Distributed Intelligence in the IoT

The Internet of Things has intelligence distributed throughout many integrated layers. Algorithms and analytics are needed to make sense of the raw data. Often this intelligence needs to reside on the edge of the network close to the source of the data–near the ‘things’ comprising the IoT.

The Internet of Things inherently calls for an extremely distributed architecture. By its nature, the real-time “thing-originating data” (i.e. sensory, ID, time, and location data that emanates from the intelligent things themselves) originates at the edges of the network — out where the ‘things’ are.¹ In many cases, a lot of processing is done on IoT data very near to the source. Hence the intelligence about the things within the IoT is very distributed.

Real-time IoT data comes from many sources including sensors, locating devices, and identification devices. The raw data generated at the lowest level in these devices generally goes through many layers and levels of filtering, summarizing, pattern detection, interpreting, and processing. This is required for many reasons. First is just the sheer volume of data. A sensor may emit data hundreds or thousands of times per second. Few if any central enterprise applications want to receive those volumes of raw data. Additionally, the data often is much easier to comprehend and use when it is interpreted into a meaningful business event by a local processor (more on this below). Also, often there is a feedback loop controlling the ‘thing’ (device or machine) based on this data as well. There is local intelligence, embedded in devices, readers, and local appliances, that makes sense of the low level data and provides intelligence to higher levels.

Figure 1: Example Multi-Layer Distributed IoT Architecture

Sensor data may be read dozens or hundreds of times per second by a local processor making various decisions based on that data. That local processor may in turn send much less frequent data about state changes or significant events up to a local appliance or SCADA device. This may in turn communicate via a local network to some plant-wide server that finally connects up to an enterprise system. The actual connection to the internet may be far removed from the end-point sensors and processors. Furthermore, IoT devices may be only intermittently connected to the network. A vehicle with many onboard sensors and processors may not connect to the network until reaching certain way points or at the end of the journey.

GIS Data (e.g. a map of field)
Guides Precision Agriculture
Source: John Nowatzki, North Dakota State University

Precision Agriculture (An Autonomous Intelligent Machinery Example)

Consider the case of precision agricultural technology that enables farming machines to perform variable-rate precision planting, precision application of fertilizers and pesticides, robotic weeding, precise parallel swathing (harvesting of grains), and yield monitoring. Agricultural equipment outfitted with precision technology very often uses GPS, with some versions having accuracy down to a couple of centimeters (about an inch) to automatically steer the machine, as well as to decide when and where to apply precise amounts of seeds, fertilizer, and pesticides.

The GPS data is combined with GIS data, as well as data from onboard sensors² (measuring things such as plant biomass, plant chlorophyll content, soil nitrogen content,etc.) enabling the machine to make constant real-time decisions and adjustments. For instance, when applying pesticides, the system may use data from sensors, as well as the GIS data, to continually adjust the boom height, nozzle flow, and droplet size.

The planting, fertilizing, or harvesting machine downloads GIS data which will guide its path and boundaries of the activity. The GIS data/map may also specify areas to skip or that should receive different types or level of treatment. Data from the sensors is used in real-time to guide the equipment, but a summarized version of this data is also recorded and goes back to the central farm record. External intelligence and data, such as weather forecasts and crop prices can be also introduced to help the systems optimize profitability and yields. Thus we see the multiple bi-directionally communicating layers of intelligence of IoT at play in the modern farm setting.

Figure 2: Distributed Intelligence — IoT in a Precision Agricultural Setting

Converting Raw IoT Data into Intelligence

I recently heard about the BrainPort, a device that helps blind people ‘see.’ It takes a low resolution video feed from a small camera mounted on a set of sunglasses and converts it into electrical impulses on a small device that sits on the wearers tongue. When first used, the sensation is perceived as just a meaningless set of tingling sensations, somewhat like the sensation of bubbles on the tongue while drinking a soda. But then over time, with practice, a rough picture emerges — the wearer can begin to recognize patterns and understand shapes and what they mean. Blind people have been able to learn all manner of useful tasks like how to locate doorways, find buttons on an elevator, recognize and pick up different silverware, and even read large letters and numbers. It demonstrates the power of the human brain at synthesizing and making sense of a confusing array of data to recognize patterns.

Similarly the raw data produced by IoT sensors is generally overwhelming, and by itself is meaningless without some level of pattern recognition to derive intelligence from the data. The pattern recognition and other processing (such as filtering of redundant data or monitoring thresholds) should be done where it is most appropriate and efficient, which in many cases means putting a processor very close to the source of the data. Thus, like our own nervous system, the IoT has intelligence infused throughout it, at all the various nodes, levels, and layers.

Deducing Intelligence from Raw IoT Data

Over a decade ago, some retailers started trying to track RFID cases in their stores by placing readers at three locations within the store: A) on the receiving dock doors, B) the doors between back and front of the store, and C) at the box crusher. They discovered that many things can interfere with RFID reads and, because read rates are never 100%, it can create some strange looking data. Normally you would expect to have four reads in the lifecycle of a case passing through the store – 1) received at dock door, 2) brought from back to front of store to replenish shelf stock, 3) empty case brought from front to back of store, 4) empty case put into box crusher. If the system saw the case received, brought to the front, and then put in the crusher (1, 2, and 4), but never read the empty case coming back from the front (#3), it can deduce that was simply a missed read. There are numerous other more complex patterns from which the system can make various deductions about what actually happened.

It becomes even more complex when the retailer is tracking individual items continuously using fixed ceiling readers with real-time locating capabilities. There are lots of sequences that seem nonsensical, such as an item that suddenly disappears or jumps instantaneously from one location to another, when in fact the tag was just temporarily obscured. Over time, these algorithms will be improved to provide high confidence and nuanced information about what is actually happening on the store floor.

Developing Distributed Intelligence in the IoT

In some cases, the data coming from sensors and decisions about what the system should do about it are pretty straightforward. A very simple example is an old-fashioned thermostat that has a set point and a hysteresis range³ to automatically turn the heat or AC on and off. However, in many cases, IoT devices require more complex algorithms to make sense of the data. One example is passive RFID being used to track cases or individual items in a retail store. Over time, with experimentation and feedback from operation in the field, engineers develop these algorithms to understand what the low level patterns of data actually mean. These ever-evolving algorithms are improved over time and can be used to provide increasingly more reliable, meaningful, and nuanced intelligence about what is actually happening in the store. In fact, the intellectual property behind these types of algorithms (across all sorts of domains, not just retail or RFID) will become some of the most valuable and differentiated aspects of IoT offerings.

Intelligence at All Levels

To be clear, we are not saying that intelligence, pattern matching, and analytics only exist down at a local level. There are plenty of opportunities at all the various layers to create the appropriate intelligence. For example, a real-time situational awareness application may take in local information from IoT devices, environmental information from local weather feeds, situational information from social media, enterprise data about shipments or facilities, and put it all together to provide intelligence at many levels.

An autonomous self-driving mining truck is continuously navigating its environment while also monitoring its own health. It might send a signal up to a central service organization about preventative maintenance needed, based on a combination of usage and early-warning measurements (e.g. excessive vibrations or temperatures). Furthermore, an aggregated and organized version of all that usage, performance, and parts failure data might be shared with the mining vehicle design engineers who are designing the next generation of truck, so that they can do a better job at usability, reliability, efficiency, and safety.

In short, the Internet of Things has intelligence inherently distributed throughout, in many different integrated layers. The development of algorithms and analytics to make sense of it all is a big part of building out the IP and value of IoT. The ability to aggregate this IP and give system builders quick access to these various layers, tools, and algorithms is becoming a key enabler of accelerated development of IoT applications and extraction of greater value from these devices and systems.

For further exploration, see IoT Distributed Intelligence Examples

________________________________________________________

¹ Note, there are many other data used by IoT applications that do not originate from the edge of the network. Examples include GIS maps and enterprise data (e.g. order or shipment information, CAD data, part numbers, employee information, etc.) and other semi-static data, which may be pulled from various databases around the network. — Return to article text above
² Leading agricultural sensor manufacturers and devices include Ag Leader/OptRx, Topcon/CropSpec, Trimble/GreenSeeker, and Holland Scientific/Crop Circle. — Return to article text above
³ The hysteresis range is generally not visible to or adjustable by the end user, but rather is designed by the engineer. — Return to article text above

To view other articles from this issue of the brief, click here.