This isn’t your father’s predictive maintenance. When predictive maintenance was first introduced, it was famously the low-hanging fruit of advanced manufacturing. Predictive maintenance was typically the first step in the move to smart manufacturing.
Technology has matured considerably. Now the goal is to gain optimal life for the machines while avoiding failure. If deployed properly, predictive maintenance predicts future machine failures, pinpoints complex equipment problems, and identifies parts that need fixing. With those principles in place, plant operators become more attuned to planning maintenance. They can reduce unplanned machine downtime while maximizing overall equipment life. This is a bigger bite than simply identifying a motor that’s running too hot or sounds funny.
According to MathWorks – an AI software company – these desired outcomes won’t happen without an algorithm capable of predicting a time window when machines will likely fail. The following four-step workflow is designed to guide manufacturers through the process of developing effective predictive maintenance:
- Assemble raw data that describes systems across several healthy and faulty conditions
- Preprocess it so condition indicators – features that distinguish healthy conditions from faulty ones – can be extracted
- Use these features to train a machine learning model to detect equipment anomalies, classify fault types, and estimate the machine remaining useful life
- Deploy the algorithm and integrate it into systems for machine monitoring and maintenance.
We caught up with Philipp Wallner, industry manager at MathWorks, to expand on these four points:
Design News: What is involved in assembling raw data that describes systems across several healthy and faulty conditions? What type of sensors are used for this?
Philipp Wallner: It doesn’t matter if it comes from a PLC or intelligence sensors. Or if it’s live or recorded. If we look at the project, it’s a combination of different sensors. A lot of different sensor data, pressure data from extruders, and temperature data. The first approach was to get this raw data and put it in the system to identify the feature. Looking at the key sensor values makes a difference in identifying whether the equipment will fail or not. Specific sensors can make a big difference. Oil extraction requires vibration data. When the client doesn’t have the right data, we often recommend an additional sensor.
DN: Preprocessing so condition indicators can be extracted. Are some of the data coming from the equipment manufacturers?
Philipp Wallner: Most of the time it’s from past use, but more and more the equipment manufacturers provide data. With automation components and drives, we do see some of the bigger vendors providing some information about the health indicators of the equipment. Some provide models, but this is in the early stage.
DN: Do you have to have machine failure to provide data on potential failure?
Philipp Wallner: The equipment does not necessarily have to fail. We do have data from equipment that reveals whether the data in front of you is off or not. That helps you to predict whether the equipment will fail. Another way is looking at component data from machines that have failed in the field. The other aspect is to watch as the condition indicators move away from a healthy area, and then define when it has moved out of the healthy area. This is the heart of predictive maintenance. This is where domain experts come into play. They can pretty well decide the window of operation you need to stay. Then you send alerts for the time when the indicators move out of this area.
DN: Using these features to train a machine learning model to detect equipment anomalies, classify fault types, and estimate the machine remaining useful life. How is the estimated remaining useful life determined?
Philipp Wallner: In general, one of the key goals is to get an understanding of how long the equipment will last before it fails. There are a couple of ways to indicate the remaining useful life. One approach is to set specific thresholds. If the hydraulic pump can’t exceed a certain temperature, that’s a good aspect to bring into the remaining life. You have models that predict how long the equipment will perform before it reaches this threshold. You can calculate that. If you have data, you can predict the degradation of the particular piece of equipment.
Our role is to provide the toolset. We have a team that helps customers ramp up, but then it is our customers who collect and analyze the data. More and more wheat farm vendors set up sensors to collect data. Then they use the data from multiple locations around the world to predict failures.
These users are sharing data with the manufacturers. We’re also seeing that in wind farms. Machine builders in Europe have service contracts with customers to share data with the manufacturer in exchange for additional service.
DN: As for deploying the algorithm and integrating it into systems for machine monitoring and maintenance, how is it integrated into systems? What systems?
Philipp Wallner: We are pretty open in terms of integrating algorithms. From a technical prospect, there are a lot of different ways. There are three main platforms. One involves controls such as PLC, and industrial PLC. Most often it’s edge devices running Linux or Windows. The third is the cloud. We see all three options. The fourth is splitting the data up. It doesn’t make much sense to send raw data to the cloud. What makes the most sense is to do processing at the equipment like the PLC, then the data is pre-processed at the edge and you send the post-processed data to the cloud with the sophisticated algorithm.