hm. Thus, when the second machine learning model NN2 corresponds to a second neural network, it can include a first plurality of specific neurons used for identifying defects on the heat map, and a second plurality of specific neurons for classifying the identified defects.

[0110] The classification 150 of at least one defect may include a classification 151 of each defect of the physical element into an acceptable defect class Ca or a critical defect class Ce. These examples are schematically represented in [Fig. 13]. Furthermore, an example of a defect classification Cd comprising the acceptable defect class Ca and the critical defect class Ce is schematically represented in [Fig. 14].

[0111] The acceptable defect class Ca may, for example, correspond to a class grouping defects belonging to a plurality of defect types, in which each defect type is associated with a corrective operation belonging to the determined list of remediation operations. In particular, a defect type can be associated with a corrective operation belonging to the list of remediation operations when this corrective operation allows the defect to be corrected. Therefore, class Ca groups together defects that can be corrected so that the physical element moves from the second class C2 (compliant subject to performing a remediation operation) to the first class Cl (compliant).

[0112] The critical defect class Ce can, for example, correspond to a class grouping other defects, that is, defects for which no corrective operation belonging to the determined list of remediation operations is associated. In other words, if the second machine learning model NN2 classifies a defect as belonging to the critical defect class Ce, it means that the classification of the physical element processed as belonging to the second class C2 was an error and that the physical element should have been classified in the third class C3 (non-conforming).

[0113] In examples where process 100 includes classification operation 150, the process may further include an operation 160 for determining a compliance score Src of the physical element by the second machine learning model NN2. Thus, in these examples, the second machine learning model NN2 allows both the classification of defects and the determination of the compliance score Src of the physical element. These examples are schematically illustrated in [Fig. 15].In examples where the second machine learning model NN2 corresponds to a second neural network, the second neural network may include an output layer with a plurality of output neurons in which one output neuron determines the compliance score Src and a group of other output neurons from among the plurality of output neurons determines the classification Cd of the physical element defects.

[0114] A first schematic example of a software architecture comprising the first machine learning model NNi and the second machine learning model NN2 is schematically represented in [Fig. 16]. In this example, at least one data point D is input into the first machine learning model NNi, which outputs the score Se associated with a state of the physical element, and also outputs a heat map Hm. The heat map Hm is input into the second machine learning model NN2, which outputs both a classification of defects Cd and a compliance score Sc of the physical element.

[0115] In examples in which process 100 includes operation 160, the second machine learning model NN2 can also receive in input is a binary compliance signal Bc. In these examples, the compliance score Src is equal to the binary compliance signal.

[0116] The binary conformity signal Bc can, for example, correspond to a signal sent by a quality control operator indicating whether or not the processed physical element can be brought into conformity. In particular, the binary conformity signal Bc can correspond to 1 (or alternatively 0) when the quality control operator determines that the processed physical element can be brought into conformity and can correspond to 0 (or alternatively 1) when the quality control operator determines that the processed physical element cannot be brought into conformity.

[0117] Thus, in these examples, the conformity score Src becomes a binary score that relies on an external operator to enable the training of the first machine learning model NNb and possibly the second machine learning model NN2. It is understood that these process examples 100 combine advantageously with those comprising operations 141 and 142, for which areas considered to be defect zones are identified on a screen displaying the heat map. The operator can therefore easily identify on the screen, or directly by observing the physical element being processed (possibly based on the observations displayed on the screen), any potentially defective zones, and indicate whether, in their opinion, the physical element can be brought into conformity via the binary conformity signal Bc.

[0118] The binary conformance signal Bc can thus be used to enable supervision of the learning of the first NNi and / or the second NN2 machine learning model.

[0119] A second schematic example of a software architecture comprising the first machine learning model NNi and the second machine learning model NN2 is schematically represented in [Fig. 17]. This is the same architecture as that shown in [Fig. 16], which also integrates the binary conformity signal Bc as input to the second machine learning model NN2.

[0120] In examples where process 100 includes operation 160, process 100 may further include an operation 170 for determining a correction factor Fc by a third machine learning model NN3. The third machine learning model NN3 may correspond to a third neural network, for example, a recurrent neural network. This correction factor Fc may be used in an operation 180 to modify the first machine learning model NNi in order to train this model.

[0121] The third machine learning model NN3 receives as input at least one data point D characterizing the physical element, the classification Cd of at least one defect in the physical element, and the compliance score Src. The third machine learning model NN3 therefore receives the input data from the first machine learning model NN3 (the at least one data point D), and the outputs (the defect classification Cd and the compliance score Src) from the second machine learning model NN3. From these elements, the third machine learning model NN3 is able to determine a correction factor Fc intended to improve the performance of the first machine learning model NNb. The third machine learning model NN3 thus acts as a reinforcement model for the learning of the first machine learning model NNb.

[0122] In examples in which process 100 includes operation 170, process 100 may further include an operation 180 of modifying the first machine learning model NNi from the correction factor Fc. These examples are schematically illustrated in [Fig. 18].

[0123] In examples where the first machine learning model NNi corresponds to a first neural network, the correction factor Fc can be applied to the first neural network so as to modify a weight associated with one or more specific neurons of the first neural network. The correction factor Fc can, for example, comprise a plurality of correction coefficients, where each correction coefficient is determined to be applied to a weight of a respective neuron of the first neural network.

[0124] An example of a software architecture comprising the first machine learning model NNb, the second machine learning model NN2, and the third machine learning model NN3 is schematically represented in [Fig. 19]. In this example, the third machine learning model NN3 is provided as input at least one data point D characterizing the physical element, the classification Cd of at least one defect in the physical element, and the compliance score Src. The third machine learning model NN3 then determines as output the correction factor Fc, which is provided as input to the first machine learning model NNb.

[0125] The model(s) can therefore be trained continuously during the implementation of process 100 on a plurality of physical elements. In particular, it is understood that the different iterations of process 100 make it possible to determine, with increasing precision, the score Se representing the state of an element to analyze, so that the classification of an element to be analyzed into one of the three classes becomes increasingly precise.

Claims

Demands

1. A computer-implemented method (100) for processing data relating to a physical element, in particular a manufactured product, the method comprising: obtaining (110) at least one data point (D) characterizing the physical element; determining (120) a score (Se) representative of a state of the physical element from processing, by a first machine learning model (NNi), the at least one data point (D) obtained; and an automated classification (130) of the physical element into a classification (Ce) representative of a state of the physical element from the determined score and a dynamically determined conformity threshold;in which the automated classification (Ce) representative of the state of the physical element comprises three classes: a first class (Cl) in which the physical element is defined as compliant, a second class (C2) in which the physical element is defined as compliant subject to performing at least one recompliance operation on the physical element belonging to a determined list of recompliance operations, and a third class (C3) in which the physical element is defined as non-compliant.;

2. A method according to the preceding claim, wherein, prior to obtaining (110) at least one data (D) characterizing the physical element, the method further comprises: receiving (101) at least one measurement signal of the physical element; preprocessing (102) a measurement signal in order to obtain a data of the at least one data (D) characterizing the physical element.

3. A method according to claim 2, wherein the at least one data (D) characterizing the physical element comprises data relating to a contour of the physical element on an image of the physical element; in which at least one measurement signal includes an image of the physical element; and in which the preprocessing (102) of a measurement signal in order to obtain data of at least one data characterizing the physical element includes a detection (1020) of the contours of the physical element in the image.

4. A method according to any one of claims 2 or 3, wherein the at least one data (D) characterizing the physical element includes data relating to a dimension of the physical element; wherein the at least one measurement signal includes an image or a point cloud of the physical element; and wherein the preprocessing (102) of a measurement signal in order to obtain data of the at least one data characterizing the physical element includes a determination (1021) of a dimension of the physical element from the image or the point cloud.

5. A method according to any one of the preceding claims, wherein the method (100) further comprises: a determination (140) of a heat map by the first machine learning model (NNi), from at least one data (D) characterizing the physical element; the heat map associating a plurality of coordinates of the physical element being processed with a respective intensity value representative of a probability of defect.

6. A method according to the preceding claim, wherein the method (100) further comprises, when the physical element is classified in the second class: a classification (150) of at least one defect of the physical element by a second machine learning model (NN2), the second machine learning model (NN2) receiving as input the heat map determined by the first machine learning model (NNi); and a determination (160) of a compliance score (Src) of the physical element by the second machine learning model (NN2).

7. A method according to the preceding claim, wherein the second machine learning model (NN2) also receives input a binary compliance signal (Bc); and in which the compliance score (Src) is equal to the binary compliance signal (Bc).

8. A method according to any one of claims 6 or 7, wherein the classification (150) of at least one defect of the physical element comprises a classification (151) of each defect of the physical element into an acceptable defect class (Ca) or into a critical defect class (Ce); the acceptable defect class (Ca) grouping defects belonging to a plurality of defect types, in which each defect type is associated with a correction operation belonging to the determined list of remediation operations; and the critical defect class (Ce) corresponding to a class grouping the other defects.

9. A method according to any one of claims 6 to 8, further comprising: a determination (170) of a correction factor (Fc) by a third machine learning model (NN3), the third machine learning model (NN3) receiving as input at least one data (D) characterizing the physical element, the classification (Cd) of at least one defect of the physical element, and the remediation score (Src); and a modification (180) of the first machine learning model (NNi) from the correction factor (Fc).

10. Product computer program comprising instructions for carrying out a method (100) according to any one of the preceding claims.

11. Data processing device (1) configured for the implementation of a method according to any one of claims 1 to 9.

12. Data processing system relating to a physical element, in particular a manufactured product, comprising: at least one sensor (2) configured to acquire at least one measurement signal of the physical element; and a data processing device (1) according to the preceding claim; in which at least one data (D) characterizing the physical element is obtained from at least one measurement signal of the physical element.

Images