Rune Allnor wrote: ...> All signal analysis methods -- all -- depend on the user specifying > in advance what "basis set" or "signal model" to evaluate the data > against, as well as a strategy for the implementation (neural net, > DFT, something else). The OP is asking whether these choices of > signal models and analysis strategies can be delegated to the > computer. > > They can not.Not yet, anyhow. :-) Jerry -- Engineering is the art of making what you want from things you can get. �����������������������������������������������������������������������
Online · Live + on demand
February 9-11, 2027
Price goes up Jan 1: register now to lock in today's rate
Register Now →
the "wavyness" question re-formulated
Started by ●November 24, 2006
Reply by ●November 26, 20062006-11-26
Reply by ●November 27, 20062006-11-27
Jerry Avins skrev:> Rune Allnor wrote: > > ... > > > All signal analysis methods -- all -- depend on the user specifying > > in advance what "basis set" or "signal model" to evaluate the data > > against, as well as a strategy for the implementation (neural net, > > DFT, something else). The OP is asking whether these choices of > > signal models and analysis strategies can be delegated to the > > computer. > > > > They can not. > > Not yet, anyhow. :-)Well, it would take a computer that can emulate the human brain to do this type of thing. The computers will not take over these sorts of jobs for acouple of months yet. Present-day computers excel at all the drudgerous repetitive tasks that drive a human being mad from boredom. Humans thrive when they can use their past exterience mixed with a bit of imagination to solve puzzels. Deciding what analysis strategy fits a particular data set and purpose, take these very skills: Experience and imagination. Of course, these factors occur in different proportions. There is little imagination involved in standard processing flows like, say, Air Traffic Control radar, where everything has to fit whatever codes are set down by governments. In other applications, like exploration seismics, one really have to use all skills and experience to get anything from the data. When one face new processing tasks, the human operator has to gain the necessary experience befor applying it imaginatively to the data. The only effective way I know of to gain that experience is to implement the processing methods, and then test them to destruction with synthetic data. After having done that, one can see features or properties in experimental data that might suggest this method or exclude that method. I used to work with a major oil company. They decided to test how various contractors performed on seismic processing, so they sent the same data set to eight or ten contractors for processing, and presented all the results side by side. If one did not know that all the processed images were based on the same data, one would never have suspected they were. That test caused quite a stir among both processors and interpreters... Rune
Reply by ●November 27, 20062006-11-27
Rune Allnor wrote:> Fred Marshall skrev: > >><lluum@yahoo.com> wrote in message >> >>>>Question 1. Is there a systematic approach (or perspectives) to >>> >>>classify signals. >> >>***There are. I believe all or most of them, including using neural nets, >>depend on using a basis set or signal model or .... something like that. > > > All signal analysis methods -- all -- depend on the user specifying > in advance what "basis set" or "signal model" to evaluate the data > against, as well as a strategy for the implementation (neural net, > DFT, something else). The OP is asking whether these choises of > signal models and analysis strategies can be delegated to the > computer. > > They can not. >Another way of looking at this is that when you analyze data like this you are throwing away information -- after all, you're starting with a vector of 1024 numbers that are essentially continuous, and you want to get one or two diagnostic criteria out of them. When you're done there's 1022 numbers lying in the trash that you didn't care about. Figuring out what numbers to keep and what ones to throw out is pretty important, in this case. -- Tim Wescott Wescott Design Services http://www.wescottdesign.com Posting from Google? See http://cfaj.freeshell.org/google/ "Applied Control Theory for Embedded Systems" came out in April. See details at http://www.wescottdesign.com/actfes/actfes.html
Reply by ●November 27, 20062006-11-27
"Rune Allnor" <allnor@tele.ntnu.no> wrote in message news:1164593372.770627.308170@14g2000cws.googlegroups.com...> > Fred Marshall skrev: >> <lluum@yahoo.com> wrote in message >>> > Question 1. Is there a systematic approach (or perspectives) to >> > classify signals. >> >> ***There are. I believe all or most of them, including using neural >> nets, >> depend on using a basis set or signal model or .... something like that. > > All signal analysis methods -- all -- depend on the user specifying > in advance what "basis set" or "signal model" to evaluate the data > against, as well as a strategy for the implementation (neural net, > DFT, something else). The OP is asking whether these choises of > signal models and analysis strategies can be delegated to the > computer. > > They can not. > > RuneRune, Well not to be *too* picky but I hedged my comment because the question is "classification" and not, specifically, "analysis". I suppose one could debate whether classification is possible without analysis and I'd not suggest such a departure here. I could only imagine someone saying that classification without a basis set is possible...... so I hedged my response. Or, to put it another way, I was being (too?) polite. I think we agree - certainly about the computer comment. I once pressed a group of engineers to explain how they were going to do a fairly sophisticated acoustic classification in a new version of their gadget and they said "well, there will be a computer in there...." Truly! Fred
Reply by ●November 27, 20062006-11-27
Fred Marshall skrev:> "Rune Allnor" <allnor@tele.ntnu.no> wrote in message > news:1164593372.770627.308170@14g2000cws.googlegroups.com... > > > > Fred Marshall skrev: > >> <lluum@yahoo.com> wrote in message > >>> > Question 1. Is there a systematic approach (or perspectives) to > >> > classify signals. > >> > >> ***There are. I believe all or most of them, including using neural > >> nets, > >> depend on using a basis set or signal model or .... something like that. > > > > All signal analysis methods -- all -- depend on the user specifying > > in advance what "basis set" or "signal model" to evaluate the data > > against, as well as a strategy for the implementation (neural net, > > DFT, something else). The OP is asking whether these choises of > > signal models and analysis strategies can be delegated to the > > computer. > > > > They can not. > > > > Rune > > Rune, > > Well not to be *too* picky but I hedged my comment because the question is > "classification" and not, specifically, "analysis". > > I suppose one could debate whether classification is possible without > analysis and I'd not suggest such a departure here.As far as I am concerned, "classification" is the result of a preliminary "analysis".> I could only imagine someone saying that classification without a basis set > is possible...... so I hedged my response. Or, to put it another way, I was > being (too?) polite. > > I think we agree - certainly about the computer comment. I once pressed a > group of engineers to explain how they were going to do a fairly > sophisticated acoustic classification in a new version of their gadget and > they said "well, there will be a computer in there...." Truly!Been there, done that... sort of. Once upon a time I was at the receiving end of a similar request. Somebody asked me to see what was possible if a certain constraint was "known" to be true. I refused to discuss the issue for two reasons: First, the constraint could easily be invalidated, either by chance or on purpose; it was basically nothing more than whishful thinking. Second, certain data had been observed a short time previosly that made no sense if the constaint was assumed to met, but that could easily be explained if the constraint for some reason should be broken. Of course, it is impossible to even dicuss that sort of question without suggesting to senior staff, both in-house and with the client, that they had no clue about how the systems worked. Not a very pleasant work environment, on other words, so I left shortly afterwards. Rune
Reply by ●November 28, 20062006-11-28
Fred Marshall wrote:> "Rune Allnor" <allnor@tele.ntnu.no> wrote in message > news:1164593372.770627.308170@14g2000cws.googlegroups.com... > > > > Fred Marshall skrev: > >> <lluum@yahoo.com> wrote in message > >>> > Question 1. Is there a systematic approach (or perspectives) to > >> > classify signals. > >> > >> ***There are. I believe all or most of them, including using neural > >> nets, > >> depend on using a basis set or signal model or .... something like that. > > > > All signal analysis methods -- all -- depend on the user specifying > > in advance what "basis set" or "signal model" to evaluate the data > > against, as well as a strategy for the implementation (neural net, > > DFT, something else). The OP is asking whether these choises of > > signal models and analysis strategies can be delegated to the > > computer. > > > > They can not. > > > > Rune > > Rune, > > Well not to be *too* picky but I hedged my comment because the question is > "classification" and not, specifically, "analysis". > > I suppose one could debate whether classification is possible without > analysis and I'd not suggest such a departure here. > > I could only imagine someone saying that classification without a basis set > is possible...... so I hedged my response. Or, to put it another way, I was > being (too?) polite.It is not clear what you mean by a basis set. I assume you mean a set of functions that serve as bases in function space. If so, then they are not necessary. Classifiers are, typically, constructed using a design set which, in general, consists of a labeled training set used to estimate classifier parameters and a labeled validation set which is used to select one of many classifier design candidates. The classifier is typically evaluated using a test set which does not contain design set members. Once you have decided on a representative design set, you should definitely take time to get familiar with the data via scatter plots and unsupervised learning in the form of clustering and Principal Coordinate Analysis. This should give some insight as to which classes will be hard to separate and which variables may be best or worst in effecting class separation. If you can label the design set by eye, you probably don't need to resort to spectral representation via the fft. On the other hand, if you prefer to use the fft, my suggestion is to use both. Trying to design a classifier without some form of predesign analysis is definitely not recommended. Hope this helps. Greg> I think we agree - certainly about the computer comment. I once pressed a > group of engineers to explain how they were going to do a fairly > sophisticated acoustic classification in a new version of their gadget and > they said "well, there will be a computer in there...." Truly! > > Fred
Reply by ●November 29, 20062006-11-29
"Greg Heath" <heath@alumni.brown.edu> wrote in message news:1164732285.180641.82460@n67g2000cwd.googlegroups.com...> > Fred Marshall wrote: >> "Rune Allnor" <allnor@tele.ntnu.no> wrote in message >> news:1164593372.770627.308170@14g2000cws.googlegroups.com... >> > >> > Fred Marshall skrev: >> >> <lluum@yahoo.com> wrote in message >> >>> > Question 1. Is there a systematic approach (or perspectives) to >> >> > classify signals. >> >> >> >> ***There are. I believe all or most of them, including using neural >> >> nets, >> >> depend on using a basis set or signal model or .... something like >> >> that. >> > >> > All signal analysis methods -- all -- depend on the user specifying >> > in advance what "basis set" or "signal model" to evaluate the data >> > against, as well as a strategy for the implementation (neural net, >> > DFT, something else). The OP is asking whether these choises of >> > signal models and analysis strategies can be delegated to the >> > computer. >> > >> > They can not. >> > >> > Rune >> >> Rune, >> >> Well not to be *too* picky but I hedged my comment because the question >> is >> "classification" and not, specifically, "analysis". >> >> I suppose one could debate whether classification is possible without >> analysis and I'd not suggest such a departure here. >> >> I could only imagine someone saying that classification without a basis >> set >> is possible...... so I hedged my response. Or, to put it another way, I >> was >> being (too?) polite. > > It is not clear what you mean by a basis set. I assume you > mean a set of functions that serve as bases in function > space. If so, then they are not necessary. > > Classifiers are, typically, constructed using a design set > which, in general, consists of a labeled training set > used to estimate classifier parameters and a labeled > validation set which is used to select one of many > classifier design candidates. The classifier is typically > evaluated using a test set which does not contain > design set members. > > Once you have decided on a representative design set, > you should definitely take time to get familiar with the data > via scatter plots and unsupervised learning in the form of > clustering and Principal Coordinate Analysis. > > This should give some insight as to which classes will > be hard to separate and which variables may be best > or worst in effecting class separation. > > If you can label the design set by eye, you probably > don't need to resort to spectral representation via the > fft. > > On the other hand, if you prefer to use the fft, my > suggestion is to use both. > > Trying to design a classifier without some form of > predesign analysis is definitely not recommended. > > Hope this helps. > > Greg >Greg, Sounds like neural net lingo to me. Is it? I'm no expert in that area. It's not clear what you mean by "design set", "labeled training set", "classifier parameters", and "validation set". Well, I can guess but .... Fred
Reply by ●November 29, 20062006-11-29
Hi, Here is my take on this...hth lluum@yahoo.com wrote:> 1. Some signals have drift -- maybe a constant drift. Some signals > may have zero drift > > 2. Signals might be classified into the following cateories (just by > inspection in the time domain) > Category (1) Almost a straight line with some noise. > Category (2) Brownian motions (random walk) overlaid with a drift. > Category (3) An overlay of harmonic waves of several frequency with a > zero or nonzero drift.> (1A) For instance, if I use the Fourier perspective, what are the > characteristics in the fourier domain for Category(1) and Category(2)For Category(1) I assume you are talking of white noise. If you subtract the mean of it you will get a straight line in the Magnitude Spectrum. For Brownian Motion (random walk) you will get brown noise, which is characterized in Magnitude Spectrum by 1/(f^2) times a proportion The factor. The drift in the brown motion would add some power to the low components but the frequency should still drop as you go to higher numbers (not the case for white noise)... These two should not be hard to differentiate....in the magnitude spectrum domain.> (1B) If I use the Random walk perspective, is there any characteristics > of Random walk that can differentiate Category(1) (2) and (3)See above for (1) and (2). For (3) you can notice than it's spectrum is non-monotic (as oppose to (1) and (2)), so a test to check if the magnitude spectrum is non-monotic (specially and mid to high frequencies) should work as long as you have harmonics in those location.> Question 2. How do I fit a random walk signal to extract its drift and > the std deviation of the normally distributed increments.Because the increments are normaly distributed, you can use the pdf of a standard Wiener process with drift coeff. u: f(x)= (1/sqrt(2*pi*t)) * exp (- (x-u*t)^2/(2*t) Where t is the time (which is known), You can then use maximum likilehod mehod to estimate u from your data.. The estiamte of the std I think can be approximated by: std=sqrt( t*(fs^2) + u^2)> Question 3. How do I extract drift if the signal is an overlay of > several harmonics with a constant driftIf you take the LOG of the signal than these components should be separable (if they are multiplied together). So you can implement the steps mentioned above using the cepstrum (ie, liftering) to get rid of the harmonics. I hope this helps. But in the end, like others have suggested, I am sure you will need to tune and twicker to fine tune whatever classification you want. I would not use neural networks, at least for the firtst steps of the feature extraction that you have asked above, since it would require alot of data for training and would add an extra degree of uncertainty (aka, hard to debug) to the process. -Ikaro
Reply by ●November 30, 20062006-11-30
Fred Marshall wrote:> "Greg Heath" <heath@alumni.brown.edu> wrote in message > news:1164732285.180641.82460@n67g2000cwd.googlegroups.com... > > > > Fred Marshall wrote: > >> "Rune Allnor" <allnor@tele.ntnu.no> wrote in message > >> news:1164593372.770627.308170@14g2000cws.googlegroups.com... > >> > > >> > Fred Marshall skrev: > >> >> <lluum@yahoo.com> wrote in message > >> >>> > Question 1. Is there a systematic approach (or perspectives) to > >> >> > classify signals. > >> >> > >> >> ***There are. I believe all or most of them, including using neural > >> >> nets, > >> >> depend on using a basis set or signal model or .... something like > >> >> that. > >> > > >> > All signal analysis methods -- all -- depend on the user specifying > >> > in advance what "basis set" or "signal model" to evaluate the data > >> > against, as well as a strategy for the implementation (neural net, > >> > DFT, something else). The OP is asking whether these choises of > >> > signal models and analysis strategies can be delegated to the > >> > computer. > >> > > >> > They can not. > >> > > >> > Rune > >> > >> Rune, > >> > >> Well not to be *too* picky but I hedged my comment because the question > >> is > >> "classification" and not, specifically, "analysis". > >> > >> I suppose one could debate whether classification is possible without > >> analysis and I'd not suggest such a departure here. > >> > >> I could only imagine someone saying that classification without a basis > >> set > >> is possible...... so I hedged my response. Or, to put it another way, I > >> was > >> being (too?) polite. > > > > It is not clear what you mean by a basis set. I assume you > > mean a set of functions that serve as bases in function > > space. If so, then they are not necessary. > > > > Classifiers are, typically, constructed using a design set > > which, in general, consists of a labeled training set > > used to estimate classifier parameters and a labeled > > validation set which is used to select one of many > > classifier design candidates. The classifier is typically > > evaluated using a test set which does not contain > > design set members. > > > > Once you have decided on a representative design set, > > you should definitely take time to get familiar with the data > > via scatter plots and unsupervised learning in the form of > > clustering and Principal Coordinate Analysis. > > > > This should give some insight as to which classes will > > be hard to separate and which variables may be best > > or worst in effecting class separation. > > > > If you can label the design set by eye, you probably > > don't need to resort to spectral representation via the > > fft. > > > > On the other hand, if you prefer to use the fft, my > > suggestion is to use both. > > > > Trying to design a classifier without some form of > > predesign analysis is definitely not recommended. > > > > Hope this helps. > > > > Greg > > > > Greg, > > Sounds like neural net lingo to me. Is it? I'm no expert in that area. > > It's not clear what you mean by "design set", "labeled training set", > "classifier parameters", and "validation set". Well, I can guess but ....I confess. I am a neural-netter. Typically nonlinear NN models designs are designed and tested using labelled data, i.e., both input and output are known. This is called supervised learning as opposed to unsupervised learning from unlabelled data (e.g., clustering or PCA). In order to avoid estimation biases that can occur when testing models by reusing design data, the following split-sample technique (very well explained in the comp.ai.neural-net FAQ and archived threads) is used: Split the labelled data into design and test sets (N = Ndes+Ntst). Further split the design set into training and validation sets (Ndes = Ntrn+Nval). The training set is used to estimate classifier parameters (i.e, weights and thresholds). In the simple regression y = b1*x + b0, b1 is a weight and b0 is a threshold. Ntrn should be large enough to obtain adequate estimates. The validation set is used to estimate the MSE of competing designs (e,g, linear, quadratic and cubic polynomials). Nval should be large enough to obtain adequate estimates of MSE. Once a design is chosen, the test set is used to estimate the expected value of MSE for unseen data. Ntst usually satisfies Ntst >= Nval since it is not necessary to obtain the same precision in estimation when just choosing between models. Usually, rule of thumb estimates of minimum values for Ntrn, Nval and Ntst are obtained. If the sum exceeds N, then additional and/or alternate approaches like input variable reduction, weight reduction or elimination, f-fold cross-validation and bootstrapping are considered. Hope this helps. Greg
Reply by ●November 30, 20062006-11-30
Fred Marshall wrote:> "Greg Heath" <heath@alumni.brown.edu> wrote in message > news:1164732285.180641.82460@n67g2000cwd.googlegroups.com... > > > > Fred Marshall wrote: > >> "Rune Allnor" <allnor@tele.ntnu.no> wrote in message > >> news:1164593372.770627.308170@14g2000cws.googlegroups.com... > >> > > >> > Fred Marshall skrev: > >> >> <lluum@yahoo.com> wrote in message > >> >>> > Question 1. Is there a systematic approach (or perspectives) to > >> >> > classify signals. > >> >> > >> >> ***There are. I believe all or most of them, including using neural > >> >> nets, > >> >> depend on using a basis set or signal model or .... something like > >> >> that. > >> > > >> > All signal analysis methods -- all -- depend on the user specifying > >> > in advance what "basis set" or "signal model" to evaluate the data > >> > against, as well as a strategy for the implementation (neural net, > >> > DFT, something else). The OP is asking whether these choises of > >> > signal models and analysis strategies can be delegated to the > >> > computer. > >> > > >> > They can not. > >> > > >> > Rune > >> > >> Rune, > >> > >> Well not to be *too* picky but I hedged my comment because the question > >> is > >> "classification" and not, specifically, "analysis". > >> > >> I suppose one could debate whether classification is possible without > >> analysis and I'd not suggest such a departure here. > >> > >> I could only imagine someone saying that classification without a basis > >> set > >> is possible...... so I hedged my response. Or, to put it another way, I > >> was > >> being (too?) polite. > > > > It is not clear what you mean by a basis set. I assume you > > mean a set of functions that serve as bases in function > > space. If so, then they are not necessary. > > > > Classifiers are, typically, constructed using a design set > > which, in general, consists of a labeled training set > > used to estimate classifier parameters and a labeled > > validation set which is used to select one of many > > classifier design candidates. The classifier is typically > > evaluated using a test set which does not contain > > design set members. > > > > Once you have decided on a representative design set, > > you should definitely take time to get familiar with the data > > via scatter plots and unsupervised learning in the form of > > clustering and Principal Coordinate Analysis. > > > > This should give some insight as to which classes will > > be hard to separate and which variables may be best > > or worst in effecting class separation. > > > > If you can label the design set by eye, you probably > > don't need to resort to spectral representation via the > > fft. > > > > On the other hand, if you prefer to use the fft, my > > suggestion is to use both. > > > > Trying to design a classifier without some form of > > predesign analysis is definitely not recommended. > > > > Hope this helps. > > > > Greg > > > > Greg, > > Sounds like neural net lingo to me. Is it? I'm no expert in that area. > > It's not clear what you mean by "design set", "labeled training set", > "classifier parameters", and "validation set". Well, I can guess but ....I confess. I am a neural-netter. Typically nonlinear NN models designs are designed and tested using labelled data, i.e., both input and output are known. This is called supervised learning as opposed to unsupervised learning from unlabelled data (e.g., clustering or PCA). In order to avoid estimation biases that can occur when testing models by reusing design data, the following split-sample technique (very well explained in the comp.ai.neural-net FAQ and archived threads) is used: Split the labelled data into design and test sets (N = Ndes+Ntst). Further split the design set into training and validation sets (Ndes = Ntrn+Nval). The training set is used to estimate classifier parameters (i.e, weights and thresholds). In the simple regression y = b1*x + b0, b1 is a weight and b0 is a threshold. Ntrn should be large enough to obtain adequate estimates. The validation set is used to estimate the MSE of competing designs (e,g, linear, quadratic and cubic polynomials). Nval should be large enough to obtain adequate estimates of MSE. Once a design is chosen, the test set is used to estimate the expected value of MSE for unseen data. Ntst usually satisfies Ntst >= Nval since it is not necessary to obtain the same precision in estimation when just choosing between models. Usually, rule of thumb estimates of minimum values for Ntrn, Nval and Ntst are obtained. If the sum exceeds N, then additional and/or alternate approaches like input variable reduction, weight reduction or elimination, f-fold cross-validation and bootstrapping are considered. Hope this helps. Greg







