Multichannel Displacement measurement via self mixing interferometry and neural network : training and test datasets

<p>Self mixing interferometry is a simple and robust sensing method which can be used (among other things) to measure the displacement of a target along the light propagation axis. While conceptually simple, the actual use of this method is less straightforward than originally envisioned because reconstructing the target&nbsp;displacement from the interferometric signal is often tricky. A small neural network can do this task very well after proper training, as described in [10.1364/OE.419844], with dataset [10.5281/zenodo.7303745].&nbsp;</p> <p>Here, the&nbsp;dataset is composed by a training set and a test set, in a specific configuration in which 3 self-mixing sensors measure simultaneously the same target displacement. Both datasets contain&nbsp;the displacement itself and the 3 interferometric signals (1 per sensing channel)</p> <p><strong>The training set </strong>relies on two python/numpy data files corresponding to <strong>harmonic displacements</strong> for different frequencies ranging from 53&nbsp;and 93 Hz&nbsp;and amplitudes from 3.5 to 7.5 &micro;m&nbsp;:&nbsp;</p> <ul> <li>Training_set_2_lostchannel_displacement.npy : 93744-elements long numpy array containing the target&#39;s displacement in units of &micro;m/ms with a 1.024 ms&nbsp;time step.&nbsp;</li> <li>Training_set_2_lostchannel_signal.npy : numpy array of shape (3, 93744, 256, 1) containing the interferometric signals. The first dimension refers to the channel (1, 2 or 3), the second dimension is the number of segments of 256 points. Each segment of 256 points correspond to a&nbsp;1.024 ms window of signal, matching one element of the displacement.&nbsp;For instance, the displacement value in `displacement[416]` corresponds to the interferometric signal segment `signal[0,416,:,0]` for channel 1,&nbsp;`signal[1,416,:,0]` for channel 2 and&nbsp;`signal[2,416,:,0]` for channel 3.&nbsp;</li> </ul> <p><strong>The test </strong>set follows the same architecture and format as the training set, but contains only <strong>random displacements&nbsp;</strong>generated by a delta-correlated signal, which we Fourier filter with a fifth order Butterworth filter between 10 and 100 Hz :</p> <ul> <li>&quot;Displacement_test.npy&quot; : &nbsp;with shape&nbsp;(246078, 1)</li> <li>&quot;Signal_test.npy&quot; : with shape&nbsp;(3, 246078, 256, 1)</li> <li>Only the displacement type has changed from harmonic to random, from the training to the test datasets.</li> </ul> <p>These datasets have&nbsp;been used to train and test a 3 channel neural network (after data augmentation) in order to emphasize the high availability&nbsp;potential of multichannel schemes, against backscattered power fluctuations.&nbsp;</p>