Non-Invasive Real Time Error State Detection for Tractors
Using Smart Phone Sensors & Machine Learning 2019-26-0217
Condition Monitoring is the process of identifying any significant change in operating parameters of a machine. If the change in signature is an indication of an impending failure, a planned maintenance can be done to reduce revenue losses. This approach, termed as predictive maintenance is widely adapted in static machinery. Same approach is now being increasingly explored in the field of automobiles and farm machinery.
In India, tractors are extensively utilized during peak agricultural seasons but often poorly maintained during off seasons. This could lead to poor efficiency and untimely failure of critical components. One such problem is the “disturbed engine tappet setting”. Operating a tractor with mis-adjusted or loose tappet can lead to reduced performance, increased fuel usage and can even lead to engine damage over time. Since this problem does not stall the tractor immediately, the operator generally ignores this problem. There are multitude of such problems which can be detected with condition monitoring.
A non-invasive condition monitoring system could be the easiest solution to the tractors already in the field. Hence, we have considered the sensors available in smartphones (accelerometer, gyroscope and microphone) for sensing and investigating the problems which could be identified by noise and vibrations. In the case of engine tappet mentioned above, a disturbed setting causes a distinguishable noise, referred to as “tappet rattle”. We have used android smartphones to record the audio from tractors in good condition as well as in disturbed condition for such common problems.
Time series data analysis is done to extract relevant features and then Discrete Fourier Transform is applied to the audio signal for frequency domain signatures. Frequency domain-based features have shown significant improvements in the model prediction accuracy. This paper compares different classification algorithms and evaluate the results on performance and accuracy. The trained model is then used along with a smartphone application to do the real-time detection without any additional sensors.