The document discusses adding intelligence to LoRaWAN devices by extracting meaningful features from raw sensor data on the device. It notes that currently, 99% of sensor data is discarded due to constraints, but extracting features could enable anomaly detection, classification, and forecasting on the device. It proposes a two-step process: 1) collecting high-resolution raw data from sensors, and 2) extracting features from the raw data on the device to perform intelligent analysis without transmitting all the raw data.
3. 3
Typical LoRaWAN sensor in 2019
Vibration sensor (up to 1,000 times per second)
Temperature sensor
NFC
Water & explosion proof
Processor capable of running >20 million
instructions per second
4. 4
But... what does it actually do?
Once an hour:
• Average motion (RMS)
• Peak motion
• Current temperature
5. 5
99% of sensor data is discarded due to
cost, bandwidth or power constraints.
https://www.mckinsey.com/~/media/McKinsey/Business%20Functions/McKinsey%20Digital/Our%20Insights/
The%20Internet%20of%20Things%20The%20value%20of%20digitizing%20the%20physical%20world/The-
Internet-of-things-Mapping-the-value-beyond-the-hype.ashx
8. 8
Interesting questions require more data
Classification
What's happening right now?
Anomaly detection
Is this behavior out of the ordinary?
Forecasting
What will happen in the future?
9. 9
Two problems
1. Very hard to answer with rule-based programming
if (accX > 9.5) {
lorawan.send('someone picked me up');
}
2. Needs to happen on-device, because
of this little pesky thing called physics
11. 11
Step 1 - Getting raw data
High-resolution data straight from devices (100 Hz)
Correct labeling
Offloading probably not over LoRaWAN (but signaling could)
https://pixabay.com/photos/factory-night-view-industrial-pipe-1769429/
19. Classification
What's happening right now?
Anomaly detection
Is this behavior out of the ordinary?
Forecasting
What will happen in the future?
19
Step 3 - Letting the computers figure it out
20. 20
Picking the right algorithm
Classification
Neural network
Anomaly detection
K-means clustering
Forecasting
Regression
21. 21
Neural networks
Now possible on device
Size of the network still matters (code size + ops)
Signal processing is key to smaller networks
22. 22
Neural networks aren't the only game in town
Classic ML algorithms are much smaller
K-means clustering is super efficient
(just compare new data against clusters)
Combining classic ML and NN is great
23. 23
Step 4 - Deploying
Don't continuously sample.
Monitor model performance.
Something weird? Send DSP result back to network.
26. 26
Collecting data
Collect on same device and same sensor
Store raw data in flash
Sync via WiFi or serial
Labeling directly on device
Capture all variations
DATA COLLECTED
12m 1s