Research On Soil Characteristics And Data Mining Based On Ultra-wideband Signals

UWB radar sensors have been widely acclaimed and applied to the research field of soil characteristic parameter extraction on account of their satisfactory range resolution,impressive radar penetration capability,and relatively low energy requirements.In this paper,two soil characteristic parameter classification algorithms and two soil characteristic parameter prediction algorithms are proposed for the extraction of soil characteristic parameter information from ultra-wideband radar signals.The application of a variety of machine learning algorithms appeared in this paper.The strategy of signal feature mining includes the method of time domain information flow,energy in different frequency bands and signal maximum separable mapping.The scientific mapping between soil echo data and soil characteristic parameters has been revealed.The main work content is summarized as follows:1.This paper proposed the UWB-LSTM and UWB-GRU soil water content classification algorithms to achieve the purpose of classifying five types of soil UWB radar data with different volumetric water content without feature extraction.The core innovation of this strategy is to use the data stream characteristics of the soil echo signal.Two special cyclic neural networks with powerful information flow processing capabilities are used to classify soil echo information.The two algorithms performed classification experiments on real measured data,and the training convergence efficiency,generalization application ability and classification accuracy of the model were compared.2.Based on confident learning,this paper proposes a special logistic regression algorithm.The error in the label is the main problem to be solved by this algorithm.The advantages of confident learning is utilized to strengthen the error label processing ability of the logistic regression algorithm.The classification advantage of this algorithm is reflected in the actual measured data with incorrect labels.This algorithm has the ability to find possible wrong labels by estimating the probability of label noise.In addition,this article compares the classification accuracy of the logistic regression model based on confidence learning and the general logistic regression model under different signal-to-noise ratios.Simulation analysis shows that The new algorithm is more capable of coping with the negative effects of noise tags under different SNRs.3.In this paper,a soil characteristic parameter prediction algorithm named WPTSVM is proposed.The core advantage of this program is to creatively establish the relationship between the data and soil characteristics with the frequency domain characteristics of the soil signal.WPT is used to decompose signals with high frequency resolution.The soil p H value and VWC prediction based on the frequency domain characteristics of soil echo data are responsible for the support vector regression scheme.This processing method not only solves the problem of soil signal frequency feature mapping,but also makes quantitative analysis of soil features possible.We think this breakthrough is very meaningful.4.This paper proposes a soil p H prediction algorithm based on PCA+ emsemble learning.A data feature mapping direction selection scheme based on the maximum separability of data is applied to soil data feature extraction.The algorithm not only can effectively predict the soil characteristic parameters,but also has certain resistance to signal noise.The anti-noise performance of the model has been verified under a variety of SNR conditions.We use the soil data in the real environment to find the best combination of model algorithms.In addition,using the characteristics of the ensemble learning algorithm to integrate multiple weak learners,the energy characteristic spectrum obtained by WPT is analyzed,and the importance of the signal energy of each frequency band to the soil p H prediction is ranked.The results show that: When using the three integrated model analysis,more soil characteristic information may be contained in the frequency range of 3.84-4.10 GHz.This paper combines traditional signal processing algorithms and uses the powerful data fitting capabilities of machine learning to extract features from the three perspectives.Solutions are proposed for the tasks of soil p H value and volumetric water content prediction,soil characteristics classification and prediction under signal noise environment.These studies are of great positive significance for the large-scale monitoring of soil characteristics.