The IoT characteristic tracker collects vibration data of a machine using an accelerometer. It uses signal processing and extracts characteristic features using two different techniques:
- Statistical Feature Analysis of the Wavelet Packet Decomposition
- Area under the Curve of the Power Spectrum Density
These methods produce an input which is fed into machine learning algorithms, a Feed-Forward Neural Network and Deep AutoEncoder.
The system can determine the state of health of the machine accurately when it was trained in a supervised manner (with output labels). Deviation from a normal operating condition can also be determined with medium accuracy without any knowledge of the output (unsupervised). The system developed can easily be enhanced for specific machine health monitoring requirements.
In this demonstration, the machine status, machine condition and fault status are all determined by the machine learning algorithms from vibration data alone. There is no voltage or current monitoring and no connection to the machine power source.
The current graph shows the health of a rotating machine in the FlexWare office with 100% being perfect health and 0% being faulty. We run the machine in a hostile environment and have exaggerated the scale to make the data more interesting.
The system would allow detection of an imminent failure, allowing longer maintenance cycles on non-critical machinery or early warning of an unplanned shutdown on critical machinery.