EFT's CORTEX™ software brings predictive analytics to the manufacturing process. Through machine learning and real-time visualizations, EFT enables manufacturing industries to predict, optimize processes, pinpoint causes and improve productivity in the operating environment to boost profitability.
Welcome to EFT's Cortex software. Cortex is a powerful, easy to use SAS platform that uses machine learning and real-time visualization to deliver insights for data-informed actions. This video will demo the platforms capabilities and an example of how it's worked for one of our clients. Sound good? Great. Let's get started.
Once we tap into Cortex, we'll want to click, "Launch Network Wizard". The first step is to select how we're going to get data into the platform. Now, I already have a data set here, so I'm going to select "Existing Data Set", but we can also upload a CSV file using File Upload Wizard, or pull straight from you Pi Historian. Cortex is equipped with EFT's exclusive data cleanse, prep, and enrich feature. This means you can load your data as is and Cortex will clean it, detect outliers, input missing values, handle categorical variables, and enrich your data with computer variables producing high-quality models. Yeah, it's pretty cool and makes your life a whole lot easier. In that same vein, EFT can build connects to other historians if needed.
Once we name and upload our data set we'll advance to selecting our network type. Here you'll see a list of machine learning algorithms that we can utilize in the platform. That includes [Embasia 00:01:22] networks and other classification and regression models, a list that will continue to grow as Cortex is enhanced. On the left, you'll see a network type called MaGE. This is an EFT proprietary algorithm that eliminates the need for you to hunt for the best model. MaGE builds a variety of models in the platform as well as an ensemble model and provides scores for all of them. It's a powerful tool, and one I recommended you utilize. However, for the sake of the demo and the data set I have uploaded, I'm going to do a top ten naive Bayes model.
Once your network type is selected, it's time to configure our test set. I'm going to use a random split percentage, which will take 70% of the rows in the data set and train the model based on that. The 30% left over will be tested after the model is built to give us our accuracy score.
Now, it's time to configure our variables. We're going to select all variables. This data set is representative of a basic distillation column. And our variable is going to be tons of steam per tons of production. From here, we have some more options. We can do a K-means clustering, or select costume bends, and adjust the bends size from one to 35 depending on your need. With bands, we split the data into ranges. I'll explain in more detail when we view the model. For this demo, we're going to use K-means and three bands for each of the variables. I've found this is a good place to start.
Once you experience Cortex a little more and gain a better understanding, feel free to further refine these variables. Then, you click submit and Cortex will begin building your model. Thanks to the intuitiveness and robustness of the platform, this building process should only take a few minutes. While Cortex is working it's magic I'm going to take you to the training sets page.
Here's the training set we just built in the network wizard. This is built from the random 70% of the data set we put aside earlier. This relationship data table is a helpful piece of information in the platform. Based upon your variables, it calculates the value of mutual information. Essentially, it rates the correlational between the selected variable and the others in your data set. In this example, we can see that production rate is highest ranked, followed by reflux ratio and so forth down the line. So, if you have a ton of variables this is going to be super helpful for you to narrow your focus and give yourself a great starting point.
Now, it's time to check out the model Cortex has created based on our data set. The green number is the score, so it was about 83% accurate when it took the test set of the random 30% we selected. Not the best, but not the worst. From here, we could try to refine our model to gain better accuracy. To view an interact with the historic model, we click on "Model Complete.". The best way to showcase the valuable outcomes that can be produced with a model is to share real data from one of our real customers, and how they used the insights gained from the Cortex platform to drive actionable solutions for their business. And while this case study is from the steel industry, our software is flexible and proven to help many other businesses in a variety of industries.
Big River Steel was having an issue with product defects, which is, unfortunately, a common occurrence in the industry. And while steel process experts understand key variables impacting defects, it's often difficult to pinpoint a clear cause. That's where we came in. Now, whereas the demo data set had five variables, Big River Steel had 4,000. So, just a little more. Using Cortex and their subject matter experts, we were able to narrow it down to 25 key variables which is what you see now. Each box represents a different variable. Together, we were able to identify longitudinal cracks, shown here as LC's, as the target variable. The numbers on the left of each box are the bend ranges. This allows us to take a continuous variable and split into ranges, so we can compare continuous variables, with categorical variables. So, this is where the K-means clustering or costume bending method options come into play from earlier.
For Big River Steel, we started with K-means and later used a costume method. The blue bars within each box represent the probability of the data being in that bend range. So, for instance, there's a 50% chance that this liquidus temperature is going to be between 2,781, and 2,789. The cool part about this is that we can perform some what-if analysis but applying evidence here. As an example, we can select a variable from the drop-down menu. I'm going to select TOnde superheat, which is the amount of degrees higher than the liquidus point of the steel. Essentially, this is how much hotter the steel is than when it would begin to go from a liquid to a solid.
From here, I'm going to select bend ranges but using the "Select State" drop-down menu. This is a What-If analysis of when the TOnde superheat is in this range and how the other variables in the model interact. From here, we can examine the change in these probabilities. We can also select "Show Previous Values" so we can see the difference. And wow, what a difference in LC's it made. It went from a 4.8% chance of having a longitudinal crack, to an almost 20% chance. We can also go through and try out different ranges and see how that affects our probability. The optimum temperature range that created the lowest probability of having an LC, is the 53-75 bend range.
Just like we applied evidence to once variable, we can apply evidence to all of the variables. As you can imagine, with numerous variables in many ranges, this becomes very complex. That's where the real-time monitor comes into play. This monitor brings your real-time data into Cortex and gives you the output of the model you created. The green line that turns to red is the probability of being in your target bend. In Big River Steel's case, the green line shows the likelihood of not having a longitudinal crack. If that decreases and goes past the threshold of 50% that we set, the line turns red and creates an alert below. The orange line is the actual value coming into the platform.
So, what did this mean for Big River Steel? Another good question. Before we had this model implemented, there was about 30-40 minutes before they realized they made a product with a defect. Now, when an alert is created they can see the possible actions for what they could change to go back to the state before it predicted they were going to have a crack.
This real-time monitor also contains a more in-depth look at the alert. All you have to do is click the "Details" tab. Here, the variables that have changed are noted by arrows. A red box indicates a drop, a green box indicates a rise, and a gray box with a line shows no change. There's also the very helpful and dynamic "What-If" analysis tab. Here, you can change variables and see how that would affect the probability of the model output in real time. For example, if we were to take this left side tapper and move it to another band, it adjusts the gauged value of the rest of your variables. The dark blue bar is the probability of the current data. The lighter blue bar represents the probability distribution of what you've changed below so you can compare and contrast.
There you have it. In a nutshell, that's Cortex. EFT's cutting-edge software that gives you real-time machine learning and process optimization so you can take action on insights that deliver profitable outcomes. We hope you enjoyed this demo. If you have any questions, or if you'd like to learn more about our software and our services, don't hesitate to reach out. Thanks again for watching, and have a great day.