Find-out abnormal behavior over the time

I want to detect abnormal behaviour in a oil pipe where the oil is flowing with some constant pressure. I have a sensor which monitors and pressure over the time and push it to my cloud server.

I got the dataset . My requirement is to do data analytics in python and find-out the abnormal pattern in the dataset over the time. And I need to suggest abnormal behaviors present in the dataset ., say for example., from 1 pm to 3:30 pm today , the pressure raises and falls may be due to some leakage in the pipe.

can we do it by simple statistical model? or machine learning is required?

Can please suggest the best suitable machine learning algorithm for this scenario?

Thanks

• Visit (sections of) the pipeline. Talk to the people who maintain it. There is only so much that you can see from data. You never know if someone is digging near the pipe that might cause fluctuations in the data which are not caused by leakage. – phiver Feb 13 '18 at 14:12

You have time series data which is used to measure the pressure using your sensor. You wish to identify when the pressure recordings are abnormal. This problem would be best solved using anomaly detection algorithms. But, there are so many ways that you can approach this problem.

I would use a sliding window approach and use that as your feature space to detect the distribution of your recordings. The window length you will select $m$ is the first of your hyper-parameters that you will need to tune.

Using a statistical model

The retained time series $X \in \mathbb{R}^m$, you can treat this signal as a queue, first in first out. When a new recording is defined then discard the oldest datapoint. For a given set $X$ of samples, get the mean and the standard deviation $\sigma$. If the new point exceeds a multiple of $\sigma$, usually this is set to $3\sigma$, but it depends on the expected variation of your sensor, then we flag a state change.

You can further extend this by using the generalized likelihood ratio test (GLRT) to determine when a new sample causes the point to fall significantly outside the distribution of the null hypothesis. In which case this indicates a state change (the new point came from a different distribution than the normal flow through the pipe).

Using machine learning

You can collect multiple instances of your time series and annotate them based on your experience as being nominal or anomalous. Then you will have a supervised 2-class classification problem. First you should attempt some feature extraction using PCA, LDA or similar techniques.. Then you can attempt to use all the fun algorithms available to you through scikit-learn (SVM, Random Forests, K-NN, etc.).

If there is a significantly higher number of nominal instances than anomalous ones, this will introduce bias to your model. Anomaly detection algorithms are better suited for these types of problems. These algorithms learn the distribution within which your nominal set should belong. Then for novel instances it evaluates the probability of it being contained in the learned distribution. If the probability is small, then the algorithm will flag the instance as anomalous.