Virtual Interfaces – Responsive Adaptive (VI-RA)

Abstract

In Industry 4.0, the user is overwhelmed with all the information that is generated from the machines which are found on site. This creates a problem as informative irregular errors can get lost in the sea of data. Thus, Virtual Interfaces – Responsive Adaptive (VI-RA) was designed to make use of anomaly detection so as to identify these abnormal entries within a multitude of data. VI-RA is then able to visualize data for each zone on site to the user through an augmented reality app, allowing for interaction with the display to reveal information about the errors. VI-RA makes use of three main methods for anomalous data identification. The first method makes use of an autoencoder which is able to identify an anomalous field by passing data through a trained model and indicating which entries return the largest error. The second method uses a denoising autoencoder to reconstruct missing fields within the data set. This allows for comparison between an expected and generated output so as to single out the anomalous field. The final function generates a scatter plot from the autoencoder’s thought vector so as to produce a visual representation of the data set. The distances between the generated data points are then used to identify anomalous fields. When comparing the three methods, the denoising autoencoder proved to be the least consistent amongst the three, with its accuracy being very high when attempting to find easy anomalous points but decreasing drastically when anomalous points are hidden deeper within patterns. The autoencoder anomaly detection proved to be very accurate when handling easily identifiable anomalous points however its accuracy also slightly dropped as the points became more dependant on pattern. The Scatter Plot however, proved to be the most consistently accurate as anomalous points became more hidden. It also resulted in the highest overall accuracy making it the most promising anomaly detection method carried out. An augmented reality app is used so as to allow the user better visualization and interaction with certain achieved results. The scatter plot co-ordinates for each zone within the data set are sent over an API and received by the app which displays a three dimensional plot over a respective QR code for each zone. The points reveal information about their relative error occurrence when interacted with, allowing for outliers to be easily identified and their cause to be immediately evaluated.