Industrial IoT Edge Architecture for Machine and Deep Learning
Figure 13: Temperatures from four devices in use case example
hard to detect at the edge. When Device 4
temperature is compared with median
temperatures of all devices (in red) the
anomaly in Device 4 is detected at the
Platform tier.
The edge model can be an anomaly detection
model on a univariate time series. The
platform model can be an anomaly detection
model on a multivariate time series.
The raw data from edge is used to train edge
models whereas feature data extracted from
the raw edge data are used for platform
models. Models are stored in the model
cache. Edge models are pushed to the edge
inference engine, whereas platform models
are pushed to the platform inference engine.
Both models are served by the model serving
engine. Figure 7 shows the edge and platform
- 66 -
architectures for training and inference as
well as the data and model caches.
Note that the human domain knowledge
mentioned here for the Deep Learning
training and inference has several
components:
1. It is used to label training samples
manually or automatically through an
algorithm that applies rules based on
human domain knowledge.
2. It is used to choose the models used
for training.
3. It is used to provide auxiliary rules in
addition to the Deep Learning models.
4. It is used to interpret the inference
results and determine if the
interpretations match human domain
knowledge.
September 2017