Big Data. That definitely was the tipping point. Over many years, starting in the 1970s, concepts around replication of human interactions through computers and the terms “neural network” and “Machine Learning (ML)” were found mostly in academia. In the 2010s, a new science blending statistics and ML was launched. We call it “data science.” The focus in all of these areas was and continued to be on managing massive amounts of data and pattern recognition of images.
For many of us, the ability to interact in an almost conversational tone with devices like Amazon’s Alexa and even our cars, phones, and other devices seems normal and not the science-fiction type leap that we saw in Star Trek (1965-1969) and 2001: A Space Odyssey (1968). Even today’s computer-generated voices have an almost human quality, unlike the robotic voices that 1990s-era computer systems had. That makes them seem as if they are coming from truly intelligent devices. Television, movies, marketing, and other sources, including our personal interactions with such devices, seem to have programmed us into believing these systems are actual Artificial Intelligence (AI).
As noted in “Life, Digital Transformation, And Everything” | THE RAM REVIEW, business schools and consultants discuss “digital transformation” and how ML and AI are implicit in their thinking. In short, that means these systems are able to make human-like decisions on their own. This capability, referred to as “generalization,” takes into account an assumption that no programming is involved. However, data scientists and developers know that ML and AI thinking and, by extension, their programming are explicit, which mean they provide specific answers based on input, and have limited generalization. In fact, every ML/AI system, by definition, incorporate programmer’s bias. The better systems simply have better programming and training that use experts in specific fields for input and interactions to improve data based upon correct responses and false positives and negatives.
When we take the first position that narrow ML/AI of a specific task or tasks is implicit, we can end up with some very serious conditions. After all, to err is human. Really messing things up requires AI. However, if we take the second position that AI/ML is explicit, then we have a powerful set of tools. Basically, the more we understand about the something works, the better our decisions based on the information provided. In fact, there are some fantastic capabilities available in properly applied expert AI and ML systems that have been developed with significant programmer bias.
The use of data science, AI, and ML for reliability engineering and physical-asset management is powerful when application of those technologies is understood. What is data science? It’s the manipulation of information from chaos to ordered to identify patterns and link what might otherwise be considered unrelated. As an example, consider, the correlation between temperature and failure modes. Then add in the correlation between vibration and bearing life .When we study information from multiple sources, those correlations and patterns can have meaning, but we have to be able to visualize the data. In comes ML to the rescue.
Machine Learning is an important item in a data scientist’s toolbox. It’s used to manage large amounts of data and pull out patterns, and then allows the pattern-recognition process to be repeated. The front end of the ML development pipeline involves human interaction with data and data selection for supervised systems, and studies information clusters for unsupervised systems.
While some people think that ML requires repeated processes and data, the opposite is true. For simple systems, we can use spreadsheets to look at trends and relationships, or quickly program an ML algorithm for analysis and pattern recognition. It’s the more complex systems where data science and ML come into play, and the reason why we use those types of processes with massive sets of seemingly unrelated data to pick out the patterns that have meaning. At the same time, new patterns can be identified, even with smaller datasets, and result in previously unrealized opportunities.
Among other things, variable equipment use based on a review of operating conditions and stressors is specifically suited for data science and the application of ML and processed AI. With AI being the trained process from data science and ML the ability to forecast equipment maintenance and failure points, the leap in reliability engineering can be managed.
THE ROAD AHEAD
What haven’t we yet done with data science and ML in the reliability and physical-asset management space? With varying degrees of success, we already have applied some level of ML/AI in routine predictive maintenance (PdM) tasks that use IIoT devices for equipment monitoring and testing. However, what many do not understand about ML/AI is the ability it has to “fill in the blanks” for us. Basically, the combined technology can be used to identify where incorrect information has been entered into our various maintenance-management systems, account for those bad data points, and modify the conclusions in a way that we could not otherwise.
Yes, you read that right: Part of the power of modern data science/ML/AI in statistical analysis involving surveys is the ability to identify bias and correct for it. The same capability can be used in reliability engineering to account for incorrect data entry, failed/failing sensors, and other conditions that normally cause us to discard data. (This will be a topic for a future article.)
Most people equate data science and ML/AI to massive amounts of data rather than to the pattern analysis part that turns reliability engineers into data scientists by default. Consider all those formulae that a CMRP needs to understand: The manipulation of the data associated with them is data science by default. The automated tools for performing that part of our work are applicable in other arenas, including healthcare (reliability engineering for humans), finance (reliability engineering of money), insurance (reliability engineering of risk), and any other industry that calculates risk and probabilities.
Even calculating something as simple as MTBF (Mean Time Between Failure) is considered data science. Any system that modifies data resulting in changes from input is, by modern definition, ML; any system that acts on such information through software is AI. For example, if you have something that provides data on MTBF and it adjusts based upon data input (manual or otherwise), it is ML. If an action results through software, it is AI.
in fact, when we perform a root-cause-failure analysis (RCFA) or anything else that involves a “fault tree” or “decision tree,” we are using a tree/forest regression or neural networks. A forest regressor (a machine learning method), is literally a fault tree with weighted or unweighted yes/no answers. When we are applying a neural network, we are making decisions based on every node of weighted probabilities with the possibility of multiple correct answers or relationships.
Think in terms of a 5-Why as a forest regression and a neural network as an advanced RCFA. A shallow neural network has only a few layers of decision-making while a deep learning neural network can have many. All of the different ML/AI systems are based on what amounts to truth tables, or decision trees. That’s it. This concept is termed as “classification” of a system in which you can identify a specific fault or response.
What’s the difference between an AI system and an expert system? Not much if the expert system in question looks at trending for Time to Failure Estimation (TTFE), which is effectively the same as Remaining Useful Life (RUL). As with Reliability-Centered Maintenance (RCM), we need to identify what “normal” is and what “functional failure” looks like (including their threshold). This is performed in ML two ways:
1. Supervised: data is developed and fed into the system. Setpoints are developed and programmed.
2. Unsupervised: data is observed and the software develops patterns identifying what is ‘normal’ for
the system as it stands when the pattern is developed.
In both cases, a system would be looking for incidents when a pattern shifts from “normal” toward a variation from the expected, or crosses a pre-selected value.
The real challenge is understanding that more than half of IIoT technologies used in PdM are traditional alert systems that flag events and send resulting data for review by humans. The second part of the challenge is that a majority of the remaining systems are developed by organizations that have no expertise beyond what appears in papers and textbooks, which results in significant system bias.
The unfortunately small percentage of systems developed by experienced personnel who have replicated their internal processes into an ML-capable AI are generally quite accurate, with a high degree of confidence. Most are similar to those used in healthcare. Because of recognized issues with IioT, especially in infrastructure systems, standards and legislative bodies are beginning to actively define data science, Machine Learning, and Artificial (or Augmented) Intelligence.TRR
ABOUT THE AUTHOR
Howard Penrose, Ph.D., CMRP, is Founder and President of Motor Doc LLC, Lombard, IL and, among other things, a Past Chair of the Society for Maintenance and Reliability Professionals, Atlanta (smrp.org). Email him at firstname.lastname@example.org, or email@example.com, and/or visit motordoc.com.
Tags: reliability, availability, maintenance, RAM, asset management, IIOTm Machine Learning, ML, Artificial Intelligence, AI