Industrial Internet: Towards Interoperability and Composability
5.2 Long-Term Approach: Peer to Peer, Distributed, Ad Hoc, Amorphous Computing
There is, however, a different way to approach Industrial Internet systems that does not lessen the effort for an initial system, but does address some of the challenges that would press us toward having to keep duplicating our first copy( and the limitations inherent therein) for future systems. This approach requires us to change our systems engineering process: Rather than depending on an offline certification of requirements like safety, we move to an online approach that relies more on bounds checking and behavior analysis than Monte Carlo methods to show deontological adherence for a particular set of( control) functions. The key change is to move the responsibility for making decisions from the cloud to the asset – making the asset itself autonomous. Assets become responsible for their own analytic processes, which, if they are deployed locally, may well be part of a real-time response cycle. More importantly, assets can distribute problem solving between them as resources become available. Such a local set of ad hoc computational services can help with the limitations of embedded computing within a particular asset to take advantage of unused resources in nearby equipment, either stationed intentionally( like a set of local servers within a factory) or unintentionally( like the processing power of a plant visitor’ s cell phone). We can still take advantage of the“ amorphous computing” concept that cloud gives us, without actually sending anything offsite!
By moving the responsibility of asset operation to the asset itself( or its analogue in the digital twin sense), we can filter any data necessary for broad-spectrum fleet level analytics, turning such approaches away from‘ big data’ problems where massive amounts of data need to be transmitted. Policy changes that are induced by such off-board analysis become suggestions, not commands – each individual asset makes the decision as to whether the new policy will improve its operation. Single points of failure introduced( e. g., by security operations) are mitigated by assets that can be suspicious of any attempts by privileged operators to effect changes – the asset, as autonomous, will have the final say as to if and when such requests will be acted upon.
Moving analytics to the asset addresses privacy( no more process information escaping the company’ s assets to non-owned resources) and resiliency( no single point of failure to cloud services can cause the asset to fail). Because network users cannot be expected to know all safety constraints, asset autonomy also enables an asset to refuse requests for safety reasons. Problems being solved in context may be simpler than those that are removed from the context. However, this does introduce challenges associated with requiring higher levels of intelligence on the device, including how situations are represented and understood, how to enable online machine learning( ML) rather than offline analytic ML approaches of today 38, etc.
38 Promising advances in hardware, such as https:// www. technologyreview. com / s / 526506 / neuromorphic-chips / should lead to new applications of online learning in industrial settings.
- 72- June 2016