I’m working on a variable length low level wireless signal classification problem. I’m wondering if someone came across interesting architectures for such a problem. I’m thinking I will need LSTMs or GRUs for the variable length signals. I’m also thinking about having convolutional layers to extract features.
The input is a 2 dimensional signal spread over time. (In-phase and quadrature components, https://en.wikipedia.org/wiki/In-phase_and_quadrature_components). The first 2560 instantaneous I/Q signal samples represent the preamble region of a transmission. This is the region of the transmission that our model uses. These physical signal samples can be of a variable length in the time domain.
The output is a 2 class response. Is the device producing the signal part of the group of trusted devices (class 1) or is it outside the group of trusted devices (class 2). We are classifying the physical signals and the factory imperfections that create differences among each device’s signal.
I’m just partway through Part 1 right now, so I don’t know much about recurrent networks. But reading your description of the problem as a 2D signal tracked over time, I was reminded of the blog post by @gesman where he classified users based on mouse movements over time. The time series data could be represented in an image, with a CNN applied to extract features. Just an idea!