Senior software engineer Kundan Karma helped build the IoT Connector and machine learning model Microsoft is using to improve the operational efficiency of its energy smart buildings. (Photo by Kundan Karma)<\/figcaption><\/figure>\nBuilt on Microsoft Azure, the IoT Connector immediately removed manual steps, reducing errors, and improving communication. Creating a ticket became a one-click process, with greater accuracy and faster processing time for technicians.<\/p>\n
\u201cIn the IoT Connector, we take care of all the data,\u201d Karma says. \u201cIt\u2019s a bridge between two systems.\u201d<\/p>\n
Designed with auto-healing and telemetry fail-safes, the IoT Connector gives RE&F confidence that faults will be captured and reported as tickets with greater accuracy.<\/p>\n
\u201cIf messages between the two systems fail, the IoT Connector will resubmit,\u201d Karma says. \u201cAfter a certain number of retries or if there\u2019s a major problem, it will create a ticket for an engineer to look at.\u201d<\/p>\n
Improved communication introduced a handful of ancillary benefits\u2014specifically, visibility.<\/p>\n
Where a technician might previously circumvent inputting information into a work order, automated copying facilitated by the IoT Connector made tickets in Facility Link a single click away.<\/p>\n
\u201cIn cases where someone just does the fix without a work order, we don\u2019t know what\u2019s been done,\u201d Obermayer says. \u201cThis left us with an incomplete history. We couldn\u2019t see the demand for certain things.\u201d<\/p>\n
Now capable of tracking work orders, RE&F has a better understanding of what\u2019s going on within specific buildings and assets. These insights are improving decision-making, especially as it relates to energy efficiency.<\/p>\n
A firehose of IoT data<\/strong><\/p>\nThe IoT Connector shines a light on some challenges that come with scaling energy smart buildings.<\/p>\n
\u201cThe target was 100 buildings,\u201d Karma says. \u201cWe were so focused on integrating Iconics with Facility Link that we didn\u2019t consider the volume of data. When we first rolled out the IoT Connector, we had to stop at 13 buildings. One building was generating approximately 2,000 faults per day.\u201d<\/p>\n
Extrapolated across Puget Sound\u2019s 100 buildings, that amounted to roughly 200,000 faults in a single day. The scale of data being generated by IoT sensors could overload Microsoft\u2019s entire Dynamics 365 system, bringing things to a standstill.<\/p>\n
\u201cThe issue was conversions,\u201d Gaurav says. \u201cOnly meaningful faults require an actionable response. We only want to check on real issues.\u201d<\/p>\n
Getting useful information out of IoT sensors is a challenge.<\/p>\n
\u201cThere are different tolerances and different polling schedules for different pieces of equipment,\u201d Obermayer says. \u201cIt changes from building to building.\u201d<\/p>\n
Microsoft Digital needed to separate the wheat from the chaff.<\/p>\n
\u201cIf you have data generated in the thousands, it\u2019s easy to miss important alerts,\u201d Gaurav says.<\/p>\n
Reducing the number of faults meant rethinking the way alerts from energy smart buildings were generated.<\/p>\n
\u201cWhat we realized is that 75 percent of the total faults were coming from one source, terminal units, and most of them were never converted to any work orders,\u201d Gaurav says. \u201cIt was taking up most of the UI and creating too much noise. The way this data is now processed has adjusted how we\u2019re digesting and prioritizing alerts.\u201d<\/p>\n
Terminal units, for example, were reordered and reprioritized to reduce the amount of noise being generated.<\/p>\n
\u201cWe tried to group faults together,\u201d Gaurav says. \u201cOne fault can trigger other alerts, but you don\u2019t need multiple work orders.\u201d<\/p>\n
We want the model to mimic the behavior of a technician. It can go through the same decisions a human being can and reach the same conclusion.<\/p>\n
\u2013 Kundan Karma, senior software engineer, Microsoft Digital<\/p>\n<\/blockquote>\n
Instead of treating all alerts as individual issues, alerts could be grouped so several related faults resulted in a single ticket.<\/p>\n
\u201cWould a technician investigate that?\u201d Karma says. \u201cWe want the model to mimic the behavior of a technician. It can go through the same decisions a human being can and reach the same conclusion.\u201d<\/p>\n
Teaching a machine to think like a technician<\/strong><\/p>\nTo get things started, Microsoft Digital looked at the history of faults and determined how they were converted to work orders.<\/p>\n
Brendan Bryant, a mechanical engineer with DB Engineering, one of Microsoft\u2019s partners, helped translate the technician\u2019s process to the team. These inputs allowed the Microsoft Digital team to build a machine learning model that could mimic the behavior of a technician.<\/p>\n
\u201cWe had key performance metrics from six to eight months\u2019 worth of IoT Connector data,\u201d Bryant says. \u201cI helped Kundan look at HVAC telemetry and all the IoT metrics to get his team the information they needed to train the algorithm the right way.\u201d<\/p>\n
But before they could get there, naming conventions for assets and structures had to be standardized.<\/p>\n
\u201cThis is one of the reasons we put in our own system,\u201d Obermayer says. \u201cHow things would work was that a vendor would decide on an asset name when the building was constructed, then we\u2019d change vendors or use a different vendor for a different building.\u201d<\/p>\n
The result was a variety of similar, yet varied, naming conventions. Facility Link meant RE&F could standardize and align all data points for energy smart buildings across campus.<\/p>\n
\u201cWe can now look at a data point and tell you the number of air valves in Puget Sound,\u201d Obermayer says. \u201cData and problem types are now the same on every system, making energy smart buildings more precise and efficient.\u201d<\/p>\n
Alignment of nomenclature also meant Bryant could better convey priority issues.<\/p>\n
\u201cThere\u2019s a lot of engineering intuition involved, especially when checking what\u2019s false and what\u2019s true,\u201d Bryant says. \u201cIt\u2019s a large amount of data provided by all of the equipment, so you have to make a judgement based on what you\u2019re seeing.\u201d<\/p>\n
To help train the model to identify real issues over false alarms, Bryant and Karma moved away from real-time response and started viewing faults in aggregate.<\/p>\n
\u201cSomething might show up on a Tuesday and be gone by Wednesday,\u201d Bryant says. \u201cThere\u2019s no value in creating a work order for that. But if it\u2019s an issue for most of a week, that\u2019s something we want to flag.\u201d<\/p>\n
Once aggregated, certain key performance metrics became strong predictors of a fault.<\/p>\n
\u201cIn order to maintain high confidence that a fault needs to be addressed, we need a longer period of data,\u201d Bryant says.<\/p>\n
As the team continued their efforts, items that would result in a work order were flagged while all others were archived. From this, the model began to predict the faults that would result in work orders, flagging them for attention and archiving the rest.<\/p>\n
\u201cThe technician can view anything flagged as \u2018false\u2019 and review it,\u201d Karma says. \u201cIf needed, the technician can pull the fault from the archive and review it on the fly. The model learns from the mistake when it\u2019s time to retrain.\u201d<\/p>\n
Thanks to machine learning and new practices, the number of faults was reduced by 80 percent to 90 percent.<\/p>\n
\u201cWhen we were onboarding, we couldn\u2019t do all of Puget Sound\u2019s smart buildings because the number of faults was huge,\u201d Gaurav says. \u201cOnce we were confident that the faults generated were manageable and convertible to work orders, we were able to quickly onboard the rest of campus.\u201d<\/p>\n
Predicting the future for smart buildings<\/strong><\/p>\nWith the IoT Connector, Microsoft\u2019s technicians are more efficient, disparate systems are better integrated, and modern infrastructure is in place to further sustain energy smart buildings.<\/p>\n
\u201cRight now, we\u2019re only looking at HVAC, but there are so many other IoT assets throughout Microsoft,\u201d Karma says. \u201cA\/V, security cameras\u2014you name it. The next phase is to integrate all of these items into the IoT Connector.\u201d<\/p>\n
Flexibility within the IoT Connector allows it to be utilized with any asset across any region in the world.<\/p>\n
\u201cIt becomes a scalable implementation,\u201d Gaurav says. \u201cWe can even use it in areas that will eventually become energy smart buildings to help support those efforts.\u201d<\/p>\n
Karma also sees the IoT Connector, which is built on Microsoft Dynamics 365, as being available to other companies looking to improve the efficiencies of energy smart buildings.<\/p>\n
\u201cWhat we\u2019re planning is to create the IoT Connector in a generic way so that other people can benefit from it outside of Microsoft,\u201d Karma says. \u201cAny other team should be able to use our learnings.\u201d<\/p>\n
The standardization of assets in Facility Link has helped spur other RE&F initiatives.<\/p>\n
\u201cHaving this data is super important,\u201d Obermayer says. \u201cThis will impact everything from procurement decisions to the management of movable assets.\u201d<\/p>\n
As Karma continues to refine the model, retraining hones prediction accuracy.<\/p>\n
With each iteration, the model gets stronger.<\/p>\n
\u201cThe big thing looking forward is helping to teach the algorithm so that we understand when it makes a decision and why,\u201d Karma says. \u201cEventually the model will be able to assign work orders automatically.\u201d<\/p>\n
Gaurav agrees.<\/p>\n
\u201cThe model is robust and converts some fixed number of alerts to tickets automatically. However, we also allow technicians to review through the list of alerts and allow them to manually create tickets as and when needed,\u201d Gaurav says.<\/p>\n
For Obermayer, all of this is a dramatic improvement.<\/p>\n
\u201cWe started with thousands of faults but could only address about one percent of the issues,\u201d Obermayer says. \u201cWe got the number of faults down so that we\u2019re actioning 10 to 20 percent, which means we\u2019re hitting meaningful faults. Artificial intelligence and machine learning are improving the business of energy smart buildings.\u201d<\/p>\n
<\/p>\n
\nDiscover how Microsoft\u2019s smart buildings showcase Azure Digital Twins<\/a>. <\/strong><\/li>\nLearn about Microsoft\u2019s new era of smart building in Singapore<\/a>. <\/strong><\/li>\nFind out how Microsoft promotes environmental sustainability from the inside out<\/a>.<\/strong><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"A new machine learning model and Internet of Things (IoT) sensors and automation enables Microsoft smart buildings to keep company employees as comfortable as possible.\u00a0Microsoft\u2019s real estate operations team relies on energy smart buildings, structures with interconnected automation and sensors, to responsibly maintain a base level of comfort. Microsoft has deployed more than 50,000 sensors […]<\/p>\n","protected":false},"author":80,"featured_media":6381,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"jetpack_post_was_ever_published":false,"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_memberships_contains_paid_content":false,"_hide_featured_on_single":false,"_show_featured_caption_on_single":true,"footnotes":"","jetpack_publicize_message":"","jetpack_publicize_feature_enabled":true,"jetpack_social_post_already_shared":true,"jetpack_social_options":{"image_generator_settings":{"template":"highway","enabled":false},"version":2}},"categories":[1],"tags":[330,76,127],"coauthors":[442],"class_list":["post-6378","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-uncategorized","tag-azure-ai-and-machine-learning","tag-azure","tag-dynamics-365","program-ms-digital-stories","m-blog-post"],"jetpack_publicize_connections":[],"yoast_head":"\n
Microsoft smart buildings bolstered by machine learning model, IoT - Inside Track Blog<\/title>\n \n \n \n \n \n \n \n \n \n \n \n \n\t \n\t \n\t \n \n \n \n\t \n\t \n\t \n