Nearly two and half years ago, I conducted some research on and wrote an article about an IIoT (Industrial IoT)-enabled horse halter in early-stage development, called NIGHTWATCH. While my family is heavily involved in horses, I had no idea that over 100,000 people would be interested in and read the article. Today, Austin-based Protequus, a biomedical engineering, and data-science firm, in collaboration with NRGXP, an engineering and IoT consulting firm, announced the U.S. and Canadian launch and shipment of the NIGHTWATCH smart halter.
When I first wrote about NIGHTWATCH, I did not have full access to all the technical details as it was in early development, but now I do, and I am blown away by the level of sophistication and amount of bleeding-edge technology organized into a very small and low-profile form factor.
Here are details of what I find the most interesting technical aspects of NIGHTWATCH, which includes its ability to do “edge computing”, how it leverages machine learning, the novel application of Ultra-Wideband Impulse Radar (UWB-IR), and overall wireless implementation. Let me first review why NIGHTWATCH exists in the first place.
NIGHTWATCH monitors heart and breathing rate, activity, motion, location and posture.
NIGHTWATCH monitors heart and breathing rate, activity, motion, location, and posture (NIGHTWATCH)
Colic is the primary cause of death in horses, and according to NIGHTWATCH, the “AAEP (American Association of Equine Practitioners) estimates that more than 900,000 horses in the U.S. will experience an episode of colic each year.” That is an astounding number, but if found early, colic is easily treatable in the majority of horses and more importantly, is survivable. The challenge with horses (unlike our cats and dogs) is they spend much of their day unsupervised in a barn, large pasture, etc. so when horses colic, there may not be anyone around to see them, or if a person is around, they may not know how to spot it.
The NIGHTWATCH halter (and optional safety collar) was designed to bridge this gap by monitoring real-time biometric & behavior data and alerting up to 5 people via phone call and text at the earliest signs of a problem so they can intervene sooner. Because this technology is “smart” it learns each horse’s unique physiology and looks for deviation that may be indicative of equine distress, such as colic, being cast, and foaling (giving birth).
The edge compute choice
Over the last two years, the amount of “edge computing” technology industry dialogue and discussion has taken off. That discussion was driven in the IoT context and the realization that not everything that computes makes sense to be processed entirely in the cloud. While no rational person thinks all compute should be in the cloud versus local, it’s really the percentage of what gets done on the edge, gateway, and in the cloud. For certain applications where responsiveness, latency, security, network costs, and even fault tolerance is important, using edge compute makes more sense. This was true for the NIGHTWATCH smart halter, as the company says it needs to process over 300MB of data every day for every device.
Therefore, NIGHTWATCH needed ways to remotely monitor a horse with and without a persistent wireless connection. It’s one thing for an Amazon.com Echo to not be able to connect to WiFi or Amazon.com, but it’s another thing entirely when it comes to a loved and expensive horse. NIGHTWATCH says that without on-device computation, the device would instead have to upload all the data, compute it in the cloud, and then notify owners or caregivers, which could introduce more latency and be very costly to energy consumption. Additionally, there is a visual distress indicator in the form of LED that can report horse distress without any connection. Therefore, a person can walk-by and quickly evaluate a horse’s general well-being via the LED without any connection.
Some data does need to be uploaded to improve machine learning training models and for archive in Amazon.com AWS S3, and the halter supports WiFi and cellular communications for that. It is likely that a lot of that data could be transmitted via an expensive, cellular connection, so pre-processing the data makes sense here, too.
- Novelda Xethru UWB radar– Operating within 3 GHz to 10 GHz, this transmits the raw UWB-IR RF signal that is then processed on the device to calculate the horse’s heart and respiratory rates. The sensor can measure heart and respiratory rates by detecting small displacements in the microvascular system and physical changes in the soft tissue behind the horse’s ears. The raw radar data comes into the MCU at about 7KB per second.
- Telit GNSS GPS module– This sensor is used to determine the horse’s location and to calculate distance moved.
- Invensense 9-axis accelerometer, compass, and gyroscope– This measures the horse’s movement and comes into the MCU at about 1 KB per second.
- TE Connectivity barometric pressure sensor– This sensor, in addition to the 9-axis accelerometer, is used to better predict posture.
So overall, with around 8KB of data being collected every second per horse, generating 300MB every 12 hours, the halter needed on-device processing.
Here are some of the notable processing silicon:
- NXP Semiconductors M0 microcontroller (MCU)– Used to accurately sample and pre-process the sensor data outlined above. It runs FreeRTOS and Arm CMSIS software to connect the MCU to the sensors.
- NXP Semiconductors I.MX6 processor (CPU)– Processes real-time fuzzy logic machine learning inference algorithms that computes the EDI (Equine Distress Index) based on the pre-processed data from the MCU. The CPU runs a custom Linux OS and software for linear algebra algorithms, frequency analysis, equation evaluation, object-relational mapping, and fuzzy logic.
- Micron Technology 4GB eMMC– Stores motion data before AWS S3 archival.
- Texas Instruments and NXP Semiconductors PMICs– Used to manage and optimize the system and power use to maximize battery life.
- Maxim Integrated Products battery gauge– This gives the user an estimate of current battery capacity.
Over time, NIGHTWATCH says it will be optimizing its algorithms to reduce power draw and hence improve battery life beyond the stated overnight usage of 12-16 hours. Longer-term, the company can see moving to an FPGA or ASIC once the algorithms have been fully optimized. Does this sound familiar to what many other leading companies are thinking related to edge computing? Absolutely.
Algorithms and machine learning
NIGHTWATCH says that it uses MathWorks MATLAB to prototype machine learning training and on-device algorithms (hand-converted/optimized to C) to predict equine distress.
Training models are created on an on-premise cluster. The company says it syncs data from AWS S3 to an in-house cluster storage, trains the models, and then pushes them up to S3 for the halters to download. NIGHTWATCH says during the training period of approximately 2 weeks, models can require over 16GB RAM of data. NIGHTWATCH has not pushed model learning to AWS as the company prefers a complete understanding of the memory, CPU, and GPU (NVIDIA CUDA libraries near future) requirements before selecting the appropriate cloud infrastructure compute needs and cost). NIGHTWATCH says it’s open to doing cloud training on Amazon.com AWS, Google GCP, or Microsoft Azure in the future as the need arises.
NIGHTWATCH can be managed on smartphone, tablet, and PC (NIGHTWATCH)
NIGHTWATCH chose MATLAB, Gaussian Mixture Models (GMM), and Expectation Maximization (EMM) algorithms for some very pragmatic purposes—when it began in 2014, frameworks like Tensorflow, Caffe, and Torch were not even available. That, plus NIGHTWATCH believes that many of them use languages and multiple 3rd-party libraries that are not yet suitable for low-level embedded processors with resource constraints.
Very hairy wireless implementation
Just because the NIGHTWATCH halter is an exceptional edge-computing IIoT example, it doesn’t mean it is immune from needing a leading-edge wireless solution. What’s even more difficult is being battery-powered and organizing all this in a small form factor like a horse halter.
- Novelda Xethru UWB-IR and Custom Antennas– Operating from 3,000 to 10,000 MHz, this is used to detect the horse’s heart and respiratory rates.
- Telit Communications GPS module and Antenova antenna- Operating at 1,559-1,609MHz, the GPS antenna is used to determine the halter location and better determine motion.
- Texas Instruments WiFi chip and TDK Corp antenna- Operating at 5,000 MHz and 2,400 MHz, enables halter connectivity to a WiFi router for alerts and data upload.
- Telit Communications PLC cellular modem and Antenova antenna-Operating at 824 to 960MHz and 1,710 to 2,170 MHz, enables halter connectivity to a 3G carrier network for alerts and data upload.
One cannot have giant antennas sticking up like a WiFi router does and therefore, antenna design is tricky. Once NIGHTWATCH got all the signals working well in a cramped space, half the work was over. Now it needed to pass FCC, IC, and PTCRB regulatory requirements. If you cannot meet these requirements and get those certifications, you are not operating in the U.S. or Canada nor are you connecting to a carrier’s network.
Catherine Moorhead with Casino Royal wearing NIGHTWATCH smart halter.
While certification for WiFi, cellular, and Bluetooth are very well known and common, almost no one knows much about UWB-IR. According to NIGHTWATCH engineers, most testing laboratories did not even have the right equipment to test, did not understand the use case or the testing requirements, and even new ground had to be broken surrounding permissible exemptions for SAR (specific absorption rate) testing. Phones are one thing, but IoT-enabled devices without voice transmission requirements for horses with UWB-IR mixed in is another thing entirely. The NIGHTWATCH smart halter received FCC, IC and PTCRB certification and passed IEC-60950 safety testing in the Fall of 2017 and the company says it will be available to horses in the U.S. and Canada starting April 26, 2018.
Edge compute, but the cloud still helps
Even though the NIGHTWATCH halter is a great edge computing example, it still uses the public cloud for various tasks. The company uses Amazon.com AWS S3 for:
- User login and security
- Web UI presentation and back-end database
- Alerts for owners or caregivers
- Storage and archival of raw biometric and behavior data, logs, and machine learning models
- Distributing new inference machine learning algorithms to the halter
As I highlighted above in the machine learning section, the company currently uses an on-premise, private cloud cluster for algorithm creation.
NIGHTWATCH is the most sophisticated animal IoT device I have ever researched and one of the most sophisticated IoT device I have ever seen. Hopefully, you get a better idea of why I think the NIGHTWATCH smart halter is a good example of modern IIoT edge computing. For use cases like this where latency, network resilience, and cellular network costs are very important, it makes better sense to do the compute closer to where the data is collected.
While I may be enamored with the compute, machine learning, wireless technology, and issued patents, Protequus and NRGXP are most interested in saving animals’ lives, which, in the end, is the better reason. The company says online sales start May 1st, 2018. You can find more about NIGHTWATCH here.