Related
I would like to find the time instant at which a certain value is reached in a time-series data with noise. If there are no peaks in the data, I could do the following in MATLAB.
Code from here
% create example data
d=1:100;
t=d/100;
ts = timeseries(d,t);
% define threshold
thr = 55;
data = ts.data(:);
time = ts.time(:);
ind = find(data>thr,1,'first');
time(ind) %time where data>threshold
But when there is noise, I am not sure what has to be done.
In the time-series data plotted in the above image I want to find the time instant at which the y-axis value 5 is reached. The data actually stabilizes to 5 at t>=100 s. But due to the presence of noise in the data, we see a peak that reaches 5 somewhere around 20 s . I would like to know how to detect e.g 100 seconds as the right time and not 20 s . The code posted above will only give 20 s as the answer. I
saw a post here that explains using a sliding window to find when the data equilibrates. However, I am not sure how to implement the same. Suggestions will be really helpful.
The sample data plotted in the above image can be found here
Suggestions on how to implement in Python or MATLAB code will be really helpful.
EDIT:
I don't want to capture when the peak (/noise/overshoot) occurs. I want to find the time when equilibrium is reached. For example, around 20 s the curve rises and dips below 5. After ~100 s the curve equilibrates to a steady-state value 5 and never dips or peaks.
Precise data analysis is a serious business (and my passion) that involves a lot of understanding of the system you are studying. Here are comments, unfortunately I doubt there is a simple nice answer to your problem at all -- you will have to think about it. Data analysis basically always requires "discussion".
First to your data and problem in general:
When you talk about noise, in data analysis this means a statistical random fluctuation. Most often Gaussian (sometimes also other distributions, e.g. Poission). Gaussian noise is a) random in each bin and b) symmetric in negative and positive direction. Thus, what you observe in the peak at ~20s is not noise. It has a very different, very systematic and extended characteristics compared to random noise. This is an "artifact" that must have a origin, but of which we can only speculate here. In real-world applications, studying and removing such artifacts is the most expensive and time-consuming task.
Looking at your data, the random noise is negligible. This is very precise data. For example, after ~150s and later there are no visible random fluctuations up to fourth decimal number.
After concluding that this is not noise in the common sense it could be a least two things: a) a feature of the system you are studying, thus, something where you could develop a model/formula for and which you could "fit" to the data. b) a characteristics of limited bandwidth somewhere in the measurement chain, thus, here a high-frequency cutoff. See e.g. https://en.wikipedia.org/wiki/Ringing_artifacts . Unfortunately, for both, a and b, there are no catch-all generic solutions. And your problem description (even with code and data) is not sufficient to propose an ideal approach.
After spending now ~one hour on your data and making some plots. I believe (speculate) that the extremely sharp feature at ~10s cannot be a "physical" property of the data. It simply is too extreme/steep. Something fundamentally happened here. A guess of mine could be that some device was just switched on (was off before). Thus, the data before is meaningless, and there is a short period of time afterwards to stabilize the system. There is not really an alternative in this scenario but to entirely discard the data until the system has stabilized at around 40s. This also makes your problem trivial. Just delete the first 40s, then the maximum becomes evident.
So what are technical solutions you could use, please don't be too upset that you have to think about this yourself and assemble the best possible solution for your case. I copied your data in two numpy arrays x and y and ran the following test in python:
Remove unstable time
This is the trivial solution -- I prefer it.
plt.figure()
plt.xlabel('time')
plt.ylabel('signal')
plt.plot(x, y, label="original")
y_cut = y
y_cut[:40] = 0
plt.plot(x, y_cut, label="cut 40s")
plt.legend()
plt.grid()
plt.show()
Note carry on reading below only if you are a bit crazy (about data).
Sliding window
You mentioned "sliding window" which is best suited for random noise (which you don't have) or periodic fluctuations (which you also don't really have). Sliding window just averages over consecutive bins, averaging out random fluctuations. Mathematically this is a convolution.
Technically, you can actually solve your problem like this (try even larger values of Nwindow yourself):
Nwindow=10
y_slide_10 = np.convolve(y, np.ones((Nwindow,))/Nwindow, mode='same')
Nwindow=20
y_slide_20 = np.convolve(y, np.ones((Nwindow,))/Nwindow, mode='same')
Nwindow=30
y_slide_30 = np.convolve(y, np.ones((Nwindow,))/Nwindow, mode='same')
plt.xlabel('time')
plt.ylabel('signal')
plt.plot(x,y, label="original")
plt.plot(x,y_slide_10, label="window=10")
plt.plot(x,y_slide_20, label='window=20')
plt.plot(x,y_slide_30, label='window=30')
plt.legend()
#plt.xscale('log') # useful
plt.grid()
plt.show()
Thus, technically you can succeed to suppress the initial "hump". But don't forget this is a hand-tuned and not general solution...
Another caveat of any sliding window solution: this always distorts your timing. Since you average over an interval in time depending on rising or falling signals your convoluted trace is shifted back/forth in time (slightly, but significantly). In your particular case this is not a problem since the main signal region has basically no time-dependence (very flat).
Frequency domain
This should be the silver bullet, but it also does not work well/easily for your example. The fact that this doesn't work better is the main hint to me that the first 40s of data are better discarded.... (i.e. in a scientific work)
You can use fast Fourier transform to inspect your data in frequency-domain.
import scipy.fft
y_fft = scipy.fft.rfft(y)
# original frequency domain plot
plt.plot(y_fft, label="original")
plt.xlabel('frequency')
plt.ylabel('signal')
plt.yscale('log')
plt.show()
The structure in frequency represent the features of your data. The peak a zero is the stabilized region after ~100s, the humps are associated to (rapid) changes in time. You can now play around and change the frequency spectrum (--> filter) but I think the spectrum is so artificial that this doesn't yield great results here. Try it with other data and you may be very impressed! I tried two things, first cut high-frequency regions out (set to zero), and second, apply a sliding-window filter in frequency domain (sparing the peak at 0, since this cannot be touched. Try and you know why).
# cut high-frequency by setting to zero
y_fft_2 = np.array(y_fft)
y_fft_2[50:70] = 0
# sliding window in frequency
Nwindow = 15
Start = 10
y_fft_slide = np.array(y_fft)
y_fft_slide[Start:] = np.convolve(y_fft[Start:], np.ones((Nwindow,))/Nwindow, mode='same')
# frequency-domain plot
plt.plot(y_fft, label="original")
plt.plot(y_fft_2, label="high-frequency, filter")
plt.plot(y_fft_slide, label="frequency sliding window")
plt.xlabel('frequency')
plt.ylabel('signal')
plt.yscale('log')
plt.legend()
plt.show()
Converting this back into time-domain:
# reverse FFT into time-domain for plotting
y_filtered = scipy.fft.irfft(y_fft_2)
y_filtered_slide = scipy.fft.irfft(y_fft_slide)
# time-domain plot
plt.plot(x[:500], y[:500], label="original")
plt.plot(x[:500], y_filtered[:500], label="high-f filtered")
plt.plot(x[:500], y_filtered_slide[:500], label="frequency sliding window")
# plt.xscale('log') # useful
plt.grid()
plt.legend()
plt.show()
yields
There are apparent oscillations in those solutions which make them essentially useless for your purpose. This leads me to my final exercise to again apply a sliding-window filter on the "frequency sliding window" time-domain
# extra time-domain sliding window
Nwindow=90
y_fft_90 = np.convolve(y_filtered_slide, np.ones((Nwindow,))/Nwindow, mode='same')
# final time-domain plot
plt.plot(x[:500], y[:500], label="original")
plt.plot(x[:500], y_fft_90[:500], label="frequency-sliding window, slide")
# plt.xscale('log') # useful
plt.legend()
plt.show()
I am quite happy with this result, but it still has very small oscillations and thus does not solve your original problem.
Conclusion
How much fun. One hour well wasted. Maybe it is useful to someone. Maybe even to you Natasha. Please be not mad a me...
Let's assume your data is in data variable and time indices are in time. Then
import numpy as np
threshold = 0.025
stable_index = np.where(np.abs(data[-1] - data) > threshold)[0][-1] + 1
print('Stabilizes after', time[stable_index], 'sec')
Stabilizes after 96.6 sec
Here data[-1] - data is a difference between last value of data and all the data values. The assumption here is that the last value of data represents the equilibrium point.
np.where( * > threshold )[0] are all the indices of values of data which are greater than the threshold, that is still not stabilized. We take only the last index. The next one is where time series is considered stabilized, hence the + 1.
If you're dealing with deterministic data which is eventually converging monotonically to some fixed value, the problem is pretty straightforward. Your last observation should be the closest to the limit, so you can define an acceptable tolerance threshold relative to that last data point and scan your data from back to front to find where you exceeded your threshold.
Things get a lot nastier once you add random noise into the picture, particularly if there is serial correlation. This problem is common in simulation modeling(see (*) below), and is known as the issue of initial bias. It was first identified by Conway in 1963, and has been an active area of research since then with no universally accepted definitive answer on how to deal with it. As with the deterministic case, the most widely accepted answers approach the problem starting from the right-hand side of the data set since this is where the data are most likely to be in steady state. Techniques based on this approach use the end of the dataset to establish some sort of statistical yardstick or baseline to measure where the data start looking significantly different as observations get added by moving towards the front of the dataset. This is greatly complicated by the presence of serial correlation.
If a time series is in steady state, in the sense of being covariance stationary then a simple average of the data is an unbiased estimate of its expected value, but the standard error of the estimated mean depends heavily on the serial correlation. The correct standard error squared is no longer s2/n, but instead it is (s2/n)*W where W is a properly weighted sum of the autocorrelation values. A method called MSER was developed in the 1990's, and avoids the issue of trying to correctly estimate W by trying to determine where the standard error is minimized. It treats W as a de-facto constant given a sufficiently large sample size, so if you consider the ratio of two standard error estimates the W's cancel out and the minimum occurs where s2/n is minimized. MSER proceeds as follows:
Starting from the end, calculate s2 for half of the data set to establish a baseline.
Now update the estimate of s2 one observation at a time using an efficient technique such as Welford's online algorithm, calculate s2/n where n is the number of observations tallied so far. Track which value of n yields the smallest s2/n. Lather, rinse, repeat.
Once you've traversed the entire data set from back to front, the n which yielded the smallest s2/n is the number of observations from the end of the data set which are not detectable as being biased by the starting conditions.
Justification - with a sufficiently large baseline (half your data), s2/n should be relatively stable as long as the time series remains in steady state. Since n is monotonically increasing, s2/n should continue decreasing subject to the limitations of its variability as an estimate. However, once you start acquiring observations which are not in steady state the drift in mean and variance will inflate the numerator of s2/n. Hence the minimal value corresponds to the last observation where there was no indication of non-stationarity. More details can be found in this proceedings paper. A Ruby implementation is available on BitBucket.
Your data has such a small amount of variation that MSER concludes that it is still converging to steady state. As such, I'd advise going with the deterministic approach outlined in the first paragraph. If you have noisy data in the future, I'd definitely suggest giving MSER a shot.
(*) - In a nutshell, a simulation model is a computer program and hence has to have its state set to some set of initial values. We generally don't know what the system state will look like in the long run, so we initialize it to an arbitrary but convenient set of values and then let the system "warm up". The problem is that the initial results of the simulation are not typical of the steady state behaviors, so including that data in your analyses will bias them. The solution is to remove the biased portion of the data, but how much should that be?
I just started a Machine learning class and we went over Perceptrons. For homework we are supposed to:
"Choose appropriate training and test data sets of two dimensions (plane). Use 10 data points for training and 5 for testing. " Then we are supposed to write a program that will use a perceptron algorithm and output:
a comment on whether the training data points are linearly
separable
a comment on whether the test points are linearly separable
your initial choice of the weights and constants
the final solution equation (decision boundary)
the total number of weight updates that your algorithm made
the total number of iterations made over the training set
the final misclassification error, if any, on the training data and
also on the test data
I have read the first chapter of my book several times and I am still having trouble fully understanding perceptrons.
I understand that you change the weights if a point is misclassified until none are misclassified anymore, I guess what I'm having trouble understanding is
What do I use the test data for and how does that relate to the
training data?
How do I know if a point is misclassified?
How do I go about choosing test points, training points, threshold or a bias?
It's really hard for me to know how to make up one of these without my book providing good examples. As you can tell I am pretty lost, any help would be so much appreciated.
What do I use the test data for and how does that relate to the
training data?
Think about a Perceptron as young child. You want to teach a child how to distinguish apples from oranges. You show it 5 different apples (all red/yellow) and 5 oranges (of different shape) while telling it what it sees at every turn ("this is a an apple. this is an orange). Assuming the child has perfect memory, it will learn to understand what makes an apple an apple and an orange an orange if you show him enough examples. He will eventually start to use meta-features (like shapes) without you actually telling him. This is what a Perceptron does. After you showed him all examples, you start at the beginning, this is called a new epoch.
What happens when you want to test the child's knowledge? You show it something new. A green apple (not just yellow/red), a grapefruit, maybe a watermelon. Why not show the child the exact same data as before during training? Because the child has perfect memory, it will only tell you what you told him. You won't see how good it generalizes from known to unseen data unless you have different training data that you never showed him during training. If the child has a horrible performance on the test data but a 100% performance on the training data, you will know that he has learned nothing - it's simply repeating what he has been told during training - you trained him too long, he only memorized your examples without understanding what makes an apple an apple because you gave him too many details - this is called overfitting. To prevent your Perceptron from only (!) recognizing training data you'll have to stop training at a reasonable time and find a good balance between the size of the training and testing set.
How do I know if a point is misclassified?
If it's different from what it should be. Let's say an apple has class 0 and an orange has 1 (here you should start reading into Single/MultiLayer Perceptrons and how Neural Networks of multiple Perceptrons work). The network will take your input. How it's coded is irrelevant for this, let's say input is a string "apple". Your training set then is {(apple1,0), (apple2,0), (apple3,0), (orange1,1), (orange2,1).....}. Since you know the class beforehand, the network will either output 1 or 0 for the input "apple1". If it outputs 1, you perform (targetValue-actualValue) = (1-0) = 1. 1 in this case means that the network gives a wrong output. Compare this to the delta rule and you will understand that this small equation is part of the larger update equation. In case you get a 1 you will perform a weight update. If target and actual value are the same, you will always get a 0 and you know that the network didn't misclassify.
How do I go about choosing test points, training points, threshold or
a bias?
Practically the bias and threshold isn't "chosen" per se. The bias is trained like any other unit using a simple "trick", namely using the bias as an additional input unit with value 1 - this means the actual bias value is encoded in this additional unit's weight and the algorithm we use will make sure it learns the bias for us automatically.
Depending on your activation function, the threshold is predetermined. For a simple perceptron, the classification will occur as follows:
Since we use a binary output (between 0 and 1), it's a good start to put the threshold at 0.5 since that's exactly the middle of the range [0,1].
Now to your last question about choosing training and test points: This is quite difficult, you do that by experience. Where you're at, you start off by implementing simple logical functions like AND, OR, XOR etc. There's it's trivial. You put everything in your training set and test with the same values as your training set (since for x XOR y etc. there are only 4 possible inputs 00, 10, 01, 11). For complex data like images, audio etc. you'll have to try and tweak your data and features until you feel like the network can work with it as good as you want it to.
What do I use the test data for and how does that relate to the training data?
Usually, to asses how well a particular algorithm performs, one first trains it and then uses different data to test how well it does on data it has never seen before.
How do I know if a point is misclassified?
Your training data has labels, which means that for each point in the training set, you know what class it belongs to.
How do I go about choosing test points, training points, threshold or a bias?
For simple problems, you usually take all the training data and split it around 80/20. You train on the 80% and test against the remaining 20%.
Before I ask my question, here's a brief summary of my project:
I'm using OPENCV's built-in function to detect a face in a
cam-feed.
After that I'm processing the image which contains the face, i.e.
converting it to grayscale, resizing it to 40X40 pixels and
equalizing its' histogram.
The pixel values of the image are then read, normalized (i.e.
divided by 256 since FANN works with values between 0 and 1, or -1
and 1, depending on the used function) and saved into an array of
1600 elements. This is the data which neural network works with.
Depending on the data, ANN decides either if the face is mine,
unknown, or not a face at all (false positive).
Neural network then returns an array of 3 elements, and the
program decides which group the face belongs to by finding the
maximum.
The thing is, apart from the pretty good detection of false positives, my code gives rather inaccurate results.
Some details on my ANN. I'm using the doublefann.h. The network contains 1600 input neurons (obviously), 3 output neurons (even more so), while the single hidden layer contains 1600 neurons, although I did try out other values in the 800-2400 range. I am using 20 samples of my face, 30 samples of unknown faces and 30 samples of random backgrounds for training. I tried both the RPROP (default in FANN, seems to overfit most of the time) and QUICKPROP (gives nice, smoothly decreasing error while training, but the results are inaccurate) training algorithms, as well as the SIGMOID_SYMMETRIC (training is done very fast, but often ends in abrupt fall of MSE to a near-zero value, and the resulting network is over-fit) activation function. I used both MSE and bit fail value as the criteria for stopping. Sadly, any given combination of those resulted in a rather poor face recognition.
So my questions would be, given the nature of my project and data which it handles, what would be the optimal:
Number of hidden neurons
Activation function
Training algorithm
Number of training samples per group (I was said that 20-30 should suffice)
Stopping criteria
Of course, any advice is welcome, it does not need to answer all or any of those five questions. I know that my problem is a pretty complex one, but I spent a good chunk of time running in circles and got tired of reading literature on the subject, and some first-hand knowledge would be highly appreciated.
iOS has an issue recording through some USB audio devices. It cannot be reliably reproduced (happens every 1 in ~2000-3000 records in batches and silently disappears), and we currently manually check our audio for any recording issues. It results in small numbers of samples (1-20) being shifted by a small number that sounds like a sort of 'crackle'.
They look like this:
closer:
closer:
another, single sample error elsewhere in the same audio file:
The question is, how can these be algorithmically be detected (assuming direct access to samples) whilst not triggering false positives on high frequency audio with waveforms like this:
Bonus points: after determining as many errors as possible, how can the audio be 'fixed'?
Dirty audio file - pictured
Another dirty audio file
Clean audio with valid high frequency - pictured
More bonus points: what could be causing this issue in the iOS USB audio drivers/hardware (assuming it is there).
I do not think there is an out of the box solution to find the disturbances, but here is one (non standard) way of tackling the problem. Using this, I could find most intervals and I only got a small number of false positives, but the algorithm could certainly use some fine tuning.
My idea is to find the start and end point of the deviating samples. The first step should be to make these points stand out more clearly. This can be done by taking the logarithm of the data and taking the differences between consecutive values.
In MATLAB I load the data (in this example I use dirty-sample-other.wav)
y1 = wavread('dirty-sample-pictured.wav');
y2 = wavread('dirty-sample-other.wav');
y3 = wavread('clean-highfreq.wav');
data = y2;
and use the following code:
logdata = log(1+data);
difflogdata = diff(logdata);
So instead of this plot of the original data:
we get:
where the intervals we are looking for stand out as a positive and negative spike. For example zooming in on the largest positive value in the plot of logarithm differences we get the following two figures. One for the original data:
and one for the difference of logarithms:
This plot could help with finding the areas manually but ideally we want to find them using an algorithm. The way I did this was to take a moving window of size 6, computing the mean value of the window (of all points except the minimum value), and compare this to the maximum value. If the maximum point is the only point that is above the mean value and at least twice as large as the mean it is counted as a positive extreme value.
I then used a threshold of counts, at least half of the windows moving over the value should detect it as an extreme value in order for it to be accepted.
Multiplying all points with (-1) this algorithm is then run again to detect the minimum values.
Marking the positive extremes with "o" and negative extremes with "*" we get the following two plots. One for the differences of logarithms:
and one for the original data:
Zooming in on the left part of the figure showing the logarithmic differences we can see that most extreme values are found:
It seems like most intervals are found and there are only a small number of false positives. For example running the algorithm on 'clean-highfreq.wav' I only find one positive and one negative extreme value.
Single values that are falsely classified as extreme values could perhaps be weeded out by matching start and end-points. And if you want to replace the lost data you could use some kind of interpolation using the surrounding data-points, perhaps even a linear interpolation will be good enough.
Here is the MATLAB-code I used:
function test20()
clc
clear all
y1 = wavread('dirty-sample-pictured.wav');
y2 = wavread('dirty-sample-other.wav');
y3 = wavread('clean-highfreq.wav');
data = y2;
logdata = log(1+data);
difflogdata = diff(logdata);
figure,plot(data),hold on,plot(data,'.')
figure,plot(difflogdata),hold on,plot(difflogdata,'.')
figure,plot(data),hold on,plot(data,'.'),xlim([68000,68200])
figure,plot(difflogdata),hold on,plot(difflogdata,'.'),xlim([68000,68200])
k = 6;
myData = difflogdata;
myPoints = findPoints(myData,k);
myData2 = -difflogdata;
myPoints2 = findPoints(myData2,k);
figure
plotterFunction(difflogdata,myPoints>=k,'or')
hold on
plotterFunction(difflogdata,myPoints2>=k,'*r')
figure
plotterFunction(data,myPoints>=k,'or')
hold on
plotterFunction(data,myPoints2>=k,'*r')
end
function myPoints = findPoints(myData,k)
iterationVector = k+1:length(myData);
myPoints = zeros(size(myData));
for i = iterationVector
subVector = myData(i-k:i);
meanSubVector = mean(subVector(subVector>min(subVector)));
[maxSubVector, maxIndex] = max(subVector);
if (sum(subVector>meanSubVector) == 1 && maxSubVector>2*meanSubVector)
myPoints(i-k-1+maxIndex) = myPoints(i-k-1+maxIndex) +1;
end
end
end
function plotterFunction(allPoints,extremeIndices,markerType)
extremePoints = NaN(size(allPoints));
extremePoints(extremeIndices) = allPoints(extremeIndices);
plot(extremePoints,markerType,'MarkerSize',15),
hold on
plot(allPoints,'.')
plot(allPoints)
end
Edit - comments on recovering the original data
Here is a slightly zoomed out view of figure three above: (the disturbance is between 6.8 and 6.82)
When I examine the values, your theory about the data being mirrored to negative values does not seem to fit the pattern exactly. But in any case, my thought about just removing the differences is certainly not correct. Since the surrounding points do not seem to be altered by the disturbance, I would probably go back to the original idea of not trusting the points within the affected region and instead using some sort of interpolation using the surrounding data. It seems like a simple linear interpolation would be a quite good approximation in most cases.
To answer the question of why it happens -
A USB audio device and host are not clock synchronous - that is to say that the host cannot accurately recover the relationship between the host's local clock and the word-clock of the ADC/DAC on the audio interface. Various techniques do exist for clock-recovery with various degrees of effectiveness. To add to the problem, the bus clock is likely to be unrelated to either of the two audio clocks.
Whilst you might imagine this not to be too much of a concern for audio receive - audio capture callbacks could happen when there is data - audio interfaces are usually bi-directional and the host will be rendering audio at regular interval, which the other end is potentially consuming at a slightly different rate.
In-between are several sets of buffers, which can over- or under-run, which is what looks to be happening here; the interval between it happening certainly seems about right.
You might find that changing USB audio device to one built around a different chip-set (or, simply a different local oscillator) helps.
As an aside both IEEE1394 audio and MPEG transport streams have the same clock recovery requirement. Both of them solve the problem with by embedding a local clock reference packet into the serial bitstream in a very predictable way which allows accurate clock recovery on the other end.
I think the following algorithm can be applied to samples in order to determine a potential false positive:
First, scan for high amount of high frequency, either via FFT'ing the sound block by block (256 values maybe), or by counting the consecutive samples above and below zero. The latter should keep track of maximum consecutive above zero, maximum consecutive below zero, the amount of small transitions around zero and the current volume of the block (0..1 as Audacity displays it). Then, if the maximum consecutive is below 5 (sampling at 44100, and zeroes be consecutive, while outstsanding samples are single, 5 responds to 4410Hz frequency, which is pretty high), or the sum of small transitions' lengths is above a certain value depending on maximum consecutive (I believe the first approximation would be 3*5*block size/distance between two maximums, which roughly equates to period of the loudest FFT frequency. Also it should be measured both above and below threshold, as we can end up with an erroneous peak, which will likely be detected by difference between main tempo measured on below-zero or above-zero maximums, also by std-dev of peaks. If high frequency is dominant, this block is eligible only for zero-value testing, and a special means to repair the data will be needed. If high frequency is significant, that is, there is a dominant low frequency detected, we can search for peaks bigger than 3.0*high frequency volume, as well as abnormal zeroes in this block.
Also, your gaps seem to be either highly extending or plain zero, with high extends to be single errors, and zero errors range from 1-20. So, if there is a zero range with values under 0.02 absolute value, which is directly surrounded by values of 0.15 (a variable to be finetuned) or higher absolute value AND of the same sign, count this point as an error. Single values that stand out can be detected if you calculate 2.0*(current sample)-(previous sample)-(next sample) and if it's above a certain threshold (0.1+high frequency volume, or 3.0*high frequency volume, whichever is bigger), count this as an error and average.
What to do with zero gaps found - we can copy values from 1 period backwards and 1 period forwards (averaging), where "period" is of the most significant frequency of the FFT of the block. If the "period" is smaller than the gap (say we've detected a gap of zeroes in a high-pitched part of the sound), use two or more periods, so the source data will all be valid (in this case, no averaging can be done, as it's possible that the signal 2 periods forward from the gap and 2 periods back will be in counterphase). If there are more than one frequency of about equal amplitude, we can plain sample these with correct phases, cutting the rest of less significant frequencies altogether.
The outstanding sample should IMO just be averaged by 2-4 surrounding samples, as there seems to be only a single sample ever encountered in your sound files.
The discrete wavelet transform (DWT) may be the solution to your problem.
A FFT calculation is not very useful in your case since its an average representation of relative frequency content over the entire duration of the signal, and thus impossible to detect momentary changes. The dicrete short time frequency transform (STFT) tries to tackle this by computing the DFT for short consecutive time-blocks of the signal, the length of which is determine by the length (and shape) of a window, but since the resolution of the DFT is dependent on the data/block-length, there is a trade-off between resolution in freqency OR in time, and finding this magical fixed window-size can be tricky!
What you want is a time-frequency analysis method with good time resolution for high-frequency events, and good frequency resolution for low-frequency events... Enter the discrete wavelet transform!
There are numerous wavelet transforms for different applications and as you might expect, it's computationally heavy. The DWT may not be practical solution to your problem, but it's worth considering. Good luck with your problem. Some friday-evening reading:
http://klapetek.cz/wdwt.html
http://etd.lib.fsu.edu/theses/available/etd-11242003-185039/unrestricted/09_ds_chapter2.pdf
http://en.wikipedia.org/wiki/Wavelet_transform
http://en.wikipedia.org/wiki/Discrete_wavelet_transform
You can try the following super-simple approach (maybe it's enough):
Take each point in your wave-form and subtract its predecessor (look at the changes from one point to the next).
Look at the distribution of these changes and find their standard deviation.
If any given difference is beyond X times this standard deviation (either above or below), flag it as a problem.
Determine the best value for X by playing with it and seeing how well it performs.
Most "problems" should come as a pair of two differences beyond your cutoff, one going up, and one going back down.
To stick with the super-simple approach, you can then fix the data by just interpolating linearly between the last good point before your problem-section and the first good point after. (Make sure you don't just delete the points as this will influence (raise) the pitch of your audio.)
My problem is as follows. I have 6 dataset (collected from different days) that I want to build a decision tree classifier for. I need to find the optimal tree depth parameter to avoid over fitting and have good prediction for new data.
My first attempt was to split each dataset by 50%/25%/25% into training/validation/testing, and merged them into three big files. I produced a graph of accuracy vs tree depth, trained using the big training set and calculated the accuracy from the big validation set. However, I noticed the accuracy of the validation set kept increasing as the tree depth increased and never went down!
The accuracy reached around 99%, at a tree depth of 20. Even the test set's accuracy was around 99%. I assumed it had to be over fitting, based on my knowledge of the data. I confirmed this by removing one of the dataset, leaving only 5. I trained on the remaining 5 dataset using the same procedure as above, splitting into training/validation/testing and training it at a depth of 20. This time I tested it against the dataset I left out and it produced a very poor result, accuracy in the low 50%.
It turns out my dataset has a lot of 'repetitive' data due to the time-series nature of it. To the point that sampling 25-50% any of the dataset is probably equivalent to sampling 100%, hence the over fitting. I decided to try a different strategy by randomly shuffling the 6 dataset and splitting them into training/validation via 3/3, repeat it 20 times and average the results to find the optimal tree depth. I'm still waiting for the result, should be done in a few days ...
So my question (finally!) is whether what I did is a sound method? What other kind of funny data should I be aware of? It seems like my dataset has a 'gotcha' that I have seen mentioned anywhere.
Turns out the problem was quite subtle. Long story short, DON'T up sample your data (for unbalanced dataset) and do cross-validation!
Long story. I had an unbalanced dataset where I had more negative samples than positive by about 7-10 times. My choice was either to up sample the positive via duplication or down sample the negative. I went with up sampling,. This means during cross validation, each fold has a very high chance of picking all of the original positive samples and thus overfit the positive samples. I'm now down sampling the negative samples, doing the cross validation a few times and shuffling the data. Seems to work now.