How to design an inpute matrix for binary classification in Keras - machine-learning

Question

I want to predict lane changes with RNN based on sensor data extracted from vehicles. Therefor I have an input matrix containing periodic bus data (about 40ms) on every row. Each columns contains data from a single sensor. In sum, I have more than 300 columns with sensor data.
My output Matrix with label-data contains for every row entry in input a resulting boolean value representing a lane change. So they have both the same length.
Here for example:
..my input matrix: ..and output matrix:
+------+---------+--------+---------+...+ +-----------+
| ## |sensor 1 |sensor 2| sensor 3|...| |lane change|
+------+---------+--------+---------+---+ +-----------+
|time 1| 15.0 | 42.0 | 1 |...| | 0 |
+------+---------+--------+---------+---+ +-----------+
|time 2| 14.8 | 38.2 | 1 |...| | 0 |
+------+---------+--------+---------+---+ +-----------+
|time 3| 14.3 | 27.0 | 0 |...| | 1 |
+------+---------+--------+---------+---+ +-----------+
|time n| . | . | . |...| | . |
Keras needs for recurrent neural Networks (RNN) like LSTM an input structure in the shape of
[samples, time steps, Features]
Now my questions:
How do I have to reshape my input matrix to get working with LSTM?
Do I have to shift the whole input matrix to the next row for each timestep?
I suppose the last 2-3 seconds before lane change could be interesting for prediction, so I need many timesteps which causes also many copies of my input matrix.
Here what I think:
Input matrix at next timestep (shift row 1 line):
+------+---------+--------+---------+...+
| ## |sensor 1 |sensor 2| sensor 3|...|
+------+---------+--------+---------+---+
|time 1| 0.0 | 0.0 | 0 |...|
+------+---------+--------+---------+---+
|time 2| 15.0 | 42.0 | 1 |...|
+------+---------+--------+---------+---+
|time 3| 14.8 | 38.2 | 1 |...|
+------+---------+--------+---------+---+
|time 4| 14.3 | 27.0 | 0 |...|
Input matrix next to last timestep (shift row 2 lines):
+------+---------+--------+---------+...+
| ## |sensor 1 |sensor 2| sensor 3|...|
+------+---------+--------+---------+---+
|time 1| 0.0 | 0.0 | 0 |...|
+------+---------+--------+---------+---+
|time 2| 0.0 | 0.0 | 0 |...|
+------+---------+--------+---------+---+
|time 3| 15.0 | 42.0 | 1 |...|
+------+---------+--------+---------+---+
|time 4| 14.8 | 38.2 | 1 |...|
Is this right?
Please note that, I will try to reduce sensor data later. This is in the moment not my first aim.

Answer 1

From what I understood what you need to do is,
Decide how much the network need to see to be able to predict, what
I mean is you show n step and the network will predict n+1 step
Split your matrix into multiple matrices with n row and save the n+1 row lane change as label
Now you should have a list of k matrices of shape(n, m_sensor) which has shape [k-samples, n_stpes, m_sensor], this is exactly what the rnn needs as input.