Striking a balance between supervised & unsupervised machine learning

How do the two models of machine learning work in industry, and what are their limitations?
12 July 2019

Supervised or unsupervised? Source: Shutterstock

Since the first use of advanced software in asset-intensive industries more than four decades ago, manufacturers have been on a journey to transform their businesses and create added value for stakeholders.

Today, a fresh generation of technologies, fuelled by advances in artificial intelligence based on machine learning, is opening new opportunities to reassess the upper bounds of operational excellence across these sectors. 

To stay one step ahead of the pack, businesses not only need to understand machine learning complexities but be prepared to act on it and take advantage.

After all, the latest machine learning solutions can determine weeks in advance if and when assets are likely to degrade or fail, distinguishing between normal and abnormal equipment and process behavior by recognizing complex data patterns and uncovering the precise signatures of degradation and failure. 

They can then alert operators and even prescribe solutions to avoid the impending failure, or at least mitigate the consequences.

Unsupervised machine learning

The leading software constructs are autonomous and self-learning. They demonstrate a capability known as unsupervised machine learning, a specific method of learning patterns of performance or behavior using clustering techniques.

Such a technique can be used to understand ‘normal’ operational behavior, based on signals from sensors on and around machines. 

Once the behavioral patterns are learned, analysis of new data can help detect deviations from the norm, called anomalies, highlighting mechanical issues and process changes that affect specific pieces of equipment.

However, the downside is that anomaly detection based on unsupervised learning may be fraught with errors and always requires human intervention. It is good at detecting correlations but less effective at working out causation. 

Indeed, unaided machine learning may find correlations that can be complete nonsense, such as the meaningless but true correlation between reduced highway deaths in the US and the number of tons of lemons the country imports from Mexico.

Unsupervised machine learning can find correlations, but without human guidance, they can be meaningless.

Unsupervised machine learning can find correlations, but without human guidance, they can be meaningless. Source: Shutterstock

When unsupervised machine learning detects an anomaly, the change in behavior patterns could be just a new operating mode, or it could be an impending failure. 

A human must take a look at the machine and decide which of the options is correct. Such manual intervention can then help machine learning learn and adapt, effectively ensuring that moving forwards it always provides analysis the business can trust.

After all, correlation is not the same as causation, so to learn properly, machine learning needs human guidance. For example, voice recognition technologies use machine learning, but cannot learn without help.

The technology assessments need to be highly-stewarded by humans, who intercept unresolved phrases and apply translations to assist learning techniques. Similarly, credit card fraud detection needs help to learn to recognize spending behavior.

The credit card company might ask “Are you attempting to purchase an air ticket in Paris?” The credit card issuer uses machine learning to understand your normal spending patterns and now recognizes an abnormal event.

A simple input of ‘yes’ or ‘no’ characterizes the detected anomaly and ensures that the technology learns to recognize any future spending as normal or fraud.

Supervised machine learning

Supervised machine learning similarly needs human involvement to work effectively.

It requires an individual to declare an event and the time and date it occurred. Then, the technology must learn the signature of the precise patterns that lead to that event, which in the asset-intensive industries could, for example, be a machine failure due to an exact cause such as bearing failure.

The technology learns and calibrates the precise degradation and failure pattern and then tests new incoming data streams to find exact pattern recurrences, to then alert well before the failure occurs, allowing an action to avoid the failure or provide time to arrange a timely repair before major damage occurs. 

The results are much lower maintenance costs and more uptime producing valuable products.

Bright prospects

Owner-operators across asset-intensive industries urgently need to start taking advantage of the many benefits that machine learning can already bring in terms of running their facilities more efficiently and optimizing asset performance. 

In today’s crowded marketplaces, there is a window of opportunity to use machine learning to predict asset performance; drive business advantage and develop a competitive edge. 

Unsupervised machine learning can take businesses part of the way by helping to identify anomalies in asset or plant performance, but organizations must be aware that to perform at optimum levels, machine learning needs human guidance and intelligence. 

It still needs guide-rails to find and solve the right problems.

Asset-intensive businesses need to act on this understanding now. Those that do will be best placed to take advantage of the new era of machine learning and ensure that by measuring actual patterns of asset behavior and extracting real insight and automatically developing foresight, they can optimize asset performance across their entire operations. 

This article was contributed by Mike Brooks, Senior Director at AspenTech.