| The knee joint is an important part of the body,which is easy to be aged and damaged.Incidence rate of knee joint disease is high.Early screening of knee joint disease is helpful to control the disease in time and prevent the deterioration of the disease.Vibroarthrographic(VAG)signal refers to the sound or vibration signal when the knee joint flexion or extension.VAG signal can sensitively and objectively describe the status of knee joint,which plays an important role in the non-invasive detection of knee joint diseases.The existing classification methods for VAG signal have low degree of automation,and the overall performance of specific disease classification needs to be improved,which needs further research.In consideration of the above problems,this dissertation cooperats with the Department of orthopedics of Dalian Zhongshan hospital to collect the clinical VAG signal data firstly,and build the VAG signal database,including 222 healthy cases,176 cases of osteoarthritis,86 cases of meniscus injury,103 cases of cruciate ligament injury,67 cases of patellar arthritis,a total of 654 cases of VAG signal data.Then,based on the database,combined with machine learning and deep learning theory,this dissertation systematically studies the two classification and multi classification of VAG signals.(1)In view of too many feature parameters in traditional machine learning,and the classification effect is not good,a classification and recognition method based on feature sorting is proposed.The dissertation extracts 20 characteristic parameters of VAG signal comprehensively based on time domain,frequency domain,time-frequency domain,nonlinearity and statistics,and uses Feast feature selection method to sort the feature parameters effectively.Combined with four different classifiers,the data sets constructed by different features and number of features are classified and studied.The best VAG signal classification results is obtained.Through this method,the redundancy and mutual exclusion of features in the classification of VAG signals are avoided,and the features are better combined to improve the classification accuracy and reduce the classification time.The experimental results show that the accuracy,sensitivity and specificity of SVM classifier are91.62%,92.54% and 91.47% in the data set composed of 8 features.This method can reduce the redundancy of feature and improve the normal or abnormal of VAG signals’ classification and recognition results.(2)In view of the low accuracy of using original signal classification in machine learning,although the method of extracting feature can improve the accuracy,it needs manual extraction,low automation and the quality of features directly affect the classification results.This dissertation proposes a classification algorithm for VAG signals based on improved convolutional neural network serial recurrent neural network(PCNN-LSTM).Firstly,empirical mode decomposition(EMD)and wavelet transform are used to transform one-dimensional VAG signal into two-dimensional time-frequency characteristic spectrum as data set;Secondly,an improved PCNN-LSTM model is constructed by fusing the parallel CNN network structure and LSTM neural network on the basis of serial neural network,so as to classify the normal or abnormal VAG signals and realize the automatic detection of knee.The experimental results show that the accuracy of the model is 96.93%,the sensitivity is100% and the specificity is 95.56%.The model has better classification effect than other algorithms,and can effectively realize the automatic classification and recognition of normal or abnormal VAG signals.(3)In view of the problems of VAG signal data shortage,and the multi classification effect is not good,a new multi classification algorithm of VAG signal based on small sample learning is proposed.After constructing two-dimensional time-frequency characteristic spectrum as data set,convolutional neural network model is used to pre train,map samples to high-dimensional embedded space,find the class prototype of each class,and propose a calculation method to measure the distance between test samples and various types of prototypes in embedded space,and classify and identify them in multiple categories.The results show that the accuracy,sensitivity and specificity of the model are 88.35%,89.33% and 88.19%,respectively.It provides a way of thinking and method for the research and clinical application of VAG signal,and promotes the realization of VAG signal applied to the noninvasive detection of knee joint to assist diagnosis. |
