Vibration Analysis Utilizing Unsupervised Learning

Abstract Many manufacturing environments have implemented methods of collecting data from their processes relating to vibration, temperature, or sound. With the data stored, manufacturers can run analytics to plan maintenance schedules and track machine health. However, in many cases, these maintenance schedules and health tracking are largely reactionary, largely implemented through experience rather than through predicting the onset of critical events and taking measures to prevent them. This paper describes a case of using time series data analytics of vibration from an automotive paint shop PVC dispensing pump (doser) attached to a robot using a novel combination of unsupervised learning and feature extraction. The goal is the determination of healthy versus unhealthy data and the implementation of predictive maintenance on the machine cell. Since the robot is a multi-axis robot, direct application of traditional health monitoring methods is lacking; instead a combination of methods suitable to the multidimensional nature of the robotic pumping process is employed. The goal of the first phase of the project is to build the tools to aid in this feature extraction using unsupervised learning and begin to establish a baseline of healthy data versus unhealthy data or fault data. The doser cell has been monitored for six months gathering data from seven sensor sets. Traditional methods of data analysis such as spectral analysis through Fast Fourier Transforms (FFTs) were used to establish the capability of reading vibration signals before moving to feature extraction of the time series data. For feature extraction, a Gaussian Mixture Model is utilized for the learning and the model building. These methods utilized not only determine the vibration of each specific component, but also help differentiate between the nozzle flow rate and angle. In extracting these features from the data, patterns can be traced from the variation of each production process and differentiation can take place based on what is healthy and unhealthy data. The goal of the continuing process phase is to inform the predictive maintenance function to improve equipment uptime.