Industrial Automation

        By Pedro Perdomo   When you want to have a digital model of your process, mainly you have two options, real three but let’s take the first two. You can use the first principles of physics and try to get a fairly, replicable, and “precise” math model. On the other hand, you can use data to train a machine-learning algorithm and get a math actuation,  a black box,  but certainly wit an embedded rich acknowledge of your process. The first option needs subject matter experts who understand the thermodynamics of the process. The second just need data experts.  Which option to take? Why? In engineering, mathematics does not have all the solutions. On many occasions, we understimete variables in an math equation looking for a practical solution, close to reality. We do NOT struggle to calculate milligrams if at the end, we can NOT measure milligrams. Making a mathematical model of an industrial process suggests posing primary equations, well-un...

by Pedro Perdomo A Virtual Sensor Implementation using digital twin models . You need data to train a virtual sensor. This was the main problem to surpass in this project.  So , how do you pretend training  a neural sensor without data collected? 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.  We had a digital model  of the column in a popular simulation software.  This digital model was recently certified by comparing with real data in a very good range of operat...

The third model in the picture fit better the predictor variable. It models the training data so well !   But it is not the best one ! Putting it into production will produce big prediction errors over new data. “Overfitting is a modeling error that occurs when a function is too closely fit to a limited set of data points. Overfitting the model generally takes the form of making an overly complex model to explain idiosyncrasies in the data under study “ (1)   . Machine Learning is not to compete with a formal math function. It is not a good idea to compare with formal math equations. Machine learning algorithms describe the world in some different way, not fully comprehensible, but for sure most of the time even more accurate than math models developed by scientist.

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 p...