In simple terms, we expect the following objectives to be met using IoT to different degrees.
Increase Throughput: “Do more”
Continuous Improvement: “Higher quality”
Mass Customization: “Better UX”
Volume, velocity and variety of input data are so great that a human learning what the data is saying, especially in real-time or near real-time, is inconceivable! Therefore, one of the key enablers of IoT has been Machine Learning (ML).
Back in 2000, when we developed and deployed Rockwell Automation’s Integrated Condition Monitoring Solution in large factories, the focus was on automating actions in the periphery: automatic collection via attached sensors, edge-processing (called “fog” computing these days) to make data transport manageable and data transport itself. There was of course, automated processing in the server (“cloud”) but the algorithms applied were simple automated deviation detection, ranking and methods for remote root-cause analysis plus manual follow-up when essential.
By 2016, improvements in networking, data transport, sensors and automated Machine Learning have proliferated industrial IoT applications and driven them into social use cases. So far, so good.
If this was the “Crawl” stage of the “Crawl-Walk-Run” evolution of IoT, what will drive the “Walk” stage from a technical perspective?
CONTINUOUS Learning! Till now, ML has gotten by with “one-shot” learning . . . and accomplished some impressive things. But the larger value-proposition of IoT is Continuous Improvement (Objective #2 above). Continuous improvement cannot happen without continuous learning! Continuous Learning happens when we view learning as “generalization from past experience AND results of new action”! Current ML algorithms cannot accommodate “results of new action”; in other words, current ML uses STATIC machine learning.
Dynamical ML is machine learning that can adapt to variations over time; it requires “real-time recursive” learning algorithms and time-varying data models such as the ones in Generalized Dynamical Machine Learning. Hence, continuous learning can be implemented using DYNAMICAL machine learning. Dynamical ML adds the following FOUR benefits:
Learning in real-time.
Just like children mature over time . . .
As machines age, IoT adapts to NEWnormal.
IoT designed for long-term use.
Underlying system “states” provide a meta-model.
A more stable description; less False Positives.
System “states” as “Digital Twin”!
Continuous “closed-loop” performance improvement.
In real-life applications, be it industrial or social, the value of the first 3 benefits are self-evident. “States as Digital Twin” requires a few comments. Wikipedia defines Digital Twin as – “Digital twins refer to computerized companions of physical assets that can be used for various purposes. “”Digital twins”” use data from sensors installed on physical objects to represent their near real-time status, working condition or position.” GE, Siemens and other industrial giants have been championing Digital Twins for some time now.
So, Digital Twin is a non-physical representation of a physical asset, or an “avatar”. “Twin” emphasizes the closeness to reality; however, there are many non-physical representations. On the one extreme, one can develop detailed equations (a large set of nonlinear coupled partial differential equations?!) that capture the total “physics” of the machine; then they can be rendered nicely on a (3D?) display using CGI to look almost real. I am not an advocate of such verisimilitude – there is little business benefit.
In the process of DYNAMICAL machine learning, the data model and the algorithms used (Generalized Dynamical Machine Learning) naturally generates what is called the “State-space” model of the machine. It may not *look* like the machine but it captures the dynamics in all its detail (there IS a challenge in relating “states” to actual machine component conditions though). So, I am a proponent of using the “State-space representation” that we get for free in Dynamical ML as the “digital twin”. This is a topic for future exploration and advancement.
We are ready for Dynamical ML today. The following publications provide the theory, algorithms, applications and MATLAB code.