As the industrial sector embraces the Smart Factory concept, executives are increasingly recognizing that applying Machine Learning to reduce unplanned machine downtime in industrial plants can directly improve operating margins and company profitability.
Unscheduled downtime occurs when machinery in an industrial plant stops working in the middle of production without any advanced notice. The result is that production is slowed or stopped until repairs can be made. Although the average industrial plant loses 17 a year of production due to unplanned downtime, until recently, this was not considered an executive level priority. According to one study, the process industry loses $20 billion a year or 5% of production due to unplanned downtime.
The Smart Factory concept is based on the principle of applying Big Data and Machine Learning to both production and factory operations and maintenance.
Within industrial plants, machinery is embedded with sensors that monitors the performance of machines and controls the process. Factory technicians monitor these sensors – including temperature, vibration, pressure and others – because significant changes in sensor readings are often indicative of possible machine failure. The problem with this approach is that by the time technicians are aware of changes in sensor behavior, in many cases, the machinery has already broken down.
This is where Machine Learning emerges as catalyst for change. Within the exabytes of sensor data generated by industrial machines are micro-patterns that can tell us when a machine is likely to fail.
Until recently, there were no tools to systematically access the insights that are found within the sensor data. With advances in Machine Learning, patterns of anomalous machine behavior can be detected across multiple sensors and machines. The correlations of anomalous sensor behavior can be used to identify the root cause of the machine failure. When repair technicians are provided with this detailed information, they can order parts and fix the problem before the downtime occurs.
At an executive level, there is a lot of support for the idea of applying Machine Learning to the data generated by industrial plants. However, the problem has been a lack of skilled data scientists to extract the insights hidden with in the data.
In research conducted by Presenso and Emory University, skill shortages of data scientists was considered the most significant factor preventing the adoption of Machine Learning for Predictive Asset Maintenance.
Underlying this concern is that in order to implement Machine Learning solutions at a plant level, data scientists will need to be hired. Today, there is a significant shortage of data scientists in the market, and most are attracted to higher paying industries such as high tech and financial services. A second issue is that given the vast amounts of Big Data that is continuously generated, it is difficult to scale Machine Learning solutions across industrial plants with a large number of assets.
An advance in data sciences known as Automated Machine Learning or AutoML is being applied to address these issues. What is not understood about the data science discipline is the extent to which much of the work done by data scientists is manually performed. These include Machine Learning model development, selecting the correct model to use for a given scenario and re-calibrating this model.
In a sense, Automated Machine Learning applies Machine Learning to the data science field. Tasks that are repetitive and labor intensive are now performed automatically. Instead of the data scientists selecting which Machine Learning algorithm to use for a given data set, an algorithm is trained to make this selection.
The impact of applying Automated Machine Learing is wide-ranging.
From an operational perspective, because repair work crews are provided with information about evolving failure in advance, it gives them extra time to order parts and schedule the repair without disrupting production.
Another outcome is that Original Equipment Manufacturers (OEM) that produce and sell industrial machinery are taking a larger role in the maintenance of the machinery that they sell. Whereas in the past, OEMs sold machinery to industrial plants that became responsbile for maintenance, this model is now changing. Companies such as Siemens realize that because the Big Data generated from sensors can be monitored remotely, there is now an opportunity for the OEM to take responsibility for asset maintenance.
A new sales model known as Hardware as a Service (HaaS) is now emerging whereby OEMs lease industrial machinery to plants and provide ongoing maintenance service and support.
Not long ago, data science was primarily a theoretical study limited to academia. However, in a short period of time significant progress has been made to operationalize the insights generated by Big Data. The impact is not only financial or production output related. Over time, it is leading to a fundamental change in how industrial plants operate and maintain production machinery.