How to apply neural pattern recognition to evaluate time-series data?

I have a time-series dataset of a dynamic system that I would like to validate. So far that was done manually, so I have available a wide variety of trained data and I would like to check out, if it is possible to automatize this evaluation through a neural pattern recognition network.

For instance, I have a numeric array that represents longitudinal velocity and its time reference. By plotting it, it is quite easy to conclude, if the behaviour is what it should.

As you see, the main point is that instead of having a set of "scalar" variables like size or weight that represent my data, I have vectors. In other words, my input dataset for the neural network would be a 3D array (features x entries x timesteps) whereas my target variable is still a binary vector (ok or not ok).

Is that possible? Or should I just preprocess every vector to characterize it as scalar magnitudes such as maximum and minimum values, mean, standard deviation, etc?

[ampaul]

Did you ever find your answer to this question? I am having the same trouble... Thank you

[Pablo]

I kept searching and found some research about "Multichannel time series convolutional neural networks", mostly used so far for human activity recognition or health diagnostics.

These are some very interesting papers I found on this matter:

• Deep Convolutional Neural Networks On Multichannel Time Series For Human Activity Recognition
• Data Augmentation for Time Series Classification using Convolutional Neural Networks
• Exploiting Multi-Channels Deep Convolutional Neural Networks for Multivariate Time Series Classification
• Time Series Classification Using Multi-Channels Deep Convolutional Neural Networks
• LEARNING TO DIAGNOSE WITH LSTM RECURRENT NEURAL NETWORKS
• Multi-Scale Convolutional Neural Networks for Time Series Classification
• Time Series Classification from Scratch with Deep Neural Networks: A Strong Baseline

Also some code in GitHub:

• cnn-timeseries
• Multilabel time series classification with LSTM

And other useful links:

• Sensor fusion and input representation for time series classification using deep nets
• Implementing a CNN for Human Activity Recognition in Tensorflow

I tried myself to start coding taking into account the hints and code shown above, but the scope of my project change and I haven't had for the moment time to keep developing this idea. I hope this is a good starting point for you and other people interested in this topic.

Answer 1

[Greg Heath]

You do not mention any dimensions. So it is hard to make an informed comment.

If you can use feature extraction to represent each input by a reasonably short vector,

then do so.

Hope this helps.

Thank you for formally accepting my answer

Greg

[Pablo]

Hi Greg,

Thank you for your answer. I am afraid that feature extraction is not the best approach in this case. Data is too complex and I think, I may miss some relevant information required for the further classification.

Anyway, I found some related papers posted above that seem interesting and may lead to the desired goal. I'll post again, if I have further information on this matter. For the moment I leave it open, in case someone can suggest a new approach or even a complete solution.