Evaluation of surface residual stresses in metals using machine learning combined with indentation technique
Residual stresses,which are formed in the structural components during various mechanical manufacturing processes(casting,cutting,and assembling,etc.),exert significant impacts on the mechanical properties of materials,particularly when dealing with corrosion cracking,fracture,and fatigue life.Therefore,it is essential to examine their magnitude and distribution.Recently,numerous methods were proposed to measure residual stresses,classified as mechanical/destructive(hole drilling and contouring methods,etc.)and physical/non-destructive(ultrasound,X-ray,and magnetoacoustics emission methods,etc.).Besides,a promising instrumented indentation technique has attracted keen academic interest for its micro-damage,low cost,and in-situ measurements.Previous studies indicate the indentation load-depth curve and true contact area increase with applied compressive residual stress and decrease with applied tensile residual stress.Based on these findings,efforts were made to measure surface residual stress,but there are still obvious drawbacks.First,the indentation is typically small,requiring the use of high-precision microscopic equipment for measurement,and is limited in structural engineering applications.Second,the established mapping model is based on the indentation curve obtained from the unstressed reference state of the same material,which is often complex in form and involves numerous fitting parameters.All of these issues may potentially affect the accuracy of residual stress assessment.To address these issues,we propose a new method for evaluating residual stresses on metal surfaces using machine learning combined with indentation techniques.First,a dimensionless functional relationship between the curvature of the indentation load-depth curve,plastic parameters,and surface residual stress is constructed based on dimensional analysis.Then,numerical simulations are conducted for conical indentation of metallic materials with 950 combinations of plastic parameters and residual stresses.The mapping network between loading curvature,plastic parameters,and residual stress are established by using BP(Back Propagation)and MLP(Multi-Layer Perceptron)neural networks,respectively.Based on this,we propose an instrumented indentation method for determining surface residual stress.Its validity is confirmed by comparing finite element predictions with machine learning results.The neural networks,trained on a dataset from finite element simulations comprising 950 combinations of material plastic parameters and residual stresses,show good agreement with finite element predictions,notably the BP network with a superior accuracy.Experimental validation on 18CrNiMo7-6 alloy steel,which had induced residual stress,along with a comparison to existing test data,reveals machine learning predictions made by both BP and MLP neural networks achieve an absolute error of less than 40 MPa.Our study offers an alternative method for evaluating surface residual stress,complementing existing residual stress detection techniques in engineering applications.
neural networkresidual stressinstrumented indentationfinite element simulation
国家科技期刊平台
NSTL
万方数据
