Study On Dairy Cows Gait Feature Extraction And Early Lameness Prediction

Inner Mongolia Autonomous Region is an important dairy cow breeding advantage region and dairy product production and processing area in China.As the second major disease of dairy cows,lameness has a great impact on the economic benefits and welfare farming.This study is aimed at the difficulty of dairy cow early lameness identification in dairy cows,and the goal is to improve the ability of early lameness identification.The acceleration signals of cows' limbs were collected by wireless three-dimensional acceleration sensor system,and the methods of data preprocessing,feature extraction and analysis and model establishment were systematically studied.The main research contents and results are as follows:(1)For the various types of noise in the acceleration signal,wavelet threshold denoising method is used to denoise the cow's limb acceleration signal.The wavelet threshold rules and threshold function are improved in our work,the experimental results show that the improved wavelet threshold method has achieves good denoising performance.(2)As the current segmentation methods have problems of large workload,low degree of automation,and low accuracy,this paper proposes a cow stride segment method based on an improved subsequence dynamic time warping algorithm.In this method,the Euclidian distance of a two-dimensional gait sequence consisting of time and amplitude is taken as the distance matrix element,and the distance function is calculated by the accumulated cost value to find the minimum warping path and determine the stride segmentation point,so as to further realize the segmentation of a single stride of cow.The experimental results show that the improved algorithm's precision,recall,and F-score average values are all higher than 95%,which are 7.92,6.93,and 7.45%higher than previous algorithm,respectively.The comparison with other methods shows that the algorithm presented in this paper has high precision and strong generalization ability.(3)In order to improve the predictive ability of the model,gait inconsistency variables were taken into account.The kinem atic-and kinetic characteristics are extracted from the segmented stride and dimensionality is reduced.The new eigenvector is used as the model input variable.(4)For the problem of learning rate fixed of the iterative algorithm of cost function,a mini-batch gradient descent method which dynamically adjusts the learning rate is proposed.This method will not cause training divergence because of high learning rate,and will not affect the convergence rate of cost function due to low learning rate.Namely,dynamically adjusts the update gradient through the learning rate,and the cost function is converged to the minimum model coefficient through the training output to complete the establishment and optimization of the logistic regression model.The experimental results show that the convergence of the cost function is better,and the recognition accuracy is 90.64%.(5)The best recognition model is determined by k-fold cross validation.The test set is identified and classified by the model and the experimental results show that the recognition accuracy is 91.83%,the sensitivity and specificity are 91.49%and 92.31%respectively,which are 4.08%,4.55%and 3.85%higher than the previous.The experimental results of diferent classifiers show that the AUC value of the model in this paper is 0.94,which is 3%and 8%higher than SVM and KNN,respectively.Therefore,the optimization model improves the ability to recognize dairy cow early lameness.