Predictive Analytics from Visualized Log Data

The SigSRF demo page now includes a predictive analytics demo that parses unstructured log files, extracts and interpolates time series data, combines into one multivariate, linearly sampled time series using a regression model neural net, then submits to a sliding FFT to produce an image representation suitable for input to a convolutional neural net.

Yea that sounds like some math and signal processing, so the short summary is converting log files into visual format data so it can be analyzed by deep learning. Part of the idea is to take advantage of the fast moving state-of-the-art in computer vision.

This demo came about due to real world (read that as "painful") experience with a system that had an error rate less than 10e17 and would occur only after several days, making the system extremely time-consuming to debug. The demo learns the pattern of stress conditions leading to the error, with the objective to analyze log data and give developers an idea of how to adjust their testing in order to increase the error rate to a few hours (instead of days !), and thus potentially save weeks of test and debug time.