Machine learning is a fundamental element of the Industrial Internet of Things; results from it can drive significant operational efficiencies in industrial companies such as identifying false positives for equipment errors so that a repair truck doesn’t have to be dispatched unnecessarily, or predicting likely equipment failures so that repairs can be made proactively when the equipment is being serviced for other reasons. However, machine learning is not magic, nor is there an omniscient system that can predict the future for any piece of industrial equipment on day one. Machine learning is a process and one that must be followed with some rigor in order to deliver meaningful results.
The first step in machine learning is to train the system; the machine actually needs to learn first. It needs to learn about the piece of equipment being analyzed, the equipment’s characteristics under normal conditions, what conditions are considered abnormal, and the equipment’s characteristics under abnormal conditions. The training process consists of analyzing an initial data set, interpreting the results, and then improving the data set based on the results – typically increasing the volume of data as well as collecting additional data elements from the equipment.
Rarely is the initial data set sufficient to provide all the desired outcomes with a high level of confidence because it’s difficult to determine how much data will be sufficient and which data points will be relevant for a given business problem up front. But over time, the iterative process of analysis, interpretation, and data augmentation will lead to an acceptable machine learning outcome.
Once training is complete using a representative data set, the system can begin to make real-time predictions based on incoming telemetry data from the equipment. But in order to validate the accuracy of the predictions, and to improve them over time, a feedback loop needs to be established which tests and validates the prediction. This feedback loop typically involves integration with operational business systems so that equipment service details – i.e. what repair work was performed – can be captured. While this type of back-end integration may not be top of mind for companies thinking about machine learning, it’s an essential component to generating predictions with a high level of confidence.
Just because machine learning requires a series of methodical, and often technical, steps doesn’t mean that it is out of reach for any industrial company. With the right guidance and support from subject matter experts and data scientists, companies can go from sifting through piles of random data, to effective predictive analytics based on machine learning. If the right science is done behind the scenes, the end results will seem magical.