Method for measuring material characteristics of a broken part

Abstract

The invention relates to a method for measuring characteristics associated with a material forming a part that has undergone fracture, comprising the steps of: - performing a topographic survey of the fracture surface of this material, - using at least one machine learning algorithm to determine, from this topographic survey, said characteristics. (Fig. 1)

Description

Title of the invention: Method for measuring characteristics associated with the material of a part that has undergone a rupture

[0001] The present invention relates to the field of statistical fractography, that is to say to the quantitative analysis of the topography of the fracture surfaces of materials.

[0002] Many previous studies have focused on such analyses: it then appeared that the topographical survey of the fracture surface of a mechanical component contains a lot of information which, correctly processed and interpreted, allows in particular a better understanding of the causes and methods of the fracture of the material forming the part.

[0003] This better understanding, particularly of rupture phenomena, is essential in many industrial fields, such as aeronautics, railways, automobiles or the energy sector where we seek to anticipate and, ultimately, avoid mechanical ruptures in service.

[0004] The present invention aims in particular to provide a new approach making it possible, from the analysis of the fracture surface of a material, to access characteristic information associated with this material.

[0005] This object of the invention is achieved with a method for measuring characteristics associated with a material forming a part which has undergone rupture, comprising the steps of:

[0006] - carry out a topographic survey of the fracture surface of this material,

[0007] - using at least one machine learning algorithm to determine, at from this topographic survey, the said characteristics.

[0008] Surprisingly indeed, it was realized, through the analysis of numerous fracture surfaces made possible by artificial intelligence tools - and in particular machine learning - that there was a predictable relationship between the topography of these surfaces and intrinsic characteristics or characteristics linked to the history of the material forming the part that broke.

[0009] According to other optional characteristics of the method according to the invention, taken alone or in combination: - said algorithm is of the deep learning type: this type of algorithm - also referred to as a convolutional neural network algorithm - has proven to be particularly effective in the context of the present invention, making it possible to achieve high reliability; • said characteristics are chosen from the group comprising: • the magnitude of the stresses leading to the rupture of the material, in particular through fatigue, • the intensity of the stresses leading to the rupture of the material, • the mechanical properties of the material such as its toughness, its elastic limit or its work hardening coefficient, • the nature of the corrosion leading to the rupture of the material, • the nature of the material, • the treatment undergone by the material, • the origin of the material.

[0010] The method according to the present invention, which can be applied in particular to metals and metal alloys, can thus find a very wide range of applications, making it possible to obtain numerous very useful pieces of information on the intrinsic and historical characteristics of the materials as well as on the stresses undergone in service.

[0011] Other characteristics and advantages of the invention will emerge on reading the description which follows, with reference to the appended figures, which illustrate:

[0012] [Fig-1]: a confusion matrix, allowing to visualize the gap between, on the one hand the values ​​determined by means of the method according to the invention and, on the other hand, of the amplitude of the stresses having led to the fatigue failure of a part formed in a given material, for different stress amplitudes, and

[0013] [Fig.2]: a two-dimensional graph allowing only points to be visualized determined by means of the method according to the invention are grouped into packets representative of the nature of the material forming the part which has undergone the rupture.

[0014] In the following, the method according to the invention will be illustrated more particularly when it is used to determine the amplitude of the stresses on a part formed from a material which has undergone fatigue failure.

[0015] As is known per se, fatigue failure is the result of repeated stresses applied to the material, having a variable or substantially constant amplitude, leading to cracking and, ultimately, to rupture of the material.

[0016] Other failure scenarios also fall within the scope of the present invention, such as sudden failure, caused by a single stress whose intensity exceeds the failure threshold of the material, or such as failure by stress corrosion.

[0017] Therefore, with regard to the fatigue failure of a part formed from a given material, the objective is to determine the amplitude of the fatigue cycles commonly noted AK and expressed in MPa.m1 / 2 which led to the failure of the part.

[0018] Indeed, for metal alloys in particular, the higher the amplitude AK, the more quickly the crack propagates, and therefore the more quickly the phenomenon of rupture of the part occurs.

[0019] Thus, when analyzing post-mortem a part formed from a given material, i.e. after rupture, it is essential to know the amplitude AK of the fatigue cycles because this information provides the mechanical loading at the origin of the rupture, ultimately allowing the replacement part to be correctly sized and / or an appropriate material to be chosen so that the part remains operational throughout the desired service life.

[0020] According to the method according to the invention, one begins by training software using at least one machine learning algorithm using a part fracture face database.

[0021] In the following, the term “machine learning software” will refer to the software using the machine learning algorithm.

[0022] This facies database consists of a large number of topographic surveys of fracture surfaces of parts formed in one or more different materials, having undergone fatigue cycles for different ranges of AK amplitude values.

[0023] These topographic surveys, obtained for example using a profilometer, typically make it possible to obtain images of a few tens of thousands of pm2 containing around 106 pixels each.

[0024] These images are divided into image classes, all the images of the same class corresponding to fracture surfaces of parts formed from the same material and having undergone fatigue cycles of the same amplitude AK.

[0025] Within a class of images, a portion of the images (typically 80% of the images) is selected to train the machine learning software, and the other portion of the images (typically 20% of the images) to validate the results provided by this software.

[0026] Concretely, during the training of the automatic learning software, for each class of images, each fracture surface is annotated, associating it with information which may typically include: the nature of the material, the amplitude AK of the stresses which led to the fracture of the part by fatigue, the treatment undergone by the material, the origin of the material, etc.

[0027] For the validation of the automatic learning software, for each class of images, images of rupture facies from the validation image database are submitted to the software, and the amplitude predicted by this software AK^^eè is compared with the actual amplitude AK corresponding to each rupture facies of the validation image.

[0028] We can thus obtain a confusion matrix such as that represented in [Fig. 1], whose abscissa axis indicates the predicted amplitude values ​​AKn-edi^ and the ordinate axis the real amplitude values ​​AK.

[0029] As can be seen from this confusion matrix, for different validation images corresponding to the same AK, i.e. to parts having undergone fatigue stresses of the same amplitude, the automatic learning software provides predicted amplitude values ​​AKpredicted slightly dispersed around the actual value AK: this is reflected by the fact that all of the squares represented in this figure are distributed substantially around a straight line with slope -1 corresponding to a perfect agreement between the predicted values ​​and the actual values.

[0030] This surprising result expresses that for different parts formed in a given material having undergone fatigue stress cycles of the same given amplitude AK, we find fracture surfaces presenting similar reliefs (fields of heights and slopes, in particular), and that these reliefs differ from those present on the fracture surfaces of parts having undergone fatigue cycles of different amplitudes.

[0031] Once this unexpected property has been identified, it is understood that artificial intelligence tools, and in particular machine learning, are particularly suitable for measuring the AK amplitude of fatigue cycles post-mortem.

[0032] These tools allow, in fact, as new fracture facies are submitted to them, to enrich their base of annotated facies, and thus to calculate predicted amplitudes AK^a^ for new fracture facies which will be all the closer to the amplitudes AK which led to the fracture of the parts concerned.

[0033] Preferably, a deep learning algorithm of the convolutional neural network (CNN) type is used as the fracture facies classification algorithm.

[0034] The back-propagation of the results provided by this algorithm for each analysis of a new rupture face makes it possible to optimize the weights associated with the neurons, for example by using a Stochastic Gradient Descent (SGD) type optimizer.

[0035] The classification of representative images of analyzed fracture surfaces can be carried out by means of cost functions, known per se.

[0036] Preferably, the training phase of the deep learning algorithm is based on the field of log-volatilities of the field of heights of the fracture surfaces, rather than on their simple topographical survey.

[0037] At each point with coordinates*' y of the rupture surface, this log-volatility is defined as _ ijOg| + £) hÇx) |, ft(x) being the height of the rupture surface measured at the point with coordinates x - (x. y), and £ being the scale at which the variation of the field is measured

[0038] In practice, by carrying out tests on different metal alloys such as 15-5PH steel, inconel, aluminium, TA6V titanium, it was found that the method according to the invention made it possible to know the amplitude AK with a margin of error of less than 10%.

[0039] According to another possibility offered by the method according to the invention, it was realized that by analyzing abrupt rupture facies and eliminating the step of classifying these facies, it was possible to transform the images into high-dimensional data (typically vectors with approximately 1000 coordinates).

[0040] By applying a dimensionality reduction algorithm such as the t-SNE algorithm to this data, this dimensionality can be reduced to a few coordinates (typically 2 coordinates), while preserving the structure of the initial data: in this way, complex data can be transformed into a more accessible format, suitable for visual interpretation.

[0041] In doing so, it is possible to establish, for fracture surfaces of parts formed from different materials, graphs such as that of [Fig.2], on which points are plotted having as coordinates those obtained after the aforementioned dimensional reduction.

[0042] Without these coordinates being in themselves representative of physical characteristics of the materials studied, it was possible to realize that the points corresponding to fracture surfaces of parts formed in the same material, are grouped in packets on the graph.

[0043] For example, in [Fig.2]: • point group 3 is representative of fracture surfaces of parts formed in a first aluminum alloy, • point group 5 is representative of fracture surfaces of parts formed in a second aluminum alloy, • point group 7 is representative of fracture surfaces of stainless steel parts, • groups of points 9, 11, 13 are representative of fracture surfaces of parts formed in three steels of different grades, • point group 15 is representative of fracture surfaces of TA6V titanium parts, and • point group 17 is representative of fracture surfaces of 15-5PH steel parts.

[0044] As can therefore be understood in light of the above, the method according to the invention therefore also makes it possible, by analyzing the fracture surface of a part formed from a given material, to determine the nature of this material.

[0045] Naturally, the invention is described in the above by way of example. It is understood that those skilled in the art are able to carry out different variant embodiments of the invention without departing from the scope of the invention.

Claims

Claims

1. Method for measuring characteristics (AK) associated with a material forming a part which has undergone a rupture, comprising the steps of: - carrying out a topographical survey of the rupture surface of this material, - using at least one automatic learning algorithm to determine, from this topographical survey, said characteristics (AK).

2. The method of claim 1, wherein said algorithm is of the deep learning type.

3. Method according to claim 1, in which said characteristics (AK) are chosen from the group comprising: - the amplitude of the stresses (AK) having led to the rupture of the material, in particular by fatigue, - the intensity of the stresses having led to the rupture of the material, - the mechanical properties of the material such as its toughness, its elastic limit or its work hardening coefficient, - the nature of the corrosion having led to the rupture of the material, - the nature of the material, - the treatment undergone by the material, - the origin of the material.

Images