Hello Venkatesh,

1) If you are concentrating on 3-10Hz, you might require a BPF for that band, i.e. lower cut-off of 3 Hz and the higher cut-off of 10Hz.

2) Alternately, you can also think of combining two different filters such as 3Hz HPF and followed by 10Hz LPF.

3) You can have Blackman, Hamming, Hanning or Windowed type filter response and then your algorithm should be written accordingly.

4) Now the main question comes of the iterations, i.e. the speed, sampling interval etc. On similar grounds you can think of implementing DWT.

5) You have not stated which of the hardware platform you are working with Analog or Texas or some other for the DSP you are keen to use with.

Hope this stimulates your thinking!

Best wishes!

-Prasanna Waichal

Thank you :)

but my question is, how it will help me to get time instant of the particular frequency?

1) band pass filter is used of 3 -10Hz.

2)My doubt is for filter bank how to decide cut-off  frequency for each stage?

Dear Venkatesh

As i understand by your problem, you want to extract low frequency vibration from acquired signal. You want to decompose signal by reference paper "Implementation of Discrete Wavelet Transform for Embedded Applications using TMS320VC5510"

In refer paper, WT is implement by convolutional and lifting scheme. You follow the flow chart given in paper for that you required WT's low pass and high pass filters with QMF characteristic to perfect reconstruct. QMF filter's coefficients can be found by "imfilter" function in Matlab. Filter coefficients are not depends on frequency cut-off frequency. Cut-off frequency will depends on sampling frequency and number of decomposition levels. In reference to fig 2 from paper, the signal is decomposed in low pass and high pass in first level decomposition then low pass is again decomposed to low pass and high pass and continue upto last sample. The signal is having frequency from 0 to fs/2. (the minimum sampling frequency is fs by Nyquist sampling theorem). In first level decomposition signal is divided in 0 to fs/4 for low pass filter output and fs/4 to fs/2 fpr high pass filter. In second level low filter output is decomposed as as first level in 0 to fs/8 and fs/8 to fs/4 and soon.

So you decide your sampling frequency ( depends on sensor characteristics), then decomposed your signal upto desired frequency band. This will solve your problem.

Thank you Gupta sir,

i believe B = imfilter(A,h) this equation you are talking about.

1.Is Parameter 'h' sampling rate?

2.Is parameter 'B' will be coefficient?

if it is sampling rate for each stage then it will be get halved, so need to take coefficient separately for each stage .

Please provide you input.

what is the filter bank relationship with Daubechies ?

How to corelate filter banks cutoff of and all with Daubechies (1 to 20, DB1....20) can any one please tell me

now about implementation in matlab -

1. you can use -

[Lo_D,Hi_D,Lo_R,Hi_R] = wfilters(wname)

where wname is wavelet type - db1, db2, db3........

           and filters are analysis - lowpass & highpass and synthesis (reconstruction) - lowpass and high pass.

2. yes it will get half for DWT and same for WT (redundant information)

Thank you gupta sir,

Let's take a one example 

I want to use db5, and 5 decomposition level and my input signal is any vibration or speech 

1. [Lo_D,Hi_D,Lo_R,Hi_R] = wfilters(db5)

Does these are ( Lo_D,Hi_D,Lo_R,Hi_R ) first stages decomposition out put? 

Or Final out put ??

And how I can chose here 5 decomposition level?  Db5 input paramete it is ok..

If my understanding is wrong then please tell me whether this [Lo_D,Hi_D,Lo_R,Hi_R] = wfilters(wname) function uses filter bank concept or Not? 

this function will give you filter coefficients, now you use flow chart given in reference paper as convolution method.

Thank you sir :)

I will try this 

Hi Gupta sir,

I have tried with this [Lo_D,Hi_D,Lo_R,Hi_R] = wfilters(wname) . The results are exactly matching for haar wavelet. but i am facing problem for DB5 wavelet. initial few samples are not matching.

i followed below procedure.

1. [Lo_D,Hi_D,Lo_R,Hi_R] = wfilters('db5') ;

cA1 = conv(Lo_D,input);

downsample_cA1 = downsample(cA1,2);

cD1 = conv(Hi_D,input);

downsample_cD1 = downsample(cD1,2);

[CA1 CD1] = dwt(input,'db5');

now my problem is downsample_cA1 is not exactly matching with CA1, initial and last 4 samples are miss matching .

it is holds good for detailed coefficient also that is CD1.

please provide input.

It is obvious because of down sampling. It will remove redundant information from CA1 and CD1. The initial and final mismatch because of boundary conditioning (added sample at initial and final) to perform convolution operation. If you want to improve this then try other padding techniques.

all the best........

Hello Gupta sir,

I have tried all padding techniques. I have not observed any positive result. I got possitive result only for first stage with padding technique.

If i am making mistakes in padding please let me know.

Input = {a, b, c, d, e, f, g, h, i, j}

Padded input signal = {e, d, c, b, a, b, c, d, e, f, g, h, i, j, i, h, g, f}

Is this method correct ?

Please give me your inputs