Industrial IoT Edge Architecture for Machine and Deep Learning
significantly less than 1 % of total video footage.
How do you distinguish between normal footage 99 +% of the time, vs. interesting events worthy of archival? It is based on image recognition and inference neural nets. Learning( generation and enhancement of the inference model / framework) happens at the Platform tier based on as many camera feeds as possible. Big Data storage and processing in action. Perhaps with human / expert assistance to train on identifying events of interest. So, gradually the system becomes better at detecting more and more important events. Periodically, upon non-trivial improvements to the inference framework, an update is pushed to all the edge devices for better inference( i. e., judgement / harvesting) of“ interesting events.” In this truck use case, this manifests as an OTA software update.
What we describe above is the application / data plane, which layers on top of and is enabled by the control / infrastructure plane, provided by Pulse IoT Center. IoT Center insures the secure, reliable and orderly functioning of the different components of the system, and the data flow from end device to platform through the edge device. IoT Center also manages the OTA software updates. Following are more details of the control plane.
EdgeX on-boards the cameras as end devices( the initial version of EdgeX does not support video; however, this could be added in the future – our emphasis in this example is on the architecture), and populates the device metadata. An EdgeX microservice processes the footage from each camera and applies inference model / incidence detection.
Another EdgeX microservice stores important event footage. Then an export microservice transmits this footage to a Platform tier service. Liota discovers on-boarded cameras through the EdgeX device metadata microservice and proxies these devices for management by IoT Center. It obtains infrastructure metrics( e. g., heart beat signal) to monitor and ensure proper function. Liota also obtains metrics from the edge device itself( e. g., storage availability and CPU load), for monitoring and managing. In addition, IoT Center is used for OTA software updates of the EdgeX inferencing microservice, and for the latest security patches.
CONCLUSIONS
We presented architectures for the edge and platform for real-time and batch data collection, model training, model management and model inference at the edge and Platform tiers. This is a comprehensive solution for Deep Learning in Industrial IoT which can be used in many use cases. One use case involving sensor-based trucking with open source and commercial solutions is also shown using our architecture.
IIC Journal of Innovation- 69-