Deep Learning hidden layers

by Pedro Perdomo

20 years ago, we implemented a virtual sensor using neural networks with no more than a couple of hidden layers. When I trained my first two-layer neural network to predict the content of methane in the bottom of a deethanizer column, I realized that more than one layer did not improve my virtual sensor results. Why no limit today in the internal or hidden layers? Why deep learning? 

First, computing power has been increasing, so training neural networks with many layers is now feasible. Second, why so many layers if not a big amount of data available . And third, if you just add layers, you will run into some issues, the neural networks will take a long time to train, some of the layers will become all zero and worst, no better results.  Researchers have developed “tricks” and techniques to get deep neural networks to work. Neural networks with many hidden layers have enabled dramatic improvements in hard problems difficult to resolve with others Machine Learning techniques, like language translation, image classification and speech understanding.  

THE VIRTUAL SENSOR
So let's get back to our virtual sensor.

We started this project with no data in advanced. Starting to collect data and wait for extreme conditions to occur, ie. operating the column and the plant  over limit conditions was not a good idea for operators. We needed to collect data, good amount of data, but with variety.  We tried to produce this by operating at limits , but it is not always possible. In each condition you have to wait for stabilization and in the other hand, you need to collect samples to the lab. These are real issues once you try it out. 

A colleague from process department handled a digital simulation of the column in a popular software. This model was recently certified by comparing with real data in a very good range of operating conditions. Wait a minute! . Why we don't run simulations offline, with no risk of real operations and get rich data.? So, we created a script to feed this offline digital model with rich amount of process conditions, but offline, no risk to get the plant down.