Anomalydetection for Radiography Images & Time Series Data

Abstract

This thesis presents and implements anomaly detection algorithms for x-ray radiography image (size: 1600 x 950 pixels), and UAV inertial sensor measurements (x, y, z linear accelerations and roll, pitch, yaw angular accelerations). Non destructive testing of fuel filled closed metal container is achieved by detection of detach and normal region in x-ray image. Whereas vibration and disturbance acting on UAV body, originating from UAV internal moving parts and wind, are identified using sensor data by anomaly detection algorithm. A dataset of 20000 image chips (size: 41 x 41 pixels) is prepared to learn radiography information model of filled metal container. A 33 layers convolutional neural network (CNN) anomaly classifier .is trained using stochastic gradient descent with momentum algorithm and cross entropy loss function. Cross-validation classification accuracy is improved to 94% from 82% by proposed threshold tunning method for CNN classifier output. For the UAV sensor data, an unsupervised three staged anomaly detection algorithm is implemented. That is based on nonparametric empirical data analytic estimators, namely Chebyshev inequality, multimodal density, cumulative proximity, and unimodal density. Results demonstrate that the implemented algorithm autonomously 10 detect both individual and group of anomalies successfully. Furthermore, fast fourier transform (FFT) is computed for two UAVs, Huua and Uaab one hour flight data to determine spectral components in frequency domain for three axes linear accelerations and angular rates. FFT of Huua sensor data identifies in x in x-axis angular rate,axis angular rate,axis angular rate,axis angular rate, axis angular rate,axis angular rate, axis angular rate, axis angular rate,axis angular rate, axis angular rate, axis angular rate,axis angular rate,axis angular rate, higher frequency components of band 18.75 – 20 Hz 20 Hz . Whereas Whereas Whereas Whereas Whereas FFT of Uaab sensor data identifies in x in x-axisaxis , y , y-axis, and zaxis, and zaxis, and z axis, and zaxis, and z axis, and z-axisaxisaxis accelerations, accelerations, accelerations, accelerations, accelerations, accelerations, accelerations, accelerations, higher frequency components of band 32 – 34 Hz, 16 – 18 Hz, and 7.5 – 17 Hz respectively. Next optimal number of clusters value Next optimal number of clusters value Next optimal number of clusters value Next optimal number of clusters value Next optimal number of clusters value Next optimal number of clusters value Next optimal number of clusters value Next optimal number of clusters value Next optimal number of clusters value Next optimal number of clusters value Next optimal number of clusters value Next optimal number of clusters value Next optimal number of clusters value Next optimal number of clusters value obtained obtained for Huua data is 3, 7, 8for Huua data is 3, 7, 8 for Huua data is 3, 7, 8 for Huua data is 3, 7, 8for Huua data is 3, 7, 8 for Huua data is 3, 7, 8 for Huua data is 3, 7, 8 for Huua data is 3, 7, 8 for Huua data is 3, 7, 8 for Huua data is 3, 7, 8; and for and for and for and for Uaab Uaab Uaab data data data data is 4, 8, 10 using is 4, 8, 10 using is 4, 8, 10 using is 4, 8, 10 using is 4, 8, 10 using is 4, 8, 10 using is 4, 8, 10 using is 4, 8, 10 using K-medoid, Fuzzy C medoid, Fuzzy Cmedoid, Fuzzy C medoid, Fuzzy Cmedoid, Fuzzy Cmedoid, Fuzzy Cmedoid, Fuzzy C-means, and Gustafson means, and Gustafson means, and Gustafsonmeans, and Gustafson means, and Gustafsonmeans, and Gustafson means, and Gustafson -Kessel Kessel Kessel clustering clustering clustering clustering clustering algorithmsalgorithms algorithms algorithms respectively. Finally visualizations of Huua and Uaab Fuzzy C-means clustered data are obtained using two major components from principal component analysis.