what would be the suitable machine learning algorithm for Pattern/Event Detection for the following Time Series Data?

Question

Time Series Data Plot

Please refer to the above link to understand the Data View.

Background for the data : It is the data of a single variable from a machine like Bulldozer(Pressure of the Hydraulics which is responsible for the movement of its Bucket), which performs actions like Loading its bucket and then move to the vehicle or place to dump the loaded material and then Dumping the material.

If you have seen the Image at the provided link, I have marked the Load Event(Loading the Bucket), Haul Event(Machine moving to dump), Dump Event(Dumping the load).

So one Load Event, Haul Event and Dump Event constitutes a Complete Cycle. In the image provide I see 12 such cycles.

Problem Statement: Detect the count of such cycles in the data provided, also eliminate the noise(have marked Noise in Red in the image). And calculate time taken by each event as How much time it took for load event, haul event and Dump event ?

Combining these three gives complete cycle time.

I tried to detect using moving average but its doesn't fit well.

Can anyone suggest a machine learning/ANN/better way which can accurately detect the event ?

Answer 1

It has been my experience with these types of problems that you try a few models and compare RMSE. Based on the initial results you can narrow the models down and attempt to refine these. As for what to alter I would refer you to "Introduction to Machine Learning" 2nd Ed by Ethem Alpaydin. This book explains the what is happening so that you can better guess at what to alter to get better results. Another resource - An Introduction to Statistical Learning with Applications in R by Gareth James, Daniela Witten, Trevor Hastie and Robert Tibshirani.

Intro to Statistic Learning PDF