阅读说明：本技术 用于增材制造部件的疲劳寿命预测的机器学习方法 (Machine learning method for fatigue life prediction of additively manufactured components ) 是由 尼古拉斯·拉门斯 于 2019-07-10 设计创作，主要内容包括：本发明涉及用于增材制造部件的考虑局部材料特性的疲劳寿命预测的方法和系统。该方法和系统被用于预测增材制造元件的疲劳寿命特性,具有数据收集步骤(1a,1b),其中收集针对元件的不同给定处理步骤的最大应力相对于疲劳破坏循环的多个数据点,具有训练步骤(2),其中利用收集的数据来训练机器学习系统,以及具有评估步骤(5,6),其中所训练的机器学习系统面对实际处理步骤并被用于预测元件的疲劳寿命特性。(The invention relates to a method and a system for fatigue life prediction of an additively manufactured component taking into account local material properties. The method and system are used for predicting fatigue life characteristics of an additively manufactured component, having a data collection step (1a, 1b), wherein a plurality of data points of maximum stress versus fatigue failure cycle for different given processing steps of the component are collected, having a training step (2), wherein a machine learning system is trained using the collected data, and having an evaluation step (5, 6), wherein the trained machine learning system faces the actual processing step and is used for predicting the fatigue life characteristics of the component.)

1. A method for predicting fatigue life characteristics of an additively manufactured element, specimen or material structure in industrial product design or manufacture, the method having a data collection step (1a, 1b) in which a plurality of data points of maximum stress versus fatigue failure cycle are collected for different given process steps, surface and volume conditions of the element, specimen or material structure,

the method has a training step (2) in which the machine learning system is trained using the collected data, and

the method has an evaluation step (5, 6) in which the trained machine learning system is faced with new actual processing conditions and is used to predict the fatigue life characteristics of the element, specimen or material structure, and

the method has an application step in which the predicted fatigue life characteristics are used to design or optimize a product or to parametrically represent a manufacturing process.

2. The method of claim 1, wherein the data collection step comprises collecting data provided by experimental tests (1 a).

3. The method according to any one of the preceding claims, wherein the data collection step comprises collecting data provided by numerical simulations (1 b).

4. The method of any preceding claim, wherein the machine learning system employs a gaussian process regression algorithm.

5. Method according to any of the preceding claims, wherein for the evaluation step (5, 6) a definition of topological properties of the element and a definition of different zones are provided, each of the zones having its own properties and conditions, and the evaluation calculates individual fatigue life properties of the different zones.

6. The method according to any one of the preceding claims, wherein the evaluating step (5, 6) comprises using a custom durability solver.

7. The method according to any of the preceding claims, wherein the result of the evaluating step (5, 6) is used to optimize an additive manufacturing process and/or a post-processing treatment step for the element, specimen or material structure.

8. A system for predicting fatigue life characteristics of an additively manufactured component, specimen or material structure for use in an industrial product design or manufacturing process,

the system includes a computer hardware component having a computer hardware component,

the computer hardware is programmed with a machine learning system,

the system is arranged for performing the training steps (1a, 1b) and the evaluation steps (5, 6) according to claim 1, the system being arranged for providing the results of the evaluation steps to an industrial product design or manufacturing process.

9. A system for predicting fatigue life characteristics of an additively manufactured element, specimen or material structure according to claim 8, the system being arranged to carry out the optimizing step according to claim 7.

10. The system of any one of claims 8 to 9, wherein the data collection step comprises collecting data provided by experimental testing.

11. The system of any one of claims 8 to 10, wherein the data collection step comprises collecting data provided by numerical simulations.

12. The system of any one of claims 8 to 11, wherein the machine learning system employs a gaussian process regression algorithm.

13. A system according to any one of claims 8 to 12, wherein for the step of evaluating, a definition of topological properties of the elements and a definition of different zones are provided, each zone having its own properties and conditions, and the evaluating calculates individual fatigue life properties of the different zones.

14. The system of any one of claims 8 to 13, wherein the evaluating step comprises using a custom durability solver.

Technical Field

The present invention relates to a method and system for fatigue life prediction for additively manufactured components taking into account local material properties.

Background

The fatigue properties of an additively manufactured part (how long the part is subjected to a certain repetitive load) strongly depend on the exact way the part is printed and post-processed. Fig. 1 shows the fatigue behavior spread of additively manufactured titanium alloy samples obtained from an extensive literature review. Fig. 1 shows the maximum stress versus fatigue failure cycle.

The spreading is obtained according to the way of the printing and post-processing components. Printing and post-processing affect both pure material properties and the occurrence of artifacts such as roughness or the presence of porosity. In order to accurately predict the fatigue behavior of a component, it is therefore necessary to treat the fatigue life according to a number of parameters, including roughness, porosity.

One problem with fatigue prediction for additive manufacturing is that the fatigue life of any experimental component is always caused by the combined influence of multiple parameters that influence the fatigue life. A second problem is that due to the large number of parameters and interactions that occur, it is difficult to develop, define and calibrate mathematical models that describe how these different parameters interact. Furthermore, the fatigue-affecting parameters are not constant over complex parts, requiring localized material properties to account for fatigue life.

The problem to be solved by the invention is therefore twofold, firstly how to describe the relationship between different parameters and the resulting fatigue life using machine learning methods, and secondly how to predict the fatigue life of complex parts where parameters such as roughness vary over the whole part (and therefore the fatigue behavior is no longer constant over the whole part).

To describe the relationship between the different parameters and the resulting fatigue life, several methods can be employed.

The printing parameters (laser power, speed … …) have been introduced into the damage method. Limitations are (i) a priori assumptions about how these parameters need to be combined in a single parameter, (ii) they can account for only a partial level of fatigue but not for variations across the part, and (iii) it is not possible to account for artifacts such as surface roughness or porosity. (see: https:// doi.org/10.1016/j.ijmecsi.2019.02.032).

For conventional ductile iron casting, artificial neural networks have been studied to classify casting quality (high, normal, and low) and predict scale factors based on the nearest neighbor method. However, this method only considers the bulk properties (i.e., tensile strength) and only attempts to improve on existing empirical rules relating static strength to fatigue properties. This method does not take into account local effects affecting fatigue life and relies entirely on existing empirical rules to relate different factors to fatigue life. In the absence of existing empirical rules (which do not currently exist in additive manufacturing), this approach does not provide the necessary insight to produce a fatigue prediction approach. (reference: in NAFIEMS conference, Dreuston, 3.2019, Lebensdauerabracschatzung von Bauteilen ausauf der Grundle der lokalen materialigenschaft (evaluation of the service life of spheroidal graphite cast iron parts based on local material properties), J.Wang (Knorr-Bremse (Kernel group))).

There are some empirical rules in conventional manufacturing that give some guidance on how to compensate for certain artifacts. The limitations are that they are empirically derived for casting or forging and there is no guarantee that conclusions can be transferred to additive manufacturing, and they only look at artifacts such as surface roughness or porosity, but there are no rules that consider the print orientation (parameters that are not present in casting or forging).

Some authors have explored the use of machine learning to predict the occurrence of artifacts (e.g., by monitoring sound output during printing). However, these papers focus only on predicting or detecting artifacts, without considering how they affect fatigue life and what mathematical model should be used to predict fatigue in the presence of artifacts. (see: https:// doi. org/10.1007/s 00170-018. sup. 1728-0).

Standard interpolation methods may be used without publications in the past, as this is the standard method. However, this approach has serious limitations. Firstly, multidimensional interpolation cannot be extrapolated outside the range of available test points, thus requiring a large amount of test data to be available over a large range, and secondly, only predictions within the bounds of the calibration points can be made.

Therefore, it is difficult to predict the fatigue life of the complex component. Several software vendors have implemented a partitioning approach in which a complex part is divided into multiple sections, with different material properties assigned to each section. However, this method does not describe how different characteristics of each section should be calculated.

Disclosure of Invention

The object of the present invention is to propose a machine learning method that solves this problem.

The described invention uses machine learning to more accurately model the fatigue life characteristics of an additively manufactured part without any a priori assumptions about how different parameters affect fatigue life. The method is implemented by a computer, and the invention thus comprises a computer program implementing the method being executed on computer hardware, in particular a personal computer, a server computer, a workstation or another type of computer hardware. In a preferred embodiment, the method and the computer program executed thereby implementing the method are used in a product design process. The product design process may employ the method in an iterative manner to optimize the product and/or product design. The method may also be used from the beginning of a product or production design. Thus, by using the system, method or computer program of the present invention, product design or production steps (i.e., additive manufacturing processes), a better product can be achieved.

In particular, this object is achieved by a method according to claim 1 and a system according to claim 8.

The solution comprises a system and a method for predicting fatigue life characteristics of an additively manufactured element, specimen or material structure in industrial product design or manufacturing, with a data collection step, wherein a plurality of data points of maximum stress versus fatigue failure cycle are collected for different given process steps, surface and volume conditions of the element, specimen or material structure, with a training step, wherein a machine learning system is trained using the collected data, and with an evaluation step, wherein the trained machine learning system is faced with new actual process conditions and used for predicting fatigue life characteristics of the element, specimen or material structure, and with an application step, wherein the predicted fatigue life characteristics are used for designing the product or for determining parameters of the manufacturing process.

In particular, the evaluation steps may be repeated in an iterative product design or production optimization process until a desired fatigue life characteristic of the additively manufactured element, specimen or material structure ("product") is achieved.

The task is also solved by a system for predicting fatigue life characteristics of an additive manufactured element, sample or material structure of an industrial product design or manufacturing process, the system comprising computer hardware programmed with a machine learning system, the system being arranged for performing training steps and evaluation steps as described before, the system being arranged for providing the results of the evaluation steps to the industrial product design or manufacturing process.

The system includes computing hardware (e.g., PC, server) with a machine learning environment and simulation software like Simcenter 3D of siemens. Additional software for the system includes a CAD system for product design. Advantageously, the system is connected to an additive manufacturing system, such as a 3D printer with 3D printing software. The system also includes conventional peripheral devices such as a display, memory, keyboard, data interface, and the like.

The features and advantages of the dependent method claims also apply in a corresponding manner to the system.

Drawings

Fig. 2 shows a diagram depicting a workflow and invention.

Detailed Description

The invention starts with test data from experimental tests or numerical simulations provided by the user in a first block "calibration" (step 1a "experimental test data" and step 1b "numerical simulation data"). The test data from steps 1a, 1b are fed into the machine learning (step 2 "train machine learning model"). In an advantageous embodiment, a gaussian process regression with squared exponential covariance function is used. The machine learning algorithm is trained in step 2. In parallel, the end-user defines a use case (second box "use case definition"), its components are defined in step 3 "complex user components", and different zones corresponding to different parameters are defined in step 4 "user definition of zones and associated parameters" (e.g. zones machined for better surface roughness, zone … … containing increased porosity).

Use cases from the use case definitions implemented through steps 3 and 4 are fed to a custom durability solver ("durability simulation for additive manufacturing") that can interpret the parameters from step 4 "user definition of zones and associated parameters". Based on the training model from step 2, an automatic calculation is made for each zone in step 5 ("automatic assignment of fatigue characteristics in each zone by machine learning model") to define fatigue characteristics. Then in step 6 ("durability prediction with local machine learning based characteristics"), the durability solver calculates an accurate fatigue life by considering these specific characteristics.

The main specific advantage comes from the combination of step 2 and step 5 as part of the workflow, where machine learning (step 2) is combined with the zoning concept AAF (step 5) to determine accurate fatigue properties in step 6 and to take into account additive manufacturing related parameters of the complex part (by allowing different zones to be defined in a single part).

The system and method provide several advantages. The present invention does not require any prior/prior knowledge or empirical rules about how the different parameters considered are linked together to predict fatigue life. The invention is fully flexible to extend to a very wide parameter range without increasing the complexity of the user. The invention allows to account for localized phenomena in an additively manufactured component to better predict fatigue life. The invention allows providing the end user with better insight as to how the material behaves and what the effect of the various factors is, since insight can be generated that is not possible based on experimental data alone. The invention can accept the combination of experimental results and numerical simulation results as the input of machine learning. The present invention automatically provides a confidence interval for predicting fatigue life, giving the end user an indication of how reliable the prediction is and whether additional tests need to be performed to ensure performance.

These advantages are mainly based on several technical features. By using machine learning (in an advantageous embodiment, a Gaussian Process Regression (GPR) method has been used for initial technical verification without having to pinpoint the application for that particular method) for fatigue prediction of additive manufactured parts, only very limited assumptions are made about how the data should behave. Thus, the proposed method enables capturing complex behaviors present in an additively manufactured part without any additional user input. The only assumptions made are the definition of the normal distribution (because of the central limit theorem and the law of large numbers) and the covariance function.

GPR has no limit on the number of input parameters that can be accepted and therefore can operate on hundreds of parameters as easily as only 3 components. No additional user input is required to work for this purpose.

By considering the localization parameters (roughness, porosity … …) as input parameters for the GPR, the fatigue properties can be calculated in an element-by-element method instead of in a complete part. Thus, the complex structure may be divided into manageable and reusable parts.

GPR allows prediction of fatigue life for an arbitrary set of input parameters. Thus, the effect of individual parameters on fatigue life can be studied, even though these tests are not reproducible experimentally.

The GPR method does not distinguish between experimental or numerical test data. Thus, it can handle purely experimental, purely numerical or mixed combinations without increasing the complexity of the algorithm.

Since GPR is inherently based on a normalized distribution, the confidence interval is an inherent part of the method.

The user may perform a more accurate fatigue simulation explaining the complex response of the additive manufactured part without requiring expertise on the subject, which means that the solution requires less expertise.

The fatigue model may have localized characteristics (possibly automatically generated) to further improve the accuracy of the model.

Based on the present invention, it is possible to quickly provide a solution for accurate fatigue modeling of AM (additive manufacturing) components with minimal requirements on the end user (in addition to providing test data).

The present invention satisfies the need for an industrially feasible solution (i.e., taking the deficiencies into account). A new and unique aspect is the use of machine learning to account for typical additive manufacturing (localized) phenomena and their impact on fatigue life (step 2 in fig. 2). The outlined solution employs a gaussian process regression with squared exponential covariance functions. Combining the machine learning method with the localized features allows the component to have varying features (step 5 in fig. 2). Working on an element-by-element basis has been supported by expert durability solvers, such as Siemens' Simcenter 3D. The present invention adds value to this element-by-element approach because it enables the automatic calculation of the appropriate fatigue characteristics for each element or zone.

An alternative solution to the proposed problem is to replace the machine learning method with any other modeling method. Several approaches can be envisaged, despite their limitations and challenges. Nearest neighbor techniques tend to have very limited accuracy. Interpolation techniques require very large data sets before any useful predictions can be made. In addition, only predictions that are completely within the calibration point range can be made. Extrapolation in multidimensional space is not possible. Parameter identification for predefined models has the disadvantage that as the number of degrees of freedom (i.e. the parameters considered) increases, it is very challenging to define a mathematical model. Furthermore, it is known that the separation of variables is not a valid assumption for additive manufacturing, which further complicates the development of mathematical models. Even if a model can be determined, parameter identification is challenging and may require an optimization method with multiple iterations before convergence. Finally, if additional parameters need to be considered, the model will be difficult to expand and may need to be fully redefined requiring expert knowledge. A problem with adapting existing empirical rules is that most empirical rules focus on only a single parameter and are designed for conventional manufacturing. These models are not possible to apply to additive manufacturing without modification, and combining multiple rules would likely result in very poor correspondence, as separation of variables is not practical for additive manufacturing materials.