Big Opportunities in TinyML : Leveraging AI at the Edge in Supply Chains

by Kumar Singh, Research Director, Automation & Analytics, SAPinsider

What is TinyML ?

If you have not heard of the term “TinyML” before, you are not alone. The good news is that you are probably already aware of the underlying technology behind this terminology but were not aware of this marketing term that was assigned to it, sometime in late 2019. So let us first start by understanding what TinyML is all about, in the simplest way possible. There is however an assumption in the subsequent sections that you are aware of concepts like Internet of Things (IoT), edge devices and edge computing. I believe some of the most innovative applications of the power of TinyML are in the areas of operations and marketing and hence I will use examples from operations and supply chain throughout this article.

Let us start by exploring an example of how a “vanilla” IoT setup works today, in a manufacturing context. A key aspect that we need to keep in mind is that sensors capturing data on manufacturing floors is not new technology. The core setup has existed for decades. The supporting technologies have definitely evolved and sensors are becoming increasingly sophisticated and cheap but the gist is that sensors have been used to collect data on manufacturing floors much before even the term “Industry 4.0” came on the horizon. But in a vanilla setup, all that these sensors do is capture a set of data/data point and relay that data, in near real time, to a central server. So it is not difficult to infer that the true value of this type of setup is not in the hardware (sensor) functionality, but in how you leverage the data you are collecting. While that is not within the scope of this article, data captured from the floor can not only be used for analytics but this near time data relay is fundamental to creating a digital twin.

Now let us visit another example from manufacturing. A camera deployed in a production line is scanning finished parts coming out from the line. There is an “inference” program installed on this camera that can flag any parts that it believes do not align with the quality requirements/specifications. Notice the difference between this example and the previous one. In the previous example, all that the edege device (sensor) was doing was capturing a data point and relaying it back. It was not applying any form of intelligence to the data point it was capturing. In the second example, our edge device (camera) has a Machine Learning (ML) algorithm installed on it, which allows it to apply an inference algorithm on the captured data. This second example, is the example of TinyML. The letters “ML” in TinyML obviously stands for Machine Learning but why the word “Tiny” ?

As you may know, true Machine Learning algorithms are computionally intense. The training and deployment mostly happens on high powered machines. Let us revisit our manufacturing sensor example. Say we want to leverage the data being collected to implement predictive maintenance capabilities. The data collected will be leveraged to build a ML model, that will be trained on a machine with high computation specs, and will run on a proper system. The ML algorithm, in this case, sits on a device with powerful computing capabilities. Now if we come back to our camera example, it is not difficult to imagine (where we are currently in terms of technology),that as compared to a full fledged system, a camera on the edge is a low power device, and hence houses a ML algorithm that is not computationally intense, but is impactful.

So, with that background, we can formalize a simple definition of TinyML:

TinyML broadly encompasses the field of machine learning technologies capable of performing analytics on data captured by extremely low power edge devices, on those devices themselves.

Between advances in hardware & computing power, and the actively engaged open source tools based TinyML community (like TensorFlow Lite)  that has generated exponential innovation in machine learning on the edge, it is now possible to deploy increasingly complex algorithms,  directly on edge devices. The possibilities will only expand from this point so now is the right time to start exploring the opportunities to leverage TinyML capabilities in your operations infrastructure in detail.

TinyML can revolutionize supply chain visibility beyond manufacturing

TinyML has made good in-roads in manufacturing. While in one of the examples that I used in the previous section the sensor was only capturing and relaying the data, sensors today are very much capable of  applying predictive maintenance algorithms on the edge. You can magnetically attach sensors these days to the outside of the turbine and it will analyze detailed data at the edge. These sensors can alert regarding potential issues even before they occur, by analyzing the data they capture. The key areas that still need to be explored deeper are leveraging TinyML for logistics and inventory management. Let us visit few areas where TinyML, if designed and leveraged prudently, can create innovative capabilities.

Warehousing : Opportunities to leverage TinyML in warehousing range from monitoring dock bottlenecks to identifying safety issues in the warehouse. The best part is that you can build these capabilities incrementally and then integrate them in a platform, to create a “Smart Warehouse” platform, powered by TinyML. Many aspects that are currently housed across multiple systems, can be centralized in this type of platform, and combined with TinyML powered edge devices, can help control warehouse applications much more efficiently. While relatively simple analytics happens on the edge devices leveraging TinyML, the central algorithm takes care of the co-ordination and centralized planning, plus strategic planning.

Inventory Management : While smart edge devices that monitor and flag storage conditions are already being used on shop floors of many warehouses, opportunities to expand usage of TinyML are plenty. When you think about possible applications of  any technology, a good starting point is what are some of the challenges those on the shop floors run into. As an example, no matter how sophisticated a warehouse technology is, misplacement of inventory within the four walls of a warehouse is a frequent phenomenon. TinyML powered edge devices can alert putaway personnel in real time if a product is being placed in the wrong aisle or within a wrong bay in the aisle. This is the simplest of applications and it is not difficult. to imagine that solving more complex problems using TinyML will yield more significant benefits, with some having the capability to transform inventory management in warehouses.

Transportation : Transportation has been an early adopter as far as digital tracking goes. Telematic tracking devices have been in use for a long time in transportation fleets and these have now been augmented with more advanced sensors and trackers. But like inventory management, there are many unexplored opportunities. We know that pattern recognition and incident detection are among the capabilities ML algorithms can help build and these can be leveraged in TinyML powered edge devices, powered and supported by a central algorithm sitting on a central server. As an example, TinyML powered edge devices, in tandem with a central algorithm, can help study traffic patterns to optimize and plan routes based on peak traffic hours, maintenance, and construction, thereby helping develop dynamic routing capabilities that do not exist today (or maybe exist in marketing white papers only).

What does this mean for SAPinsiders ?

  • TinyML is not a new concept, the marketing term associated with the technology is relatively new. SAP ecosystem has partners that have helped clients leverage SAP technology portfolio and open source tools like TensorFlow Lite to build TinyML based solutions. Opportunites to leverage this galore in both operations and marketing.
  • There is a fundamental foundation that needs to be built before embarking on building this type of advanced analytics solution. Setting up that foundation in terms of people, processes and technology is critical. Think about it this way: You can’t make a leap from having BI capabilities to building a solution like this. While you may end up implementing the technology, your people and processes will not be able to support the solution and the entire initiative may turn into a failure.
  • Plan an incremental roadmap, based on where you are in terms of your current state, in functions mentioned above. While external partners can definitely provide support, your people are the ones who are closest to the process. Start leveraging their expertise much earlier in the planning process.
  • Stay tuned for our “Modernizing Inventory and Logistics Tracking” research report coming out in July 2021. That will provide you useful insights on the impact IoT is making on the world of digital tracking in supply chain operations. The report also explores other digital tracking approaches that SAPinsiders are exploring, like RFID technology.

 


 

Kumar Singh, Research Director, Automation & Analytics, SAPinsider, can be reached at kumar.singh@wispubs.com