How to design a ANN for signal processing properly?

Question

I don't really understand how I should implement following problem with help of ANN:

Let's say I have a few data sets, which represent the force on a car. (The test car drives on a special road with simulated road damages. All in all there are 8 different ones).

Now I want to train my model to recognize these road damages on data from cars, which drive on normal roads. That's the goal.

The question for me is, how do I classify the road damages correctly for the ANN?

Should I just cut out the special patterns of the road damages and feed my ANN with it?

Thanks in advance!

Answer 1

First of all, if you want to traing an ANN for a any specific task you generally have to complete the following steps:

1. Define the problem
2. Collect and correctly arrange training data
3. Design a suitable network topology
4. Train the network
5. Validate the training result on real data

That means for your case:

1. It seems like you are dealing with a binary classification problem. For a given sensor input you want to predict either damaged or not damaged.
2. You arrange your data as input-target pairs. Input is your sensor data (probably with some temporal context) and output is for example either 1 for damaged and 0 for not damaged.
3. If you are dealing with temporal sensor data you should check out recurrent neural network or convolutional neural network topologies (or a combination of both). Design your network to have the input layer the same amount of neurons as the dimension of your input data and the output layer with a single neuron with a sigmoid activation.
4. This step should then be straight-forward depending on the libraries you use
5. Finally, you can validate the result with some new real sensor data and check if the predictions match your expectations

You will probably need some iterations of step 3-5 until you come up with a well working classifier.