In manufacturing today, data analysis tools can give management the information they need to make better decisions in areas such as maintenance and labor. Unfortunately, however, many data analytics systems require large sets of historical data in order to generate accurate and useful results.
According to Dr. Rebecca Grollman, Data Scientist at Bsquare, anomaly detection is different. These algorithms can begin generating useful information without needing to be trained on historical data. While simple, anomaly detection can be used for applications such as detecting machine stoppage, sensor malfunctions, tracking production output, and more. Engineering.com recently spoke with Grollman about this solution.How essential is historical data in typical data science applications?
If you want to answer questions that are very specific, historical data is essential for starting a data science project. Especially if you want to use something like machine learning or deep learning, you really need to have that historical data set. However, if you just want to monitor the data and look for things that look different or not in the ordinary, then that's where anomaly detection can come in. For anomaly detection, you don't need the massive amount of data as you might for some of the other algorithms that can be used.
What is special about the way Bsquare does this?
Bsquare has developed an internal anomaly detection tool that will help us quickly see your data. We can quickly determine which type or anomaly detection algorithm is best for your data, and what types of parameter settings we should use for that algorithm. So, with these things we can quickly see what's going on with your data, where the anomalies are, and we also a have an interactive tool that will let us visualize the data and possibly even share it with the customer to evaluate what it shows.
That being said, working with our customer’s subject matter experts is key here. I can receive a data set that is just a lot of columns and numbers, but I don't know how it relates to the manufacturing equipment, and it’s just all meaningless. So being able to work with the subject matter experts and understand what their numbers mean, what data are they collecting, what sensors are they looking at, being able to see the whole picture is critical. So that's another thing that we highly value when working on anomaly detection project, or any data science project.
This internal tool is used when you’re setting up a new project to identify their right algorithm to use, but it wouldn't be used throughout the project, correct?
Correct, and then from there we've got other tools to implement the algorithm in production.
What kind of business impact can anomaly detection have for manufacturers?
First is understanding what your normal operating parameters are. Especially when you’re starting with very little data and you don't know what to expect—from a new piece of equipment, for instance. Just being able to pinpoint the normal operating parameters of the data is very important. Once you understand what is normal and what anomalies there are, you can start to find patterns in the anomalies and possibly even build rules that act on the anomalies that you're detecting.
There's actually another thing that we have a tool for. We can generate potential rules that a customer could use. For example, if an anomaly is being detected at around the same time each day, the subject matter expert can use that information to create a rule. That’s one way to start creating value from detecting anomalies.
If the use case is right, you might be able to start increasing revenue, or save costs, or manage labor. For example, if a correlation is shown between the anomalies you're detecting, you may be able to predict when the machine requires maintenance, allowing you to schedule that maintenance.
It sounds like while some data analytics solutions can help identify problems with existing equipment, it can also be used for new machines to get a clearer picture of normal and abnormal operation.
Yes, exactly. Also, with some of the manufacturing data sets that I've looked at, there can be so many different sensors and data points being collected, and you can't have someone just sitting there monitoring all these dozens of sensors for every single piece of equipment. Anomaly detection is a nice way to tie things together so you're monitoring less and gaining more insight from the data.
For more information, visit the Bsquare website.