Research On Nondestructive Testing Of Wood Bending Mechanical Properties By Near Infrared Spectroscopy

Wood is an important construction material for shipbuilding and sports equipment.However,traditional mechanical experiment is destructive and time-consuming.It causes serious waste of wood resources.Therefore,it is necessary to propose a nondestructive and effective method for wood detection,which is an important scientific issue of engineering application.Aiming at the problem,near infrared spectroscopy nondestructive detection technology is used to establish quantitative model to predict modulus of rupture and modulus of elasticity for Quercus mongolica wood.The part of this paper includes sample preparation,sample selection,spectra preprocessing,feature optimization and nonlinear model.Firstly,for obtaining the necessary experiment data,the samples of bending mechanical properties were manufactured,and spectra as well as measured value by traditional mechanical tests were collected.Secondly,for using better samples to establish model,the method based on mahalanobis distance was proposed to eliminate abnormal samples,and Kennard-Stone method was used to divide calibration and prediction sets automatically.Then,for eliminating the influence of scattered light,baseline shift and noise,we proposed the spectra pre-processing method based on multiplication scatter correlation,first derivative and Savitzky-Golay smoothing.Thirdly,for removing the redundancy and irrelevant factors,the methods of SiPLS and PSO-SiPLS were used to obtain the better wavelength variable,and the methods of LLE-PLS and Isomap-PLS were used to obtain nonlinear dimension data.Finally,the linear PLS model has the limitation of nonlinear prediction.In view of the complex nonlinear relationship between near infrared spectra and the measured value,BP neural network and extreme learning machine(ELM)for nonlinear model were proposed.The results showed that ten abnormal samples were eliminated by mahalanobis distance.The calibration set was selected reasonablely,which samples were representative and uniform.The method of MSC-1stDer-SG was used to pre-processing,and the effect was the best when smooth window size setting 9.The spectra profile was more clear,and the absorption peak was more obvious.The methods were used to optimize the input data for PLS model,including SiPLS,PSO-SiPLS,LLE-PLS,and Isomap-PLS.They were better than traditional PLS model.Among them,Isomap-PLS was the best.For modulus of rupture,the prediction correlation coefficient was 0.924,and SEP was 9.610,and RPD was 2.617.For modulus of elasticity,the prediction correlation coefficient was 0.923,and SEP was 0.839,and RPD was 2.594.Isomap optimization feature was used as input data.BP neural network and ELM were used to establish nonlinear model.Compared with BP neutral network,ELM has the advantage of higher speed and generalization ability.For modulus of rupture,the prediction correlation coefficient was 0.952,and SEP was 7.671,and RPD was 3.279.For modulus of elasticity,the prediction correlation coefficient was 0.946,and SEP was 0.706,and RPD was 3.089.The near infrared model can be used in practical application of quantitative analysis and prediction.