This dataset utilizes a machine vibration analysis trainer (MVAT-6 from V-TEK Associates) commonly used to train factory maintenance and operations technicians how to manually diagnose machine fault states and running performance.
Instead, we can use this equipment to train the QuickLogic QuickAI development platform to automatically recognize the same machine states autonomously using a SensiML generated endpoint AI sensing algorithm running continously on the QuickAI endpoint processor node. For input signals, we use a low-cost Analog Devices ADXL335 3-axis accelerometer remotely wired to the point of interest on the machine and sampled at 1 kHz.The QuickLogic QuickAI development board is housed in the white enclosure in the front of the machine. The ADXL335 analog +/-3G sensor IC is housed in the red metal enclosure box affixed to the rightmost bearing block of the MVAT-6 (accompanying detail image). Connectivity between the remoted vibration sensor and the QuickAI processor board is accomplished using micro-coaxial leads and SMA connectors to minimize analog noise. Results of the model can be visualized in real-time using an Android smartphone running the SensiML TestApp. All classification is accomplished on the QuickAI endpoint processor and only classified state events are sent to the smartphone (using a Bluetooth Low Energy wireless link).The MVAT-6 machine itself consists of a DC motor and drive electronics connected to a load shaft via a replaceable flexible coupler, with the drive shaft supported by two bearing blocks. The right end of the shaft can accommodate a variety of removable aluminum disc rotors to simulate different inertial loads. One such rotor is machined to be well balanced and concentric within tight tolerance. Another such rotor is bored with a slightly offset center hole to induce an imbalanced load vibration. Yet another has a misaligned center bore hole to produce a wobbly or “cocked rotor” rotational load.Collected Machine StatesThe following table lists the combinations of data collected in different machine configurations. Several files exists across each configuration within the dataset.
|Machine State>||Off||Running, No Rotor||Running, Balanced Rotor||Running, Eccentric Rotor||Running, Cocked Rotor|
For each machine configuration, raw sensor data was collected and labeled for classification states per the table above. Of the 3 available orthogonal sensors available, only two axes (the ones orthogonal to the motor shaft axis) were required and collected.Additionally, other relevant metadata was also collected which could impact the sensor data classification. These include:
- Whether or not the machine’s protective lexan cover shield was in place or not during operation (theory being that its vibrational resonance might have an effect on the signal)
- Which QuickAI unit was used for processing (not ordinarily an issue but collected in case problems arose)
- The motor speed in RPM (as shown on the LED tach readout)
- The operator who performed the collection (always a good idea to annotate in case differences in technique or protocol between operators is subsequently discovered)
The chart below show a representative view of the signal waveforms for one of the collection files as viewed within SensiML Data Capture Lab: