While there is ever-increasing use of industrial IoT and we are moving closer to a world where machines will be able to take care of themselves, those machines will not be fully robotised. This is according machine learning expert Mike Brooks of Aspen Technology. His company looks at large machinery used in manufacturing such as compressors and assesses the health of those compressors by taking readings from their sensors that are streamed out.
The senior director and asset performance manager at Aspen Technology explained the things he looks for in those readings. He said: “Some of these compressors have 160 signals that get streamed out to us. We combine that with the event data we get from maintenance systems to understand what are the relationships that make them work and what are the patterns that lead to failure.”
The advantage of machine learning, said Brooks, is that is can look across those 160 dimensions, whereas a human can usually only identify issues in the sensor data in retrospect if something happens.
"Patterns give us information that humans cannot see."
“There are a lot of places where the patterns that lead to failure are not evident to humans.
We found that patterns give us information that humans cannot see,” he added. Brooks likened this activity to the way that financial services companies would carry out fraud detection by looking for anomalous behaviour.
If some errant behaviour is found, Aspen Technology will request that the equipment owner take a look and confirm that either the machine is indeed breaking or that there is a change in the process. If the latter is the case Aspen Technology will need to modify its knowledge of the machine’s behaviour for the future.
His company can assess, interpret and read those patterns from signals coming from a variety of mechanical equipment including locomotives, mobile trucks as well as compressors. There are set parameters of what normal should look like and any deviations from the norm will lead Aspen Technology to inform the owner so they have time to avoid the negative consequences of mechanical failure of at least plan to mitigate those consequences.
“Understand that machine learning is not a silver bullet."
With this in mind, Brooks said machine learning is not a fix-all technology and humans need to be on hand to guide it. “Understand that machine learning is not a silver bullet. It needs guide rails to make sure it is working to solve the right problem,” he said.
The people tasked with programming the machine learning software should be very experienced in the behaviour of the equipment and Brooks also said that person, who might be called ‘a reliability engineer,’ should be in charge of the ongoing condition of the equipment.
However, it seems that there are still many decades to go before the machines can do it all by themselves.
“Fully robotising this equipment? I believe we are a long way from that, certainly in our industry,” said Brooks. “I do think if there is a problem occurring and we can recognise that there is a failure pattern coming, we can offer prescriptive advice to the operator. That’s heading towards what I call self-healing systems.”