A neural network approach to describing the fretting fatigue in aluminium-steel couplings
Introduction
The problems that occur with mechanical joints which are subjected to dynamic loading are very complex and heterogeneous. One of the most important of these problems is fretting fatigue, which occurs when a minute tangential surface motion arises between components pressed together by normal forces. A significant decrease in fatigue life—by a factor of 10 or more—occurs as a result of fretting, and fretting is one of the main causes of failure in a wide range of engineering applications, including the following: blades and disks in turbine engines; clutches; riveted compounds; bolts and couplings; drive shafts; bearings; wheel hubs; steel ropes; etc. [1], [2], [3].
There have been numerous studies of fretting phenomenon [4], [5], [6], [7] and several key factors—surface pressure; relative motion; loading frequency; stress ratio and amplitude; loading waveform; specimen size; metallurgical composition; surface finish, and many others—have been investigated to determine their influence on fretting fatigue. The influence of these factors on the fretting involves complicated synergistic interactions [3], [4] and accurate databases. In order to establish a functional dependence in a conventional way it would be necessary to vary one factor while all the others are kept constant. Unfortunately, despite remarkable progress, the existing fatigue databases are incomplete for most factors, and the fretting fatigue of mechanical joints made up of steel and aluminium alloys is particularly difficult to describe [2], [5], [6].
Artificial neural networks (ANNs) have evolved over the past few years as a relatively new area of artificial intelligence [8], [9], and they are suitable for applications in a wide range of fields. ANNs are especially useful for simulations of correlations that are difficult to describe with physical models because of their ability to learn by example and to recognise patterns in a series of input and output values from example cases.
ANNs are very capable when it comes to describing fretting-fatigue phenomena. The main advantages of ANNs over other, more conventional, types of description are that the description is based purely on data from the database, not on preconceptions. ANNs can generalise and learn trends and patterns and even though factors affecting the fretting fatigue exert synergistic effects on one another, ANNs can detect and utilise these effects in their descriptions [10].
The aim of the research was to show, using examples of fretting fatigue with steel and aluminium-alloy couplings, that ANNs can acquire the same knowledge that it took the combined efforts of many researchers to acquire, but in a shorter time and with the use of a smaller database.
Section snippets
Experimental details
The fretting-fatigue database for training and testing the ANNs contained 114 experiments. The database included studies of the effects of surface pressure and surface treatment on the fatigue life of aluminium alloys in contact with steel that were subjected to dynamic reverse bending. It covered a modest range of factors: two aluminium alloys (AlSi7Mg-T6 and AlSi11MgSr), two surface-finishing conditions (machined and machined plus shot-peened), different stress amplitudes (ρΔ = 48–138MPa) and
Neural network approach
ANNs are computing systems composed of highly interconnected, but very simple, processing elements (or neurons) that process data using the ANN's response to external inputs. The primary characteristics of ANN models are: massive parallelism, non-linearity, processing by multiple layers of neurons, and, finally, dynamic feedback among neurons [16], [17]. Due to their parallel structure, which is inspired by the parallelism of the human brain, ANNs can be used for a range of multivariable
Neural network estimates
To describe the fretting fatigue of aluminium alloys in contact with steel that were subjected to dynamic reverse bending, a model based on ANNs was created. The ANN was trained and used to predict the occurrence of fretting fatigue for various sets of input parameters. A confusion matrix (Table 5) shows the prediction of fretting fatigue for 28 tests from the database never seen by the ANN. The ANN incorrectly classified three fretting and one no-fretting tests and the overall classification
Results and discussion
A small database was the major handicap that prevented good ANN training and, as a consequence, a good fretting-fatigue description. In order to prevent overtraining of the ANN a cross-correlation subset was used. The use of the cross-correlation subset, at the expense of 15% of the entire database, as an indicator for the best training solution, turned out to be the optimum solution. The 84% accuracy level of the fretting-fatigue prediction in the training and cross-correlation subsets (Table 6
Conclusions
The neural network approach shows the tremendous potential that ANNs have for describing the fretting-fatigue phenomena of aluminium–steel couplings that are subjected to constant dynamic reverse bending. The ANN we used gave us an accurate description of the influence of different parameters on fretting fatigue and also good predictions for the occurrence of fretting fatigue based on input parameters. With the trained ANN it is possible to predict the occurrence of fretting fatigue for
References (34)
- et al.
A review of analytical aspects of fretting fatigue, with extension to damage parameters, and application to dovetail joints
Int J Solids Struct
(2001) Peak contact pressure, cyclic stress amplitudes, contact semi-width and slip amplitude: relative effects on fretting fatigue life
Int J Fatigue
(2001)- et al.
Mechanics of fretting fatigue crack formation
Wear
(1996) - et al.
The use of neural networks for the prediction of fatigue lives ofcomposite materials
Composites Part A
(1999) Analysis of learning rate and momentum term in backpropagation neural networks algorithm trained to predict pavement performance
Adv Eng Softw
(1999)- et al.
Fatigue life prediction of unidirectional glass fiber/epoxy composite laminae using neural networks
Compos Struct
(2001) - et al.
Artificial neural network technology for the data processing of on-line corrosion fatigue crack growth monitoring
Int J Pres Ves Pip
(1999) - et al.
Fatigue life calculation of flawed structures–based on artificial neural network with special learning set
Int J Pres Ves Pip
(1998) Artificial neural network technology as a method to evaluate the fatigue life of weldments with welding defects
Int J Pres Ves Pip
(1995)- et al.
Prediction of corrosion-fatigue behavior of DP steel trough artificial neural network
Int J Fatigue
(2001)
Neural networks in mechanics of structures and materials—new results and prospects of applications
Comput Struct
Neural network applications in determining the fatigue crack opening load
Int J Fatigue
Failure of carbon/epoxy composite tubes under combined axial and torsional loading, 1. Experimental results and prediction of biaxial strength by the use of neural networks
Compos Sci Technol
Artificial neural networks for predicting the macromechanical behaviour of ceramic-matrix components
Comput Mater Sci
Simulation of microhardness profiles for nitrocarburized surface layers by artificial neural network
Surf Coat Technol
Modeling the correlation between processing parameters and properties in titanium alloys using artificial neural network
Comput Mater Sci
An example of the use of neural computing techniques in materials science—the modelling of fatigue thresholds in Ni-base superalloys
Mater Sci Eng
Cited by (44)
A machine learning method for HTLCF life prediction of titanium aluminum alloys with consideration of manufacturing processes
2023, Engineering Fracture MechanicsArtificial neural network combined with damage parameters to predict fretting fatigue crack initiation lifetime
2022, Tribology InternationalCitation Excerpt :A new computational tool, artificial neural network (ANN), has been successfully developed and applied in the field of fatigue [13–18]. Orbanić and Fajdiga [19] proposed a four-input ANN method to predict fretting fatigue life under constant dynamic reverse bending. Nowell and Nowell [20] studied the influence of contact size on fretting fatigue through an artificial neural network model, where the input parameters are common variables in fretting.
A hybrid ANN-multiaxial fatigue nonlocal model to estimate fretting fatigue life for aeronautical Al alloys
2022, International Journal of FatigueCitation Excerpt :However, with respect to fretting fatigue there are only a few articles that consider the use of ANN tools. For instance, the work by Orbanić and Fajdiga, [24] proposed a four input ANN model to predict the occurrence of failures by fretting fatigue for two aluminum alloys under constant dynamic reverse bending and for two different surface finishing conditions, for various levels of pressure and load ratios. Majzoobi and Kazemi [25] used ANN to analyze the effect of re-shot peening on the fretting fatigue behavior of Al 7075-T6.
Machine learning predicts fretting and fatigue key mechanical properties
2022, International Journal of Mechanical SciencesCitation Excerpt :In the last few years, machine learning based models were developed to predict the fatigue life [22–24], fatigue crack growth [25–29], and fretting wear [30, 31]. Orbanic and Fajdiga [32] trained an ANN to describe the fretting fatigue phenomenon based on 114 experiments for different material-surface-finishing conditions and surface pressures under various stress amplitudes; their model could successfully predict almost 85% of the experimental dataset. Anand Kumar et al. [33] developed a neural network model to describe fretting wear behavior as a function of normal load and the surface harnesses of both contacted materials.
Estimation of seismic wave velocities of metamorphic rocks using artificial neural network
2019, Materials Today: Proceedings