The Temboo team was inspired to tackle the challenge of building system performance, in part, by our own office space. Similar to the typical New York City property, our office building was 118 years old with aging systems that were expensive to replace. From meeting with our building supervisor and other local property managers, we determined the main challenges of retrofitting old building systems with monitoring equipment were:

  • Lack of technical expertise to install sensor networks and internet gateways
  • Lack of interoperability between sensor and hardware vendors
  • Prohibitive cost
  • Making meaningful conclusions based on the data

We set out to give facility operators an affordable building system performance platform that could be self-installed and provide meaningful insights.

What it does

Kosmos enables property managers and facility operators to gain insight into the condition, performance, and overall operation of their building systems. First, property managers complete a wizard-like setup flow where they identify what sensors and internet gateways they will use. Based on those inputs, our software machine generates code to connect the sensors and internet gateways to our servers. After users upload the code onto the sensors and gateways, they will see data on our online dashboards, available from any computer or mobile device. The dashboards provide a number of features:

  • Live graphs
  • Detailed activity logs
  • Remote over-the-air updates
  • The ability to set rules for email & SMS alerts
  • The ability to remotely control actuator

Predictions and advanced analytics, powered by Machine Learning, also enable property managers to detect anomalous events, and forecast maintenance requirements.

How we built it

To build Kosmos, we expanded upon an existing product that enabled sensors connected to a microcontroller to send data to our servers. The challenge with our existing architecture is that it required the device to be configured with Temboo libraries. By using a gateway architecture, not only can we perform data analysis locally, but devices do not need any Temboo code running locally; the internet connected gateway handles directing device messages to our backend servers. The backbone of our server infrastructure is a message handler that forwards device messages to all relevant recipients like a webui for display, a database for storage, TensorFlow for machine learning, and an email server for alerts. We selected TensorFlow as the framework for the machine learning-based prediction tools. It had the benefit of an active community for developer support, allowed for customization, and abstracted different neural network classes in an easy to use manner. TensorFlow Serving API also allowed us to quickly configure and launch our own serving node, giving us control over how we serve our models in production.

To make Kosmos compatible with all industrial and off-the-shelf sensors, we had to build a common messaging framework. The advantage of building our own framework was that we treat all messages, regardless of originating protocol, as objects with a common set of functions. We currently support four protocols: Bluetooth 4.2 LE, HTTP, Modbus TCP and MQTT; any new protocol can be added, provided there is an appropriate business case.

Challenges we ran into

Some of the key challenges our team faced in building Kosmos were related to hardware and building machine learning algorithms. Enabling Kosmos to work with industrial and off-the-shelf sensors from any vendor required months of testing. We also encountered a number of issues when trying to build with TensorFlow since it is a young library. For example, it took a significant amount of trial and research to ensure that TensorFlow met our security expectations with data transmission.

Accomplishments that we're proud of

We are particularly proud of leaping from providing users access to the TensorFlow API to making it an integrated component of our end-to-end system. Once we overcome the challenges of working with TensorFlow, we were able to project a graph into the future, detect anomalies, and predict necessary maintenance. Not only does the software make predictions based on sensor data, but it can also incorporate external data i.e. from a weather forecast, to predict future system states, such as pipes freezing.

What we learned

In speaking to property managers about the challenges they faced, our team learned just how much creativity was required to work with old buildings. For example, to build an internet-connected system in a former army terminal, we had to learn how to work with a networking technology from the 1920’s.

We also learned a lot about building machine-learning powered models. Guaranteeing idempotent behavior allowed for redundancy in downsampler execution, preventing against unexpected failures. Also, pre-downsampling improved the software performance and allowed representative data to be quickly visualized in our dashboards.

What's next for Kosmos Industrial IoT Platform for Smart Buildings

By the year’s end, we will have launched multiple pilot projects in preparation for our public launch in early 2019. For example, we are in final talks with a major New York City developer to use Kosmos to monitor a building complex’s freight elevators and water pumps. We will also continue to refine existing and add support for new features including:

  • A redesigned of our application onboarding process
  • Batched sensor data uploads
  • Over-the-air updates updates for the gateway application
  • Additional support for Zigbee, Bluetooth 5, Bluetooth Mesh and Sub 1GHz-enabled sensors
Share this project: