Wood Mechanical Strength Evaluation Method Study Using Near Infrared Spectra Technology Based On Deep Transfer Learning

Wood is an important component material in household,construction,transportation and other field.Rapid and accurate detection of wood mechanical strength can not only ensure the safe use of wood,but also improve the efficiency of wood using.With the increasing demand for wood in the global market and the urgent demand for saving wood resources under the background of environmental protection,it is urgent to find a method of wood mechanical strength detection with high detection efficiency,high accuracy,environmental protection and safety.Wood detection technology based on near infrared spectroscopy has the advantages of nondestructive,fast and pollution-free,and can effectively detect various properties of wood.In order to improve the detection efficiency,accuracy and generalization ability of wood mechanical strength based on near infrared spectroscopy,a method of wood near infrared mechanical strength detection based on deep transfer learning was proposed.In this paper,the mechanical strength of Acer mono and Tooth oak was studied as the main experimental material,and Birch was used as auxiliary experimental material.By choosing suitable spectral pretreatment method,the automatic classification method of mechanical samples was designed,and the effective feature extraction method was selected.The transfer-ability of mechanical strength knowledge between Acer mono and Tooth oak based on near infrared spectroscopy was demonstrated.Based on the knowledge of transferability,the optimization method of sharing layer of multi-task learning model were developed,and the prediction model of multi-species mechanical strength was constructed.The main research contents are as follows:In view of the negative effects of random noise,baseline drift,light path scattering and background interference in the original spectrum,the original spectrum was pretreated.The experimental results showed that the SG-SNV as the spectral pretreatment method can effectively suppress the high frequency noise,light path scattering and other noise interference in the original spectrum.Aiming at the shortcomings of BP neural network,support vector machine and gradient boosting decision tree classifiers in near infrared spectroscopy-based sample classification,such as inadequate classification accuracy and slow modeling speed,an improved sub-sampling gradient boosting decision tree classification model based on MDS-KS was proposed.Random sub-sampling method was replaced by MDS-KS sub-sampling method.Experiments showed that the improved subsampling gradient boosting decision tree based on MDS-KS can effectively improve the generalization ability and accuracy in the learning progress.Experiments showed that the improved gradient boosting decision tree had a better classification efficiency and accuracy than BP neural ntwork and support vector machine.Three kinds of wood specimens could be classified correctly.The average classification time of each specimen was 1.562 seconds.Because of the high dimension of near infrared spectroscopy and the large amount of redundant information in the whole spectrum,the accuracy of wood mechanical strength detection based on near infrared spectroscopy is vulnerable.A feature extraction method of near infrared spectroscopy based on particle filter is proposed.The effective characteristic spectra and the corresponding mechanical strength are regarded as non-linear dynamic systems,and the redundant spectra are regarded as noise.The PF-PLS prediction model is established.The coefficient matrix of PLS model was regard as the system state,the optimal estimation of the current system state and the effective characteristic wavelength points were obtained through continuous iteration of particle filter.Experimental result showed that the PF-PLS prediction model proposed in this paper had a better prediction effect than other commonly used feature spectrum selection methods.When the ratio of test set to training set was 2:1,the correlation coefficients of compression strength prediction of Acer mono and Tooth oak were better than 0.828,the root mean square error of prediction was less than 7.845,the prediction results of tensile strength were better than 0.860,and the root mean square error of prediction was less than 33.489.Due to the limited number of training samples,poor ergodicity and other negative factors,the mechanical strength prediction model based on single tree species training samples has mediocre generalization ability and narrow application scope.In order to improve the generalization ability and accuracy of the model,this paper focuses on the transferable of knowledge of inter-species mechanical strength based on near infrared spectroscopy.Aiming at the problem that spectral knowledge can not be transferred directly among tree species,a training method based on TCA-PCA spectral feature fusion and depth transfer learning was proposed,and a prediction model of wood mechanical strength for TCA-PCA depth transfer learning based on near infrared spectroscopy was constructed.When the prediction set ratio of training set was 1:9,the correlation coefficients of mechanical strength prediction of both wood were all above 0.85,the root mean square error of compression model was less than 3.7,and the root mean square error of tensile model was less than 26.8.Experiment showed that the knowledge of wood mechanical strength based on near infrared spectroscopy could be effectively transferred by this depth transfer learning method based on TCA-PCA.The prediction model had a good generalization ability and prediction accuracy.In order to effectively reduce the over-transfer phenomenon in TCA-PCA deep transfer learning prediction model and improve the generalization ability and output efficiency of the model,combining with the classifier designed in the previous paper,according to the proven tree species spectrum with transfer-ability,a multi-task learning based near-infrared mechanical strength prediction model for wood of multi-tree species is proposed.The model also proposed setting up an early stopping mechanism in the shared layer to optimize the generalization error caused by the over-training of the shared layer and further enhanced the generalization ability of the model.Experimental results showed that the proposed multitask prediction model had a good generalization ability and prediction accuracy.When the test set of training set was 1:1,the average correlation coefficients of compression and tensile strength prediction models optimized by shared layer were 0.926 and 0.925,respectively,which were 13%and 8%higher than those of non-optimized prediction models.The mean square root error of the prediction was 2.135 and 15.681,respectively,which was 7%and 38%lower than that of the non-optimal prediction model.