That is going to be difficult.
First, the downward deflection needs to be converted into a peak, so use the negative of the signal to do that.
Second, it will be necessary to use the MinPeakHeight name-value pair (and perhaps some others, such as MinPrakDistance) to detect the peaks that you do not want.
After that, run findpeaks with no restrictions except MinPeakHeight (to avoid ths high-frequency noise, although you may want to filter that out first), then use setdiff or other functions as appropriate to isolate the locs output to the peak you want. (The filter step will require taking the Fourier transform of the signal to see the best frequency for the high-frequency cutoff. A lowpass filter will do this, however a bandpass filter can also eliminate the DC offset, making further analyses easier.)
The height of the peak you want (and all the others) will be the negatives of the peak heights that are returned, because the signal was negated first.
Automating that will require that all the signals have similar characteristics. You will have to experiment to see if that is true.
