You can read about Dynamic Time Wrapping, which can overcome problem of varying lengths.

I think this paper (which deal with more complicated problem) might give you some ideas how to tackle the problem

Park, Alex S., and James R. Glass. "Unsupervised pattern discovery in speech." IEEE Transactions on Audio, Speech, and Language Processing 16.1 (2008): 186-197.

You can read about Dynamic Time Wrapping, which can overcome problem of varying lengths.

I think this paper (which deal with more complicated problem) might give you some ideas how to tackle the problem

Park, Alex S., and James R. Glass. "Unsupervised pattern discovery in speech." IEEE Transactions on Audio, Speech, and Language Processing 16.1 (2008): 186-197.

Thank you so much. I will surely try Dynamic Time Wrapping and the unsupervised pattern discovery paper.

Thanks a lot. This is really helpful.

Recurrent neural networks such as Echo State Network and Liquid State Machines have their ability to "remember" past events, a sort of fading memory and so if you use one of these tools for classification then you do not need to window your data at all, unless the number of neurons is not large enough to "remember" your longest sequence.  If your longest sequence is relatively small, say less than 100, then a reasonably sized neuron pool of say 500-1,000 neurons will easily cover all cases.

However, as others have said, a knowledge of the waveform and where/how it was collected may suggest that extracting features via a windowed approach may be more productive.

If you are interested in recurrent neural networks then "The “echo state” approach to analysing and training recurrent neural networks-with an erratum note" by H Jaeger, 2001 is the starting point, while "A practical guide to applying echo state networks", Lukosevicius, 2012 provides practical tips on using this technique.  

My own work, "Using Echo State Networks to Classify Unscripted, Real-World Punctual Activity" DP Hunt and D Parry, 2015 may also provide some insights.

http://organic.elis.ugent.be/sites/organic.elis.ugent.be/files/PracticalESN.pdf

http://minds.jacobs-university.de/sites/default/files/uploads/papers/EchoStatesTechRep.pdf

http://link.springer.com/chapter/10.1007/978-3-319-23983-5_34

Dear Yaakov

You can use statistical feature and also you can use 1D-DWT .

Regards

Thank you so much. Echo State Networks seem quite interesting and appropriate for my problem/task. I really appreciate the help.

We developed a software named Bedforms-ATM which performs a scale-based discrimination of bedforms. You can download the beta versión from this link:

http://ronaldrgutierrez.com/data-software/

We'll be launching the final versión very soon.

Regards 

Dear Haider Adnan Khan, 

Look the link, please.

https://papers.nips.cc/paper/40-neural-networks-for-template-matching-application-to-real-time-classification-of-the-action-potentials-of-real-neurons.pdf

http://www.hindawi.com/journals/tswj/2014/647216/

http://www.ncbi.nlm.nih.gov/pubmed/15683526

Regards, Shafagat

Thank you so much.

Thanks a lot. I am matching Power Spectral Density between two signals. It's working much better than matching in time domain.

You will be better in  considering  wavelet in case of feature recognition. The key is multi-resolution principle. Haar wavelet is simple and powerful.

 Thank you Professor Sekine. Your suggestion is really helpful for me.

I agree with Andreas,

There's a lot of manners to classify EEG signals. The correct or more convenient way depends on the context and the purpose of your research. e.g. a "shape based" could be useful for evoked potential, features based on the power spectral density (PSD) are commonly used to describe changes on the dynamics associated with task realization. And the PSD could be estimated by several methods, FFT may be the less accurate given its high noise sensitivity, some other methods like periodogram, wavelet, or even short time fourier analysis achieves better results. Some other characteristics could be non-linear like entropy, scaling exponents, lyapunov coefficient, etc.

I suggest that first you make all signals of same length using Dynamic Time warping by appropriately padding the signals.

After that, you can use Sparse Coding to create dictionary of each waveform and feed it to a non-linear SVM classifier as a feature vector.

The advantage of using this method is that you are extracting feature vectors based on an unsupervised learning method where the objective(surrogate) function is minimised such that local minima is found for the population rather than the 'dataset'. Hence, overfitting problem is reduced, significantly.

Please let me know if this suggestion helped you.

Reference:

J. Mairal, F. Bach, J. Ponce and G. Sapiro. Online Dictionary Learning for Sparse Coding. International Conference on Machine Learning, Montreal, Canada, 2009

@Ankit Tripathi, Thank you so much. I will surely keep this in mind.