Computationally Efficient Optimal Video Comparison for Machine Monitoring and Process Control

Abstract

Real-time algorithms are needed to compare and analyze digital videos of machines and processes. New video analysis techniques, for computationally efficient dimensionality-reduction, for determination of accurate motion-information, and for fast video comparison, will enable new approaches to system monitoring and control. We define the video alignment path (VAP) as the sequence of local time-and-space transformations required to optimally register two video clips. We develop an algorithm, dynamic time and space warping (DTSW), which calculates the VAP. Measures of video similarity, and therefore system similarity, are estimated based on properties of the VAP. These measures of similarity are then monitored over time and used for decision-making and process control. We describe the performance, structure, and computational complexity of a DTSW implementation, which is parallelizable and which can achieve the processing rates necessary for many video-based industrial monitoring applications. We describe two case studies of unsupervised monitoring for mechanical wear and for fault detection. Results suggest opportunities for boarder applications of video-based instrumentation for real-time feedback control, wear and defect detection, or statistical process control.