0
$\begingroup$

I'm looking into using clustering-based method to detect anomalies in vibration signals. The idea is to extract features for every single sliding time-window of a time series of normal vibration data, and then find clusters of the feature vectors. Given a new period of vibration data, the model will classify the period as anomaly if its corresponding feature vector does not belong to any of the normal clusters.

Is there any literature on under what condition such methods will fail? My intuition is that clustering does not capture time-dependence of data. For instance, if I use Gaussian Mixture Model for clustering and find two clusters $N_1$ and $N_2$, the sliding windows of my normal vibration data may well form an alternating pattern, such as $N_1 N_2 N_1 N_2...$, while an anomalous vibrational time series has pattern $N_1 N_1 N_1...$, even though the data points in the anomalous vibrational data will belong to the normal clusters.

I'm not sure if this intuition is valid, though, for natural vibration. Say, if I plan to use the clustering-based approach to detect anomalies in machine vibration, is it even physically possible to have different vibrations that have same clusters as described above? If it's possible, is there a way to systematically generate realistic vibration data to show the weakness of such clustering algorithms?

$\endgroup$

1 Answer 1

0
$\begingroup$

Anomaly detection for time-series is often done using an autoencoder, and using the reconstruction error as the anomaly score. This has the advantage that it can learn feature representations, which makes it able to handle more complex patterns than a GMM is able to do.

The autoencoder architecture would be some variant of a Convolutional Neural Network or Recurrent Neural Network. For vibration data it can operate directly on X,Y,Z amplitudes, or on a time-frequency (spectrogram) input.

Note that it is possible to also use an autoencoder only for the representation learning, and passing the encoded features to a model such as a GMM.

$\endgroup$
1
  • $\begingroup$ Thanks, @jonnor. I do have plan to use a basket of algorithms. GMM is one such candidate, and I'd like to really understand where it falls short before investing in it. $\endgroup$
    – user159566
    Feb 19, 2020 at 22:00

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.