DNN methodology and feature concatenation I'm using someone else's job and I have a question that I cannot solve.
This work uses a DNN to match an electrical resistance to a bend angle. This is not very important, just for the context.
So, here is the methodology :

*

*Choosing features

*Filtering and segmentation of the signals

*Concatenation of the features

*Finding best architecture (# hidden layers and # neurons) for the DNN

*Learning

The following features are used :

*

*Discrete Cosine Transform II (DCT)

*Sum of the absolute value of the FFT components (Sum FFT)

$$\mathrm{S} = \sum_{i=1}^M |C_i|^2$$


*Power Spectral Entropy (PSE)

*Signal Vector Magnitude (SVM - do not confuse with support vector machine)

$$\mathrm{SVM} = \frac{1}{n} \sum_{i=1}^n \sqrt{x_i^2}$$


*Differential Signal Vector Magnitude (DSVM)

$$\mathrm{DSVM} = \frac{1}{t} \left( \int_0^t \Big(\Big|\sum \mathrm{SVM}'\Big|\Big) \mathrm{dt} \right)$$


*Correlation coefficient (R)

*Root mean square (RMS)

In this work, the window size for segmentation is 24 samples (0.8 secondes at 30Hz). This gives us the following features size for the two signals (electrical resistance and angle): 24 x 2 (DCT), 1 x 2 (Sum FFT), 1 x 2 (PSE), 1 x 2 (SVM), 1 x 2 (DSVM), 1 x 3 (Correlation + 2 auto correlations), 1 x 2 (RMS). In total, the concatenated feature size is 61.
Then, the work I'm using concatenates the 61 size feature of 13 windows (called frames) : the centered window, 6 windows before and 6 windows after.
My question is : What's the point of concatenate the 61 size feature of 13 windows to create a 61 x 13 = 793 size feature ? I don't understand why we couldn't use only the 61 size feature of one window.
 A: Maybe there was some preprocessing involved before the training of the DNN?
After  computing the features, we usually regularise them (using normalization or standardization for instance). This will help the training process since all the inputs will have the same dimension. 
A: If I understand correctly, the input is a Concatenation of the 61 features x 13 frames to a vector 1 x 793.
This kind of concatenation is common for DNN. Eg in image recognition, one converts a 2d picture into a vector. It's useful since the next layer is a vector of neurons.
I think the reason why your colleague included the +/- 6 frames was to try to include signals from before and after into the learning model. Even as a vector, the DNN can learn relationships on the 61 features between ANY of the 13 frames. Not sure if doing that makes sense since you might be dealing with a signal in time and so I would think that a convolutional layer would better capture the relationships between adjacent frames. By concatenation all the 13 frames, the DNN may learn some relationship between the -6th and +6th frame to predict the bend angle, and using your domain knowledge you'll have to judge whether that makes sense.
