Research On The Detection Method Of Farmland Soil Shear Strength Parameters Based On Machine Learning

Soil shear strength parameters,including cohesion and internal friction angle,are important mechanical indexes of soil,which can be used to evaluate soil erosion sensitivity and the tillage performance of the cultivated layer of soil.Soil shear strength is closely related to the deformation,resistance and compaction of the soil,and determines tillage quality and energy consumption to a large extent.The existing methods for measuring soil shear strength parameters,such as indoor direct shear test and triaxial compression test,are complicated,time-consuming and difficult to be applied on a large scale,and the stress conditions of field test methods,such as shear test of cross plate,are not easy to master.It is a difficult problem for agricultural production management to obtain soil shear strength information quickly and accurately.Aiming at the above problems,this study carried out research on soil multi-sensor data acquisition device and shear strength parameters prediction model.Multi-sensor fusion technology was used to obtain data information related to soil shear strength parameters,and machine learning algorithms were used to build prediction models to achieve quick and accurate prediction of soil shear strength parameters.The main research contents are as follows:(1)Based on multi-sensor fusion technology,a soil parameters acquisition system is designed by using STM32 single chip microcomputer as the core processor.The system is mainly composed of a self-propelled mobile platform,an actuator and a measurement control system.The actuator includes conical penetration component and soil moisture content detection component.A column pressure sensor and a film pressure sensor were used to collect the cone tip resistance and cone side pressure of the conical penetration component penetrating into the soil,and an FDR sensor was used to collect the soil moisture content.The mobile phone monitoring software was designed based on wechat developer tools to realize wireless transmission,storage and real-time display of device measurement data.The sensors calibration tests results show that the fitting coefficients between the input-output response relations of the column pressure sensor,the film pressure sensor and the FDR sensor are 0.9974,0.9867 and 0.9694,respectively,which meet the data acquisition requirements.(2)The sensing data of 83 soil samples were obtained by the soil parameters acquisition system,and the cohesion and the internal friction angle of these soil samples were measured by the direct shear tests.The data preprocessing research was carried out,and the filtering effects of four different filtering methods,including median filtering,mean filtering,S-G convolution filtering and Butterworth low-pass digital filtering,were compared and analyzed.The results show that Butterworth low-pass digital filtering has the best filtering effect.The maximum response value of the sensor in the data study area was extracted as the eigenvalue,and the eigenvector space of soil shear strength was constructed.And the correlation between each feature vector and soil cohesion and internal friction angle was analyzed based on Pearson correlation coefficient method.The results showed that the correlation coefficients(r~2)of soil moisture content,cone tip resistance and cone lateral pressure with soil cohesion were-0.69,0.74 and 0.77,and the correlation coefficients(r~2)of soil moisture content,cone tip resistance and cone lateral pressure with soil internal friction angle were-0.68,0.65 and 0.84,respectively.The eigenvectors have obvious correlation with soil shear strength parameters.The detection effects of three different abnormal sample identification methods,namely Monte Carlo Cross Validation(MCCV),Leave-One Out Cross Validation(LOOCV)and Mahalanobis distance,on soil eigenvectors space were compared and evaluated.The results show that the MCCV method has the best effect on detecting abnormal samples,and the number of excluded abnormal samples is 1,35,47 and 83.The data acquisition and preprocessing provide basic data for the next step of building the prediction models of soil cohesion and internal friction angle.(3)The main influencing parameters of soil shear strength prediction models constructed by extreme learning machine(ELM),support vector machine regression(SVR)and partial least squares regression(PLSR)were analyzed.Based on the average coefficient of determination and root mean square error,the network structures of ELM cohesion model and internal friction angle model were determined as 3-8-1 and 3-10-1,respectively.The penalty factor(p)and kernel parameter(g)of the SVR cohesion model and the SVR internal friction model were optimized by the5-fold cross-validation method combined with the cross-validation mean square error(MSECV),respectively.For the cohesion modeling parameter p and g are 2.06 and0.004,respectively,and the modeling parameters p and g for the angle of internal friction are 0.20 and 0.45,respectively.The number of modeling principal component factors(PCF)of the optimal PLSR cohesion model and PLSR internal friction angle model is determined by Akaike Information Criterion(AIC),which are 3 and 2,respectively.(4)A combined ELM-PLSR prediction model was proposed by combining ELM algorithm and PLSR algorithm in parallel weighted combination method.The prediction performance of four different machine learning algorithms,ELM,SVR,PLSR and ELM-PLSR,for establishing soil cohesion and internal friction angle were compared and evaluated.The modeling comparison results of the four algorithms show that the ELM-PLSR algorithm has the best modeling performance.When detecting soil cohesion,the ELM-PLSR model evaluation grade was"excellent",and the prediction indexes of the test set were 0.92 for R_v~2,0.92 for RMSE_v and 3.47 for RPD_v,respectively.When the soil internal friction angle was detected,the ELM-PLSR model evaluation grade was also"excellent",and the prediction indexes of R_v~2,RMSE_v and RPD_v were 0.91,1.54 and 3.30,respectively.The ELM-PLSR model can effectively improve the measurement accuracy of the PLSR model,make up for the generalized defects of the ELM model,and provide a reliable relationship model for the detection of soil shear strength parameters.