Training NARX neural network with variable length input

I want to train a NARX with 3 set of inputs each having 5 features (The output is categorical, so I am dealing with classification problem with dynamic inputs). the problem is each input set have variable length, for example:

input series 1: 5x150 (5=number of features) (150=collected data)

input series 2: 5x135

input series 3: 5x121

I know NARX works for constant length inputs, but is there anyway to train NN with variable length inputs?

P.S. I cannot use interpolation to adjust the lengths, since it changes the dynamic of data series. I cannot also use dynamic time warping, since I have more than 2 sets. and I prefer not to add zeroes to align the lengths

[Greg Heath]

Not clear.

Describe the target series .