Using Recurrent neural networks for real-time pattern recognition.

Good evening everybody,

I'm developing a Neural Network (patternnet) for recognition of phases of a standardized movement composed by 4 sequential steps.

Now basically my observations are composed by the signal of 8 accelerometers sampled at 100 ms (since i want a real time classification) and the model works pretty fine. However my next attempt is to use just one accelerometer and since i'm losing information about the movement, i would like to introduce the information about the sequence of motion (A -> B -> C -> D).

Now i was thinking about using a RNN to develop this but i'm not sure how to prepare my data:

• the dataset is composed by 6xN°observation which correspond to 1xN°classes (which are basically phases A B C or D);
• every 100 ms i should classify one class based on the actual inertial signal and (possibly) the past phases detected;

Does someone have some hints or suggestion on how to implement it ?

UPDATE:

In last weeks I considering use various option, and it seemed that starting from patternnet implementation of a classification problem represent a viable way to solve my question.

So basically i create a default network for pattern recognition:

net=patternnet(10);

and then i modify some options:

net.outputs{2}.feedbackMode='open'    % create an openloop structure
net.inputConnect(1,2)=1 % connect the feedback input

now i have two inputs one external and one feedback supposedly from previous output states of my classifier, let's say I want my model being able to recognize a real time input relying on previous 4 classified outputs.

My first plan is to modify the feedbackDelay property of outputs{2}:

net.outputs{2}.feedback Delay=4;

But from looking at documentation i can also work directly on the input line, setting:

net.inutWeigths{1,2}.delays=4;

So my questions are :

1- There is a difference between these two modalities?

2- How can I set not a 4 tapped delay line but a 1:4 tapped delay line ?