Question Of The Week: How Buildings Learn

Behind their walls buildings have long supported an interplay of systems that belies their static presence. Now, the ability for buildings to learn and proactively impact operations expands the horizon of possibilities.

By David Karpook

When the legendary modernist architect Le Corbusier wrote in 1923 that “a house is a machine for living,” he was referring metaphorically to the utilitarian purposes that a home serves. Nearly 100 years later, buildings are becoming more and more like machines — and, increasingly, machines that learn.

Buildings have long supported a rich, constantly churning interplay of systems that belies their static presence. Behind the walls, above the ceilings, under the floors, is a continual flow of air, water, electricity. Other building installations — for fire safety, emergency lighting, etc. — stand poised to act as soon as there is a distress signal. These dynamic systems are the life blood, the energy sources, and the nerve centers of buildings.machine learning

What now is evolving rapidly is the ability of buildings to sense their occupants: Our comings and goings, our activities, our needs and desires. Further, we can see that through aggregation and analysis of the “sensed” data, the technologies are moving beyond sensing to understanding. These capabilities are becoming highly developed enough that the IT analyst group Gartner Inc. has predicted a shift “from technology-literate people to people-literate technology.”

How does this happen? How do machines, and by extension buildings, develop an understanding of the people who use them? How do they learn?

The short answer is that machine learning mimics human learning. It accumulates data, recognizes patterns in that data, and then uses those patterns to make predictions. A key difference is that what humans do as a core part of our being, machines must be programmed to do. That programming takes the form of complex mathematical formulas known as algorithms. Algorithms allow machines to recognize patterns, form clusters of similar data, and extrapolate from these to reach conclusions or make predictions.

For example, facial recognition by machines and buildings uses a set of algorithms to compare measurable characteristics such as the shape of the head, distance between eyes, length of nose, and fullness of lips. By accumulating enough discrete measurements, it can determine the identity of a person whose image is picked up by a camera lens.

For a piece of building equipment, measurable data might include operating temperature, vibration, fuel consumption, gas and fluid pressures, even noise level. An algorithm might analyze these measurements to make recommendations about the need for maintenance. As with the facial recognition example, actions that would otherwise be done by humans — taking readings, comparing measurements, and correlating measurements to other data such as service calls from occupants — are done instead via automation.

If you then think about other kinds of sensors that are appearing more frequently in buildings, such as those that register people entering rooms, sitting down at workstations, opening doors, turning on lights, adjusting thermostats, or pushing a button at a restroom exit to comment on its cleanliness, it may be easier to see how buildings could become “people-literate technology.”

Machine Learning: Supervised And Unsupervised

Although all machine learning requires use of algorithms, there are two distinct branches of machine learning: Supervised and unsupervised.

Supervised learning uses what are called labeled responses. This means that for a set of inputs, such as the equipment metrics listed above, an algorithm could yield a specified result, such as predicting operating costs for a specified period of time or, better, predicting when a replacement should be budgeted based on rising operating costs. This could be done by having the algorithm incorporate known factors, such as the cost of a belt that typically needs to be replaced every 100,000-runtime hours. Humans could certainly do the same analysis, but having the machine analyze its own performance saves a lot of spreadsheet crunching and yields the result many times more quickly and consistently (machines aren’t prey to “human error”).

Unsupervised machine learning proceeds in a different way. In unsupervised learning, the machine looks at raw data — including unstructured data such as photographs, videos, and sound recordings — and spots patterns that may be difficult or extremely time-consuming for humans to discover. Clusters, layers, colors, movement, and anomalies are among the discoveries that can be made about a data set. Once patterns have been discovered, further analysis can determine their significance. This may require human intervention, but the discovery of the pattern itself provides value. And in many cases, there are algorithms waiting to be developed to make greater sense of those patterns.

Traffic patterns in a building might be discerned through unsupervised machine learning based on sensor or security camera data. Whether and how these correlate to utility consumption, cleaning requirements, or revenue generation could be a secondary or tertiary exploration that could be human, algorithmic, or, most likely, a combination.

The ability of our buildings and equipment to collect data has increased exponentially as the Internet of Things has made inroads into every aspect of our existence. Our smartphones track and report on our every step and heartbeat. Amazon Alexa, the voice recognition system, is moving from tabletops into walls — including the walls of the workplace. You may see all of this monitoring as good or bad, but it indisputably generates data to be aggregated, segmented, interpreted, and analysed. And while the volumes may be overwhelming to humans, they are simply processing time to machines.

Now, computers are being taught to write their own code to solve their own problems. That suggests that human intervention will become less important to many facility management tasks as buildings take advantage of artificial intelligence.

Ultimately, the goal of machine learning in buildings is to move from guesswork and approximations into quantitatively supported decisions that improve efficiencies, control costs, and spur performance. As we move forward, humans, machines, and buildings will likely learn from each other and find new ways to optimize their shared intelligence and support the occupants and the work that inhabit them.

Karpook is Strategic Business Consultant for Planon Corporation, a provider of technology solutions for facility management and real estate. A 25-year industry veteran, he has been a customer, vendor, system implementer, trainer, and strategist, managing workplace technology projects around the world. Karpook is chairman of OSCRE International, the Open Standards Consortium for Real Estate. He is also past chairman of IFMA’s Real Estate Advisory & Leadership community, has been a member of the IFMA Information Technology community, and has worked with the IFMA Foundation’s Global Workforce Initiative. Additional experience includes seven years as a facility manager and construction project manager at the University of Florida.

Do you think machine learning holds promise for your facilities? If so, how? What are your questions on machine learning as it relates to the FM and the real estate industry? Please share your experiences and other comments below.