By Dr. Shi-Wan Lin, IoT Technologist and Co-Chair of the Industrial Internet Consortium Technology Working Group
This is Part 3 of a 3-part blog series. Read Part 1 and Part 2 here.
As we work to realize the benefits of Smart Maintenance and Global Optimization in operations, it it tempting to build systems with a centralized architecture in which data from assets are collected to and analyzed in a centralized system, likely in a public or private cloud, where operational decisions are made based on the results of the analytics. This approach is likely a based model for many initial IIoT deployments, because of the overall simplicity of design and the efficiency offered by a large sharable pool of generic computational resources. This model would work well for use cases in which the operation of the assets involves few and infrequent decision-making from the centralized system. On the other hand, a centralized architecture may not work well in other cases. There could be technical constraints such as the latency involved in the decision-making loop and an undesirable dependency of the asset operations on the quality and availability of the network. There could also be economical constraints such as the cost of uploading a large volume of data over wireless networks.
One less obvious concern in a centralized architecture is the ever-increasing complexity of the system. The problems we seek to solve in Industrial Internet are complex problems that are only getting more complex. As we drive toward a higher level of global optimization, more systems are connected together and the resulting complexity in the system of systems only goes up. The relationship and dependence among the systems increasingly become untenably entangled. As such, centralized decision-making in such an environment would become ever more challenging. Any solutions we come up with would risk becoming less stable. It might bring about too many bottlenecks, hotspots, weak-links and worse, even single-points of failure, in the overall system thus making it harder to achieve the level of resilience that is critical to industrial operations.
To address the issue of growing complexity in the overall system, it is advisable to distribute complexity and decision-making across the system to achieve Local Autonomy. This calls for solving the problems closest to its sources and to the context in which the solution is most conducive. This also calls for distributed decision-making to the entities where it is most appropriate and effective (e.g. situational awareness is most achievable). This finally calls for autonomy in the assets and collaboration among them in proximity. In this regard, we have much to learn from the apparent global intelligence and resilience that emerge from swarm-like collaboration in schools of fish, flocks of birds, colonies of ants or bees in which the autonomous constituents execute only simple rules with peers in proximity.
Along this line of thought, as we pursue the next level of effectiveness and efficiency in operations, we may want to reexamine our current automated systems. There may be opportunities to increase the level of flexibility and adaptability of these systems to deal with conditions unforeseen when the automation systems were created. Additionally, there are many systems that are largely automated but still involve a human-in-the-loop in the flow of operations. The presence of a human-in-the-loop could be due to situations in which cognitive capability may be required to solve specific complex problems. As we become more apt in applying cognitive capabilities in these systems, we may want to reduce the reliance on humans in operations in order to increase reliability, efficiency and safety. Think about the examples ranging from robotic aids to human workers in warehouses, underground autonomous mining machines, and the well-publicized autonomous vehicles. In these autonomous systems, we elevate the human role from the operational level to the mission control level - setting objectives and handling exceptions. These systems would have an increasingly higher level of capability to learn and adapt - capable of extrapolating parameters possibly outside the range of the original test set, trying out solutions according to its risk evaluation and confidence level. In the process, they would learn and expand the range of the test set for safe operations, autonomously.
How can we go about building these collaborative systems with distributed autonomy (decision-making capabilities)? Solving complex problems requires a correct framing of the problems and the right approaches to the solutions. In order to achieve distributed problem-solving and decision-making, in other words, we need the right world modeling, analytics, computation platforms and system architectures distributed across the assets. We need well-distributed and well-interconnected computation capabilities, lots of them, closest to the assets.
Fortunately, the continuing advances in computation and communication that have ushered us into the era of Internet of Things are also enabling us to achieve distributed and collaborative autonomous problem-solving. The greater embeddable computational capabilities, increasingly packaged in miniaturizing sizes, consuming less energy and available at lower cost, are enabling the execution of - closer to the assets - more advanced analytics and better modeling of the world. This will help to transition the assets from merely automation to autonomy. At the same time, the ubiquitous connectivity among the assets and from them to the broader systems is making it possible for the autonomous assets to collaborate with each other seamlessly within proximity and to be coordinated across the network by entities with a purview of a larger scope - the idea of Global Optimization.
With such advances, we can foresee that the patterns of distributed problem solving will be dynamic in response to the changes in the nature and complexity of the problems as they develop. The systems will become increasingly capable of dynamically distributing computation for analytics and problem solving close to where the problems are - in a way similar to the Big Data architecture principle of moving the computation to where the data are, however, in a much larger scale with much greater complexity. In these systems, newly developed and verified algorithms, models and approaches for problem solving are shared with and adapted by the individual autonomous assets (not through configuration as in the centralized paradigm prevalent in today's systems).
As we advance in this direction, we can recognize the emerging role that software plays at every level of a system, in both enabling autonomous operations in the assets as well as collaboration among themselves and coordination by broader systems. In the near future, we can envision a world in which the fabric of the physical world including the industrial assets is bound together and being constantly sensed and optimally operated by software driven by advanced analytics at various levels - creating a "Software-Defined World".
A key task of achieving collaborative autonomy in the overall system is to recognize the pattern of complexity of the problem, on which the scheme of distributing the complexity across the system is based. Another task is to determine the nature and level of autonomy to be embedded in the assets or other sub-systems and the types and levels of collaboration needed among them. This is to be followed by the task of identifying the algorithms, analytics, models and frameworks required to support the autonomous decision-making at the asset level in addition to the types of sensing required to gather the necessary data for the computation.
As we are at the very first stage of this revolution, the biggest challenge we face, I would argue, is that we do not yet know enough of how to implement collaborative autonomous systems that are not only functional in operation but are also secure, safe and resilient. We expect a hastening pace in advanced research and development in both technologies and frameworks in comprehensive physical modeling, Machine Leaning, Cogitative Computing and Artificial Intelligence to meet this great challenge. Technology advances in autonomous vehicles and other robotic systems have taken great strikes in this direction and what we learn in these areas will be of great value as we move forward in other industrial areas even though most of the industrial assets have a fixed location. Another key challenge is the long lifecycle of the industrial assets that last for decades during which the cyber (computational) portion of the assets would likely evolve several generations while its physical counterpart may remain relatively stable. Therefore, as we build new smart assets (cyber-physical systems - CPS) or retrofit existing brown-field assets, we need to be keenly aware this fact and strike to design CPSs with cyber components (computational hardware and software) not only seamlessly integrated with their physical counterparts but at the same time ‘pluggable,’ allowing them to be upgraded over time at a pace different from that of their physical counterparts.
The idea of distributing complexity and problem solving (involving decision-making) is one of the many topics that are getting strong interest and attention within the IIC Technology Working Group. For example, the topic of distributed analytics is under active discussion by the Industrial Analytics Task Group. The topic of how to advance from integrability, to interoperability and finally to composability is being explored. Dynamic Composition and Automated Interoperability, a chapter in the recently published IIRA, outlines the concept of an agent-based design allowing clear abstraction of models, capabilities and controls from the details of implementation and infrastructure complexity, and then providing real-time binding between them. This topic is expected to develop further in IIC to address some of the challenges we are facing in this area.
To conclude this series of blogs, Smart Maintenance, Global Optimization and Local Autonomy are the three main themes in Industrial Internet of Things system implementations, all aiming for realizing value through achieving operational efficiency while enhancing safety and resilience. A given IIoT deployment may emphasize some of these themes or some combination of them; it may even evolve through them in a different order. As with any attempt to abstract systems with great complexity and diversity, the analysis of these themes inevitably over simplifies some aspects or omits some others in specific systems. Nevertheless, they may still be useful as a starting reference to evaluate specific IIS deployments to understand its values and objectives and to anticipate its potential challenges. To have a clear understanding of these values is of foremost importance since after all, it is the business values in IIoT systems that drive their development.
(Acknowledgment: I have immensely benefited from many stimulating discussions with many of my colleagues in IIC, some of which are reflected in this series of blogs and to whom here I would like to express my gratitude, especially to Brad Miller from General Electric Global Research and Eric Harper from ABB Group. Brad is an unabashed and fervent believer in and promoter of autonomous systems – please stay tuned for his blogs on this and other topics.)