STMicroelectronics 2022
15
ALGORITHM SELECTION
Once the data is collected and processed, it is used
for algorithm selection and training. An engineer
will train a set of candidate algorithms and compare
their performance to choose the most optimal one.
This process is bounded by algorithm availability and
computing resources to run the search.
The limited resources of small footprint edge
devices can severely restrict algorithm selection. It is
counterproductive to select an ML algorithm that runs
flawlessly but requires multiple times the available RAM or
processing power. The NanoEdge™ AI Studio automates
library searches for anomaly detection, condition
monitoring, and predictive maintenance, and provides
algorithm comparison analysis and asset performance
management support to achieve optimal solutions.
ALGORITHM VALIDATION
At this fourth step the ML algorithm has been selected
for implementation. However, further testing is
required to know how the algorithm will perform
when deployed. This testing can be conducted
with real world data in the field, with a subset
of the data not used in previous steps, or by
finding edge cases that were not accounted
for previously. This validation step may require
earlier steps to be repeated if gaps are found
in the data collection, data processing, or
algorithm performance.
STM32MP1 MCU
SERIES
NANOEDGE
AI
TM
STUDIO
Learn More
Learn More
ALGORITHM OPTIMIZATION &
LIBRARY GENERATION
The last two implementation steps
are algorithm optimization for the
target platform and library generation.
Optimization takes time and requires
algorithm expertise, but it can be entirely
skipped with NanoEdge™ AI Studio, which
automatically selects and optimizes
for available resources, data, and use
case needs. Once the model library is
generated, corresponding control logic
is written for calling the algorithm and
taking subsequent action according to
predictions.
Conclusion
Edge-AI puts ML to work everywhere. ML
algorithms are becoming ever more indispensable
for local data analysis and decision making. STM32
AI solutions allow any developer to select, train, and
deploy ML algorithms for independent learning and
inference at the edge.
The NanoEdge AI™ Studio solution handles the time-
consuming tasks of algorithm selection, testing, and
compilation to reduce project effort, time, and cost.
Once deployed, ML algorithms continue learning
from new data, thereby improving the machine's
predictive and prescriptive capabilities. ■
