IoT Machine Learning: Vibrational Anomaly Detection using Tri-Axial Accelerometer Data
From bridges to cranes to servers to trains, the modern world is filled with essential equipment and infrastructure. Breakages endanger safety, halt operations, and incur high financial costs. Intelligent vibration analysis enables preventative maintenance, detecting degradation or faults before disaster strikes. Combining this algorithm with Internet of Things (IoT) technology allows for sensitive analytics on easily deployable, inexpensive, small form-factor devices.
This project introduces a statistics-based, time-series Vibrational Anomaly Detection algorithm that relies on extreme memory efficiency to fit onto highly constrained devices. Many other machine learning algorithms are too resource intensive to fit inference on severely resource-constrained devices. However, this project fits both the training and inference phases on a true edge device while running analytics in real time. The development and testing was run on the BBC Micro:bit. This device runs at only 16MHz with only 16kb of RAM, classifying the device as a Class I IoT device.
An accelerometer captures tri-axial accelerometer data accounting for the x, y, and z dimensions in a series of time windows. These data points within each window are manipulated into a collection of features used to detect anomalies. Features include averages, mins, maxes, and frequencies for x, y, and z data as well as magnitude, which takes each dimension into consideration.
To fit within the tight memory constraints, training is broken into two phases. Phase I collects data for a configurable number of training windows and computes the averages. Phase II uses these averages and, using incoming data, computes the standard deviations for each feature. Again, the number of training windows is configurable. Finally, an infinite inference loop analyzes new time windows, computes averages, and determines how many standard deviations the new value is from the trained average. If this z-score value is greater than a threshold, that feature is marked as suspicious. If the number of suspicious features exceeds a threshold, an anomaly is reported.
In this project, an anomalous event triggers a radio transmission. This message is received by another Micro:bit, which flashes its LED grid to notify a human.
See the attached paper for more thorough documentation.
Code running on the Micro:bit must be flashed onto the device. The development environment provided by BBC Micro:bit requires internet connection; therefore it is recommended to download and install the MU editor (https://codewith.mu/). To flash code to the Micro:bit, simply copy and paste the code into the editor and press the "flash" button. Note that each time you flash, the entire filesystem is wiped and you will lose any changes/additional files.
There are two versions of the vibrational anomaly detection (vib_detecion_mini_d1.py and vib_detecion_long__d1.py). This is because the extreme memory requirements necessitate code minification to avoid running out of memory.
Flash "vib_detecion_mini_d1.py" onto device #1 for vibrational anomaly detection
Flash "anomaly_alert_d2.py" onto device #2 to recieve anomalous alerts over radio
This code was designed for a BBC Micro:bit. This repo involves a system of two Micro:bits working together; one runs anomaly detection and the other listens for radio transmissions containing alerts. The second device lights up its LED screen (in this case with a skull image) to alert a human.
Using only one Micro:bit -- Comment out the radio transmission functions in vib_detecion_mini_d1.py
Using different hardware -- Comment out the radio transmission functions, as these are Micro:bit specific. Commend out all LED display functions. (Note that this project requires micropython on-device to run)
Questions? Contact Heather Whittaker at hwhittaker@technicacorp.com