Inertial Friction Welding Simulation

Abstract

The invention relates to a method and system for simulating inertial friction welding of parts to be welded, comprising: - an acquisition module (3) configured to acquire an input configuration (9) consisting of a set of physical parameters relating to the parts to be welded and the welding process, - a machine learning module (5) comprising a predictive neural network (51) configured to establish a friction curve (11) representative of the weld between said parts, taking as input said acquired input configuration, - a simulation module (7) configured to simulate the evolving behavior of the welding operation using the friction curve (11) associated with said input configuration (9). Figure for the abstract: Fig. 2.

Description

Title of the invention: SIMULATION OF INERTIAL FRICTION WELDING technical field

[0001] The present invention relates generally to a method and system for simulating inertial friction welding and in particular to its modeling using a friction law determined by machine learning. PREVIOUS STATE OF THE ART

[0002] Weld reliability is a major concern in several industrial sectors. In particular, inertial friction welding offers very good mechanical properties, preventing the formation of cracks, porosity, or any reduction in strength in the weld zone.

[0003] The evolving behavior of the friction welding operation depends on several physical parameters and therefore it is advantageous to model the welding process in advance in order to choose the physical parameters generating the most optimal and reliable welding process.

[0004] Currently, friction welding process modeling is essentially carried out in descriptive mode. This means that it is necessary to obtain input data for the weld to be modeled before performing the numerical simulation. In other words, it is not currently possible to simulate welds that have not yet been made.

[0005] Initial tests predicting the results of inertial friction welding have been carried out. These tests make it possible to predict an inertial friction weld with a welding configuration consisting of a set of well-defined input parameters. However, the results become incorrect as soon as one of the input parameters is modified. Thus, it is necessary to identify a friction law for each welding configuration encountered. This identification is based on a trial-and-error method, increasing the pressure in increments until success or until the tests are stopped. This identification procedure is not very precise while also being very long and repetitive.

[0006] The object of the present invention is therefore to remedy the aforementioned drawbacks by proposing a method and a system for predictive inertial friction welding simulation that allows the results of welding to be anticipated on each configuration while being precise, efficient and fast. Description of the invention

[0007] The present invention is defined by a system for simulating inertial friction welding of parts to be welded, comprising:

[0008] - an acquisition module configured to acquire an input configuration formed a set of physical parameters relating to the parts to be welded and the welding process,

[0009] - a machine learning module comprising a predictive neural network configured to establish a friction curve representative of the weld between said parts, taking as input said acquired input configuration,

[0010] - a simulation module configured to simulate the evolving behavior of the welding operation using the friction curve associated with said input configuration.

[0011] This makes it possible to predict the friction law that is most representative of the evolution of the welding parameters and, consequently, to predictively simulate the welding procedure in order to anticipate the results on each configuration while being precise, efficient and fast.

[0012] Advantageously, the simulation module is configured to determine the evolving behavior of the welding operation by simulating the evolution of temperature, and / or deformation and / or stress in the parts to be welded.

[0013] Advantageously, the physical parameters relating to the parts include parameters relating to the materials of the parts to be welded, the thicknesses of the parts to be welded, and the average diameters of the parts to be welded. The physical parameters relating to the welding process are point values ​​of surface welding energy, torque of the parts to be welded, rotational speed, and pressure.

[0014] Each input configuration is selected from a set of input parameters, thus enabling the prediction of the most optimal weld. Apart from the material and geometry of the parts, all the physical quantities used as input to the learning module are derived from data acquired routinely during welding.

[0015] Advantageously, the predictive neural network is configured to take as input a point abscissa z of a considered point of the friction curve to be predicted and to give as output the friction coefficient h corresponding to said abscissa z used as input, the point abscissa z being a dimensionless number determined by the values ​​of said physical input parameters.

[0016] This allows us to obtain, by iterative prediction, the friction curve p=f(z) that is most representative of the weld.

[0017] Advantageously, the predictive neural network comprises a succession of totally interconnected layers, the neurons having the same logistic activation function of the sigmoid function type.

[0018] Advantageously, the system includes a training database configured to train the neural network, the training database comprising a set of uniformly weighted learning friction curves, the learning friction curves being pre-processed experimental friction curves.

[0019] Advantageously, the system includes a preprocessing module comprising preprocessing neural networks having experimental friction curves as input and training friction curves as output, the preprocessing neural networks being configured to perform the following steps:

[0020] - smoothing and resampling of each experimental curve to form smoothed and resampled experimental curves, and

[0021] - extrapolation of said smoothed and resampled experimental curves for form the learning friction curves with identical x-axis ranges.

[0022] Advantageously, the simulation module is configured to use each friction curve as input to simulate the evolution during welding of the temperature of the parts, and / or the deformation of the parts and / or the stress on the parts, for each input configuration associated with said each friction curve.

[0023] Thus, the simulation becomes predictive and eliminates the need for the physical execution of the weld. Indeed, the simulation uses the curve predicted by the neural network and no longer requires welding a configuration to obtain its friction curve by inverse identification.

[0024] Advantageously, the simulation module is configured to perform the simulation using a finite element calculation technique.

[0025] The invention also relates to a method for simulating inertial friction welding of parts to be welded, comprising the following steps implemented by computer:

[0026] - acquire an input configuration consisting of a set of physical parameters relating to the parts to be welded and the welding process,

[0027] - to establish, using a predictive neural network, a friction curve representative of the welding between said parts for said input configuration,

[0028] - simulate the evolving behavior of the welding operation using the curve of friction associated with said input configuration.

[0029] Advantageously, the simulation of the evolving behavior of the welding operation includes the simulation of the evolution of the temperature, and / or the deformation and / or the stress in the parts to be welded.

[0030] Advantageously, the predictive neural network is pre-trained using a mini-batch gradient descent optimization technique, using a training database containing a set of training friction curves corresponding to pre-processed experimental friction curves.

[0031] The invention also relates to a computer program product comprising code instructions for implementing the steps of an inertial friction welding simulation process according to any of the above characteristics, when said code instructions are executed by a processing unit. Brief description of the drawings

[0032] Other features and advantages of the invention will become apparent upon reading preferred embodiments of the invention made with reference to the accompanying figures, among which:

[0033] Fig. 1 schematically illustrates an inertial friction welding process between two parts to be welded;

[0034] Fig. 2 schematically illustrates a system for simulating inertial friction welding of parts to be welded, according to a preferred embodiment of the invention;

[0035] The [Fig.3] is a graph schematically illustrating a friction curve determined by the machine learning module, according to the invention;

[0036] Figure 4 schematically illustrates a simulation system comprising a training device, according to a preferred embodiment of the invention;

[0037] Fig. 5 schematically illustrates the steps of a simulation process that can be implemented by a system according to the embodiment of Fig. 4;

[0038] The [Fig.6A],

[0039] [Fig.6B] and [Fig.6C] are graphs illustrating the steps performed on an experimental friction curve to obtain the corresponding learning friction curve, according to the invention; and

[0040] Fig. 7 is a graph illustrating several friction curves predicted by the machine learning module, according to the invention. DETAILED DESCRIPTION OF THE INVENTION

[0041] The principle of the invention consists of using a neural scheme to predict a friction curve representative of the weld to be simulated, without needing to actually perform a friction weld.

[0042] Fig. 1 schematically illustrates an inertial friction welding process between two parts pl, p2 to be welded.

[0043] Inertial friction welding is a solid-state welding process consisting of rotating one of the parts before bringing them into contact under the effect of pressure. This process is described by time evolution curves of the speed V, the pressure P and the material consumption M.

[0044] More specifically, this welding process comprises three phases. The first phase (Phase 1) is a start-up phase during which one of the parts pl is rotated R and then disengaged. After disengagement, the part pl continues to rotate by inertia with a decreasing rotational speed R.

[0045] The second phase (Phase 2) is a welding phase comprising two periods. During the first period, the two parts pl, p2 are brought into contact under the effect of a pressure force E that increases rapidly before stabilizing at a constant pressure. The rotational speed R continues to decrease freely. During this braking period, the parts pl, p2 heat up locally due to friction. During the second period, the parts pl, p2 deform under the effect of heat and pressure force. The assembly of two parts shortens due to increased material consumption. The rotational speed R continues to decrease.

[0046] The third phase (Phase 3) is a forging phase. This phase begins after the complete deceleration of the workpiece pl, i.e., after its rotational speed has reached zero. During this third phase, there is no further material consumption, and the workpieces pl, p2 cool down while the pressure force E is kept constant or possibly increased.

[0047] Figure 2 schematically illustrates a system for simulating inertial friction welding of parts to be welded, according to a preferred embodiment of the invention. Figure 2 is also used to illustrate a method for simulating inertial friction welding, according to an embodiment of the invention.

[0048] This simulation system 1 comprises an acquisition module 3, a machine learning module 5, and a simulation module 7.

[0049] The acquisition module 3 is configured to acquire an input configuration 9 from a set of input configurations. Each input configuration 9 comprises a first set 91 of physical parameters relating to the parts to be welded and a second set 93 of physical parameters relating to the welding process.

[0050] The first set 91, corresponding to the physical input parameters relating to the parts, preferably includes parameters relating to the materials, thicknesses, and average diameters of the parts to be welded. The second set 93, corresponding to the physical input parameters relating to the welding process, preferably includes point values ​​from the welding process comprising quantities of surface energy of welding, inertia or torque of the parts to be welded, rotational speed, and pressure force.

[0051] The machine learning module 5 is configured to establish a friction law (or curve) 11 representative of the weld between the parts to be welded pl, p2 for each input configuration.

[0052] It should be noted that the friction law is a so-called Stribeck curve, which defines the evolution of the friction coefficient 1* as a function of a dimensionless variable denoted z characterizing the parameters of the process under study. This law or curve is known to those in the prior art.

[0053] More specifically, the machine learning module 5 comprises a predictive neural network 51 which takes as input a point abscissa z of a considered point on the friction curve 11 to be predicted, and as output the friction coefficient ■u corresponding to said abscissa z used as input. This makes it possible to obtain, by iterative prediction, i.e. point by point, the friction curve p=f(z) characterizing the weld according to the input configuration.

[0054] Indeed, [Fig.3] is a graph schematically illustrating a friction curve determined by the machine learning module, according to the invention.

[0055] The point abscissa z is a dimensionless number determined by the point values ​​of the input physical parameters. Thus, the friction law p=f(z) incorporates the influence of all the input physical parameters relating to the parts to be welded pl, p2 and the welding process. Apart from the material and geometry of the parts, all the physical quantities used as input to the learning module 5 are derived from machine data acquired during the welding process.

[0056] By way of example, the dimensionless variable z is equal to the viscosity of the material multiplied by the linear velocity divided by the pressure.

[0057] Thus, the neural network 51 makes it possible to obtain a friction curve 11 for each weld, incorporating the influence of various parameters such as the material, the geometry of the parts, and the specific parameters of the welding process. The friction curve 11 is then used as input to a numerical simulation to model the welding process.

[0058] The predictive neural network 51 comprises a succession of fully interconnected layers. Each neuron in the preceding layer is connected to all the neurons in the following layer. By way of example, the total number of layers is between 5 and 25 and the number of neurons per layer is between 20 and 50.

[0059] Advantageously, the neurons all have the same non-linear activation function such as the sigmoid logistic function.

[0060] The training of the predictive neural network 51 is carried out using a database of training data constructed beforehand and comprising a set of learning friction curves (see Figs. 4 and 5).

[0061] The simulation module 7 is configured to numerically simulate the evolving behavior of the welding operation using the friction curve 11 associated with each input configuration 9.

[0062] The simulation module 7 comprises a computing unit 71, a memory 73 and an output unit 75 including, for example, a screen.

[0063] More particularly, the simulation module 7 preferably has as input data the rotational speed of the rotating part, the pressure on the rotating part, the inertia of the plates of the inertial friction machine, and the friction curve 11 obtained via the machine learning module 5.

[0064] The computing unit 71, in conjunction with the memory 73, uses the friction curve 11 as input to a finite element simulation to determine the evolution of the temperature, deformation, and stress of the simulated weld for the input configuration 9 associated with the friction curve. This allows access to areas of the workpiece that are experimentally inaccessible. For example, the simulation makes it possible to model the temperature rise at the interface of the parts to be welded pl, p2, as well as the deformation of the parts, and in particular, the formation of the weld bead created by heating and pressure. This simulation can be displayed on the screen or printed by the output unit 75.

[0065] Numerical finite element simulation no longer requires the physical fabrication of a weld according to a given configuration. Indeed, it is no longer necessary to obtain a friction curve corresponding to a physical weld according to the given configuration in order to perform a descriptive simulation by inverse identification. The simulation according to the invention eliminates the need for the physical fabrication of the weld. In fact, it simply uses the curve 11 predicted by the neural network 51 and thus becomes predictive.

[0066] The method according to the invention is defined as follows in two complementary steps. The first step involves the use of a neural network to predict the friction curve p=f(z) representative of the weld to be simulated. The objective of the first step is therefore to obtain a friction curve or law incorporating the influence of various parameters associated with each input configuration of the weld.

[0067] The second step involves using the friction law predicted by the neural scheme as input to a numerical simulation, for example by finite elements, to predict the behavior of the welding.

[0068] Figure 4 schematically illustrates a simulation system comprising a training device, according to a preferred embodiment of the invention.

[0069] The simulation system 1 comprises an acquisition module 3, a machine learning module 5, a simulation module 7, and a training device 13.

[0070] The training device 13 includes a preprocessing module 15 configured to build a training database 17 intended to be used for training the automatic learning module 5.

[0071] The preprocessing module 13 comprises preprocessing neural networks 151 consisting of a succession of layers, the number of neurons in each layer being, for example, between 20 and 50. The total number of layers may be between 5 and 25. The layers are preferably fully interconnected. The neurons preferably have the same activation function, which may be a sigmoid function.

[0072] It should be noted that, alternatively, the training device 13 may not be integrated into the simulation system.

[0073] Fig. 5 schematically illustrates a simulation process that can be implemented by a system according to the embodiment of Fig. 4.

[0074] In step El, crude experimental friction curves 110 are constructed from the physical welding processes.

[0075] In step E2, the preprocessing module 13 is configured to determine learning friction curves from experimental friction curves.

[0076] More specifically, the preprocessing neural networks 151 perform preprocessing operations on the experimental friction curves 110 to provide the training friction curves 111 as output. These preprocessing operations include smoothing, resampling and extrapolation of the experimental curves 110.

[0077] Figs. 6A-6C are graphs illustrating the steps carried out on an experimental friction curve to obtain the corresponding learning friction curve, according to the invention.

[0078] Each graph represents the friction coefficient h as a function of the dimensionless variable z and more precisely of the logarithm of z.

[0079] In particular, [Fig. 6A] illustrates the smoothing and resampling effects performed by the preprocessing neural networks 151 on a raw experimental curve. The star-shaped points 110a represent the raw curve, while the simple points 110b represent the smoothed and resampled curve according to a rough 50-point estimate.

[0080] Fig. 6B illustrates the extrapolation of the experimental curve by extending the abscissa range over the low 120a and high 120b values ​​of z. This allows the extension of the abscissa ranges of the smoothed and resampled experimental curves, thus forming learning friction curves with extended abscissa ranges.

[0081] Indeed, the extension to low 120a and high 120b abscissas z is carried out so that the discretization of the points of each of the curves is on an identical z range for each of the curves in the training database 17.

[0082] By way of example, for the low 120a and high 120b abscissa z extension, the preprocessing neural networks 151 use at most 40% of the points located on either side of the ends of the smoothed curve sampled at 50 points in order to deduce the extension locally.

[0083] Figure 6C illustrates the points of the experimental curve and the corresponding learning friction curve 111. The points of the experimental curve are represented by stars 110a. According to this example, the learning friction curve 111 is the result of smoothing and resampling operations at 200 points after an extrapolation operation.

[0084] In total, according to this example, each learning friction curve 111 represents 200 data points. All these learning friction curves 111 defined on the same abscissa range z form a relevant training database 17.

[0085] Advantageously, the learning friction curves 111 are uniformly weighted. Thus, the weights of each point on a curve depend on the total number of points for the curve. For example, for a curve described by 50 points, each of the 50 points will have a weight of 1 / 50 during training.

[0086] The training database thus comprises all the points of each of these uniformly weighted friction curves. It should be noted that the training database 17 can contain several hundred training friction curves with different input configurations.

[0087] Advantageously, an inverse identification of the friction curve is performed by the simulation module 7 for each of the welds in the training database 17.

[0088] In step E3, the training database 17 is used to perform the training of the machine learning module 5.

[0089] Advantageously, a mini-batch gradient descent optimization algorithm is used to train the predictive neural network 51 of the machine learning module 5. The neural network weights are updated incrementally. During an entire training cycle, the entire set of mini-batches is used. For example, the size of the mini-batches ranges from 32 to 128 points. During training, a common technique called "momentum" can be used at each update of the neural network weights. The factor used when updating the weights is, for example, between 10% and 90%. An example of the "momentum" technique is described by P. Kim in MATLAB Deep Learning with Machine Learning, Neural Networks and Artificial Intelligence, 2017.

[0090] Advantageously, once the training is completed, blind tests are carried out to validate the accuracy of the friction curve predicted during simulation calculations. The success criterion is the correct prediction of material consumption at the end of the weld and of the interface temperature during welding.

[0091] After training, the machine learning module 5 is ready to predict friction curves from different configurations.

[0092] Indeed, at step E4, for each input configuration 9 acquired by the acquisition module 3, the machine learning module 5 uses this input configuration to establish a corresponding friction curve 11 representative of the weld.

[0093] More specifically, the machine learning module 5 uses the input configuration comprising the physical parameters relating to the parts to be welded (materials, thicknesses, and average diameters) as well as the point values ​​from the welding process (welding surface energy, inertia, rotational speed, and pressure) to determine the point abscissas z and to predict the friction curve p=f(z). This friction curve, representative of the weld associated with the input configuration, is iteratively predicted by the already trained predictive neural network.

[0094] Fig. 7 is a graph illustrating several friction curves predicted by the machine learning module, according to an example of the invention.

[0095] This example shows three curves lia, 11b, 1 le for three different inlet configurations 9, namely three different pressures, 48 ​​MPa, 68 MPa and 88 MPa respectively, all other physical parameters (geometry, materials, surface energy and velocity) being almost constant between these three curves.

[0096] For each configuration 9 of a geometry, a material pair and / or a welding parameter, the machine learning module 5 provides a friction curve 11.

[0097] Next, in step E5, the friction curve 11 representative of the weld of interest is used to feed the simulation module 7.

[0098] With input data the rotational speed of the rotating part, the pressure on the rotating part, the inertia of the plates of the inertial friction machine, and the friction curve 11, the simulation module 7 determines the evolving behavior of the welding operation.

[0099] More particularly, the simulation module 7 preferably has as input data the rotational speed of the rotating part, the pressure on the rotating part, the inertia of the plates of the inertial friction machine, and the friction curve 11 obtained via the machine learning module 5.

[0100] The calculation unit 71 of the simulation module 7 uses finite element calculation to determine the evolving behavior of the welding operation by preferably simulating the evolution of temperature, deformation and stress in the modeled parts.

Claims

Demands

1. Inertial friction welding simulation system for parts to be welded, characterized in that it comprises: - an acquisition module (3) configured to acquire an input configuration (9) formed of a set of physical parameters relating to the parts to be welded and the welding process, - a machine learning module (5) comprising a predictive neural network (51) configured to establish a friction curve (11) representative of the weld between said parts by taking as input said acquired input configuration, - a simulation module (7) configured to simulate the evolving behavior of the welding operation using said friction curve (11) associated with said input configuration (9), the simulation of the evolving behavior of the welding operation being used to anticipate the welding results on each input configuration.

2. System according to claim 1, characterized in that the simulation module (7) is configured to determine the evolving behavior of the welding operation by simulating the evolution of temperature, deformation and stress in the parts to be welded.

3. System according to any one of the preceding claims, characterized in that the predictive neural network (51) is configured to take as input a point abscissa z of a considered point of the friction curve (11) to be predicted, and to give as output the friction coefficient >u corresponding to said abscissa z used as input, the point abscissa z being a dimensionless number determined by the values ​​of said physical input parameters.

4. System according to any one of the preceding claims, characterized in that the predictive neural network (51) comprises a succession of totally interconnected layers, the neurons of said predictive neural network (51) having the same logistic activation function of the sigmoid function type.

5. A system according to any one of the preceding claims, characterized in that it comprises a training database (17) configured to train the predictive neural network (51), the training database (17) comprising a set of uniformly weighted learning friction curves (111), the learning friction curves being pre-processed experimental friction curves.

6. System according to claim 5, characterized in that it comprises a preprocessing module (15) including preprocessing neural networks (151) having as input experimental friction curves (110) and as output learning friction curves (111), the preprocessing neural networks being configured to perform the following steps: - smoothing and resampling of each experimental curve to form smoothed and resampled experimental curves, and - extrapolation of said smoothed and resampled experimental curves to form the learning friction curves (111) with identical abscissa ranges.

7. System according to any one of the preceding claims, characterized in that the simulation module (7) is configured to use each friction curve (11) as input to simulate the evolution during welding of the temperature of the parts, and / or the evolution of the deformation of the parts, and / or the evolution of the stress on the parts, for each input configuration (9) associated with said each friction curve (11).

8. A method for simulating inertial friction welding of parts to be welded, characterized in that it comprises the following computer-implemented steps: - acquiring an input configuration (9) consisting of a set of physical parameters relating to the parts to be welded and the welding process, - establishing by a predictive neural network (51) a friction curve (11) representative of the weld between said parts for said input configuration, - simulating the evolving behavior of the welding operation using the friction curve associated with said input configuration, the method preferably being implemented by a simulation system according to any one of claims 1 to 7.

9. A method according to claim 8, characterized in that the predictive neural network (51) is pre-trained using a gradient descent optimization technique by mini-batches, using a training database (17) comprising a set of training friction curves (111) corresponding to pre-processed experimental friction curves.

10. Product computer program comprising code instructions for implementing the steps of an inertial friction welding simulation process according to any one of claims 8 or 9, when said code instructions are executed by a processing unit.

Images