Skip to content

Latest commit

 

History

History
135 lines (100 loc) · 4.82 KB

logs.md

File metadata and controls

135 lines (100 loc) · 4.82 KB

logfile

01/28 1h

created repository structure in cluster
added authentification
added chatgpt functionality for commenting

aims
quick restructure of the processing functions
hilbert trafo iteration convergence behavior comparison

01/29 1h

added clearer mode extraction from dimensionality reduction results for the iteration convergence routine
started cleaning up the toolbox

01/30 2h

mode extraction not working properly
just use the directly saved data, maybe build some quick struct to work on
protoperiod not working properly, possibly wrong vari
iteration scheme itself works

naive iteration of hilbert trafo does not converge
there is little amplitude modulation on all modes

02/01 2h

embedding length of 1 year works better
minor amplitude modulation on harmonics 2,3,...
starts to incorporate different information then resolution of oscillation
need some plots to explain this
a lot of 'trend' modes in NLSA, sometimes split up of oscillatory  into 2 'linear trend'

maybe new approach to code logging: new analysis script, instead of rework

02/02 2h

strategic paper: the main thematic issue is the disentangling of the non-sinuisoid oscillatory trend
(that is distributed among different harmonics and is assumed to have persistent amplitude)
from the rest of the information (such as trend, amplitude modulations)

BUT: this is possibly not the goal of the work on the paper

it is confirmed now that pure hilbert iterations do not converge
there is a tiny tiny portion of modes, mainly first and second harmonic
that have persistent zero_counts up to 3 iterations, e.g. the amplitude modulation is low enough

dirty hard coded filter to find these modes works
this 'kind-of-works', but really shuldnt be published ?

02/06 1h

combined plotting on the results of the naive hard filtering

02/08 1h

since iterated hilbert trafo does not naturally converge as a frequency estimation criterion,
lets use spectral power by fft  to show the harmonic structure

the question of ridicolous modes in NLSA: compare W=7a with W=1a for all spectral results

single fft implemented
repaired access to the broken metadata file

need to create large figures for individual spots w. signal, and multiple modes, spectra

02/09 2h

put the fourier spectra for individual modes together into large overview plots for W=7 and W=1

02/13 8h

created the cauchy based sorting algorithm for frequency harmonicity estimation based on threshold
due to the steepness of the cauchy peaks, the maximum of the residuals couldnt be used for threshold differentiation
finally, the un-normalized gauss peaks with free width yielded the best spectral fits
finished some threshold estimation table, mediocre results, keep it at conservative 10%
finished the sorting and plotting routine to compare SSA with NLSA

results are pretty miserable: the goal of steady-amplitude trend with individual trends is missed by the selected modes
there is still so much amplitude modulation in the RC
are the actual EOF responsible for it?
jump: they are, lets do a high resolution W=1 re-calculation with half hourly data ATTENTION this would also induce some day night dynamic?
so maybe check them manually

the FFt of the NLSA trends have the additional harmonics, this is looking good so far
NLSA sucks in the case of some signals, the parameter tuning might not have been good enough

02/15 8h

gaussian based harmonicity algorithm:
unnormalized spectral power
and free third gaussian parameter
lead to pretty stable results

02/22 2h

presentation pictures

03/02 4h

figures work: fft peak & seasonality behavior

03/04 2h

figures work: comparison ssa/nlsa and fundamental/harmonic

03/05 2h

bigf figure
modes figure
reconstructions figure

03/06 1h

started working on the year diagram for amplitude and phase

03/07 2h

refined figures from 03/04 and 03/05

03/08 3h

finished poster thoughts
run on the year seasons

03/13 8h

put togetehr figures for
    reconstruction + FFT
    EOF shapes + FFT
    seasonality + FFT (not with fundamental yet)
    change of seasonal behavior
found out the additional variables

03/14 8h

experimented with the projection possibility on dummy RC
in order to increase FFT resolution of W=1a EOF
didnt work out, can only do projection by duplicating them
didnt check spectra yet

03/19 1h

prepared Float32 data tensor for computation of the SSA and NLSA
of the additional variables

03/22 2h

ran all the variables
started to work on automatized display functions
no: think what we really need