Dataset Description

This dataset is intended to accompany a hardware demo kit available from QuickLogic to show the potential for using the QuickLogic / SensiML solution to rapidly build time-series smart sensor data classification algorithms to translate raw physical sensor data into application-specific insight events using pattern recognition.  The hardware kit consists of the QuickLogic QuickAI evaluation board contained in a plastic enclosure and mounted atop a basic axial fan motor along with various fixtures that can be used to generate physical vibration events as might be deemed of interest for a fan monitoring application.  Such events include blade impingement, hub imbalance, and chassis shock.  Classification of motor state includes these events along with vibration detection of motor on and off states.

 

All classifications of fan events are based on monitoring only vibration/motion as detected by the QuickAI on-board inertial measurement unit (6-axis accelerometer and gyro sensor IC).  In such way, the demo simulates a common desired use case in industrial predictive maintenance applications for an over-the-top sensor system that can detect equipment usage, fault states, and degrading performance with minimal or no incursion into existing electromechanical systems used for control and operation.

 

The dataset consists of captured sessions of raw IMU sensor values (STMicro LSM6DSL on the QuickAI eval board) sampled at 104Hz.  Each session contains labeled segments of a given fan state/classification type from amongst the following class map of events for recognition:

Event #NameDescription
0UnknownBoard is undergoing some unrecognized movement
1FaultFan unit has an obstruction in the blades
2OffFan unit is turned off
3OnFan unit is turned on and in ‘normal’ operation state
4ShockSomething is tapping the fan unit from the top side

Depending on various factors like table surface, AC frequency, and other unknowns in your work space the demo may not work well under your specific conditions. This is a perfect example of how-to retrain the demo with more variance data so that it can build a more advanced algorithm that works under all conditions.

Below is a typical labeled input file (this one is Fault_01.csv) showing raw accelerometer and gyro data as collected during a capture session: