Please help with arduino fft library

From the ATmega data sheet:

"By default, the successive approximation circuitry requires an input clock frequency between 50 kHz and 200 kHz to get maximum resolution. If a lower resolution than 10 bits is needed, the input clock frequency to the ADC can be higher than 200 kHz to get a higher sample rate."

"To get full resolution you need to run the ADC at 15,000 samples per second or slower."

"You can get 76,900 samples per second at reduced resolution."

In free-running mode it takes 13.5 cycles to per sample. The pre-scaler options are:

2 = 8 MHz = 592,592 samples per second (too fast)
4 = 4 MHz = 296,296 samples per second (too fast)
8 = 2 MHz = 148,148 samples per second (too fast)
16 = 1 MHz = 74,074 samples per second (low resolution)
32 = 500 KHz = 37,037 samples per second (reduced resolution)
64 = 250 KHz = 18,518 samples per second (reduced resolution)
128 = 125 KHz = 9,259 samples per second (full resolution)

Looks like your loose a couple of digits of accuracy if you run at 1 MHz (74,074 samples per second)
Absolute accuracy, ADC clock = 200 kHz: 2 LSB
Absolute accuracy, ADC clock = 1 MHz: 4.5 LSB

Magician:

Can you suggest one that has the juice to get the job done?
More memory, Mega - 8K 'd make 4x difference in resolution, compare to UNO.
I saw a tutorial before, that a girl was able to set the sampling rate to 38.5kHz. I was hoping that's the rate I will use.
Rate is not an obstacle, memory is.
When you say the resolution is too pure for instrument tunning, is that good or bad? 4Hz difference sounds very reasonable.
Do your research, I think it's about 1%, 100Hz note 'd requre 1Hz

Piano key frequencies - Wikipedia
http://coolarduino.wordpress.com/2011/09/03/arduino-musical-note-recognition-pushing-the-limits/

I see your point. accuracy is very important for piano tuning, maybe for all the musical tuning. From the wiki page, piano only produces frequencies from 27.5hz up 4186hz, which means I need to have at lease 8372hz sampling rate to do this type of tuning. I read that sometime it requires a little bit higher than twice of highest frequency to get the optimum result somewhere. But that's a lot lower than what I intended to set at.
According to john from reply #24, if I set pre-scaler = 128, I can have a sample rate of 9259 hz and full-resolution. The frequency is enough for piano tuning, right? but for resoolution, I hope full-resolution is at least 10bits, do I still need larger memory?

Thank you.

johnwasser:
From the ATmega data sheet:

"By default, the successive approximation circuitry requires an input clock frequency between 50 kHz and 200 kHz to get maximum resolution. If a lower resolution than 10 bits is needed, the input clock frequency to the ADC can be higher than 200 kHz to get a higher sample rate."

"To get full resolution you need to run the ADC at 15,000 samples per second or slower."

"You can get 76,900 samples per second at reduced resolution."

In free-running mode it takes 13.5 cycles to per sample. The pre-scaler options are:

2 = 8 MHz = 592,592 samples per second (too fast)
4 = 4 MHz = 296,296 samples per second (too fast)
8 = 2 MHz = 148,148 samples per second (too fast)
16 = 1 MHz = 74,074 samples per second (low resolution)
32 = 500 KHz = 37,037 samples per second (reduced resolution)
64 = 250 KHz = 18,518 samples per second (reduced resolution)
128 = 125 KHz = 9,259 samples per second (full resolution)

Looks like your loose a couple of digits of accuracy if you run at 1 MHz (74,074 samples per second)
Absolute accuracy, ADC clock = 200 kHz: 2 LSB
Absolute accuracy, ADC clock = 1 MHz: 4.5 LSB

Hey, john, thank you for the great explanation.
I found out that I don't need to go all the way up to 20kHz for musical tuning. For piano, I need to have at least 8372 hz sampling rate. According to your calculation, I can use pre-scaler 128 and get the full resolution. Magician from reply #18 pointed that out! which is a great news for me.

Could you also help me out on set the simulation sampling frequency rate, please?

"I have 256 samples = sampleCount; I have no idea what the simulated sampling rate is. How can I find out? I thought all I need was to filled up the array with numbers. Please help!
If I change " k = 127sin(23.14 * 100 * t ) " to "k = 127sin(23.14 * 100 * 256 * t )" would that make a difference?
I thought I understood a little about this stuff. Now I'm really confused."

Thank you

How did you find out free running sampling frequency is 9 kHz for UNO? Can you post a link here, please? I want to know where or how to find those infos. Also, I looked at arduino reference page, UNO has 16 MHz ceramic resonator. Why does it only sample at a max rate of 9kHz? Can you teach me how to figure it out?
Download datasheet AtMega328 from atmel.com look into ADC section, you will find all information
Why would you think sampling rate is not a problem? Even if I used UNO, it still isn't a problem or you meant that if I use mega, It won't be a problem anymore? It looks like mega has the same cpu speed as UNO, it does have a lot bigger memory. You think the memory is a problem, is it because I need to get better frequency resolutions?
Exactly, don't confuse ADC resolution (which is least of your worry, as low as 8-bit o'k), and FFT resolution, that defined as Freq. range / Bin numbers. Having 4 kHz sampling and / 512-fft = 8 Hz. Paradox is , you need to lower sampling rate, to get better resolution! (I mention this in musical note recognition post). Other parameter, increasing fft-size, isn't an option , except you have Due board.

The wikipedia article shows note frequencies with six significant digits. In the low notes that's a resolution of 0.0001 Hz. To get that resolution at 9,259 samples per second you would need an FFT size of 92.5 million samples (185 megabytes). That is not going to happen on an Arduino in any reasonable amount of time.

Perhaps you can use zero crossings to measure the frequency. Have you looked at any frequency counter sketches?

http://interface.khm.de/index.php/lab/experiments/arduino-frequency-counter-library/

Magician:
How did you find out free running sampling frequency is 9 kHz for UNO? Can you post a link here, please? I want to know where or how to find those infos. Also, I looked at arduino reference page, UNO has 16 MHz ceramic resonator. Why does it only sample at a max rate of 9kHz? Can you teach me how to figure it out?

Download datasheet AtMega328 from atmel.com look into ADC section, you will find all information

This is what I found in the datasheet:
"Using the ADC Interrupt Flag as a trigger source makes the ADC start a new conversion as soon
as the ongoing conversion has finished. The ADC then operates in Free Running mode, constantly
sampling and updating the ADC Data Register."
It doesn't talk about how fast the free running mode is. Did you use some conversion factor? I'm completely new to micro-controllers. Look like I have a lot reading to do. Can you suggest a good place to start?

Why would you think sampling rate is not a problem? Even if I used UNO, it still isn't a problem or you meant that if I use mega, It won't be a problem anymore? It looks like mega has the same cpu speed as UNO, it does have a lot bigger memory. You think the memory is a problem, is it because I need to get better frequency resolutions?

Exactly, don't confuse ADC resolution (which is least of your worry, as low as 8-bit o'k), and FFT resolution, that defined as Freq. range / Bin numbers. Having 4 kHz sampling and / 512-fft = 8 Hz. Paradox is , you need to lower sampling rate, to get better resolution! (I mention this in musical note recognition post). Other parameter, increasing fft-size, isn't an option , except you have Due board.

I I think arduino FFT library only provide max # of bin of 256 to filled in the real data. It doesn't go up to 512. I was definitely confused by ADC resolution and FFT resolution. Thank you for clear that out. Amanda from instructable.com was able to turn down ADC resolution to 8 bit in order to speed up the sampling rate, but you're saying to lower it down to get better resolution, and I can clearly see your point. If I were to lower it down, how would I be able to cover all the frequency range? from 27hz to 4186hz.
I read your blog about pushing arduino to the limit. That was a very impressive project you made! A few questions after reading your blog:

  1. How could you have 1024 bins for the FFT? The one comes with arduino IDE is limited to 512 bins as the max. Half of the bins are filled with the data taken from analog input, the other half is set to zero manually. Do you use different library?
    2.The shadow masking technique you used, is it setting a range for the detected frequencies to fall in, and then light up corresponding LED even though the resolution limited to 7.8hz?
  2. If you already pushed arduino mega to the limit in that project, does it mean I have to either get a Due board or call off my project, or do something similar to yours just to understand how micro-controllers works? What if I use multiple mega boards? having each of them detecting different frequency ranges?
    4.From arduino FFT documentation, it suggests to set imaginary part to zero if the samples are taken from analog input. Is that what you did?

I know it's a lot of questions, but I'm really into micro-controller now, especially this topic. Please bear with me if I ask something that is too trivial.
Thank you

johnwasser:
The wikipedia article shows note frequencies with six significant digits. In the low notes that's a resolution of 0.0001 Hz. To get that resolution at 9,259 samples per second you would need an FFT size of 92.5 million samples (185 megabytes). That is not going to happen on an Arduino in any reasonable amount of time.

Perhaps you can use zero crossings to measure the frequency. Have you looked at any frequency counter sketches?

http://interface.khm.de/index.php/lab/experiments/arduino-frequency-counter-library/

Thank you for the suggestion. I've tried a code made by Amanda from instructable.com. It works really well up until the frequency gets up to 1khz. I will definitely look into the one you suggested here.

I also want to know how FFT works. I've read many article about how it works, but it seems like I didn't really understand everything.
Can you explain how you get 92.5 million samples is necessary to have six digits of resolution at 9259hz sampling rate? and how you figure out the size if 185 mB?

Can you also please teach me how to set the simulated frequency? I think that's what I was missing when I tested out the FFT library. I thought I was filling up the array with sine wave data generated by k = 127sin(23.14 * 100 * t ), where t was incremented 256 times, which was the same size of the FFT real bin, imaginary bins were filled with zeros. I thought with this method, the only bin that will contain numbers greater than zero is bin 100, but it is not the case. I have a lot numbers in other bins.

Thank you

I I think arduino FFT library only provide max # of bin of 256 to filled in the real data. It doesn't go up to 512. I was definitely confused by ADC resolution and FFT resolution. Thank you for clear that out. Amanda from instructable.com was able to turn down ADC resolution to 8 bit in order to speed up the sampling rate, but you're saying to lower it down to get better resolution, and I can clearly see your point. If I were to lower it down, how would I be able to cover all the frequency range? from 27hz to 4186hz.
No, you wouldn't. My project covers 4 octaves only
I read your blog about pushing arduino to the limit. That was a very impressive project you made! A few questions after reading your blog:

  1. How could you have 1024 bins for the FFT? The one comes with arduino IDE is limited to 512 bins as the max. Half of the bins are filled with the data taken from analog input, the other half is set to zero manually. Do you use different library?
    Code is linked, not a library, written by myself.
    2.The shadow masking technique you used, is it setting a range for the detected frequencies to fall in, and then light up corresponding LED even though the resolution limited to 7.8hz?
    Requirements to distinguish two notes and for tunning are different
  2. If you already pushed arduino mega to the limit in that project, does it mean I have to either get a Due board or call off my project, or do something similar to yours just to understand how micro-controllers works? What if I use multiple mega boards? having each of them detecting different frequency ranges?
    You may try Due, I used UNO btw, not Mega.
    4.From arduino FFT documentation, it suggests to set imaginary part to zero if the samples are taken from analog input. Is that what you did?
    Yes

Using fft may be appropriate if there are a few notes play simultaneously, which is not the case for you project. Freq. measuremets library linked in reply #18 and #28 is better approach for tunning.

hongp:
Can you explain how you get 92.5 million samples is necessary to have six digits of resolution at 9259hz sampling rate? and how you figure out the size if 185 mB?

It's the formula I showed earlier. The resolution of the FFT is sampleRate/sampleCount. To get resolution of 0.0001 Hz (1/10,000th Hz) you need sampleCount to be 10,000 times the sampleRate.

If each sample is a 2-byte integer you need 2*sampleCount bytes to store them.

johnwasser:
It's the formula I showed earlier. The resolution of the FFT is sampleRate/sampleCount. To get resolution of 0.0001 Hz (1/10,000th Hz) you need sampleCount to be 10,000 times the sampleRate.

If each sample is a 2-byte integer you need 2*sampleCount bytes to store them.

Got it! 2100009259/(2^20) = 176 mB.
Thank you so much!

To have "bin 100" represent 100 Hz you have to use a sample rate equal to the FFT size:

512 sample FFT at 512 samples per second.

I think the code would be something like this:

    for (int i = 0 ; i < 512 ; i += 2) {
      float time = i/512.0;       // Time in seconds from start of sample
      float cycle = time * 100;  // Cycles since start of samples
      fft_input[i] = 127 * sin(2 * Pi * cycle);; // put real data into even bins
      fft_input[i+1] = 0; // set odd bins to 0
    }

Yo, I got a few messages asking about the same problem I had with fft library.
If you still have problems getting the correct output from that library, look at my code.
This is the working code that I demo to my professor.

#define LOG_OUT 1 // use the log output function
#define FFT_N 256 // set to 256 point fft
#include <FFT.h> // include the library

//UNO clock Frequency = 16 MHz
//It takes 13 clock cycles to finish analog reading
int refFrequency = 0;
int currentFrequency = 0;
int previousFrequency = 0;
int errorFrequency = 0;
int refState = 0;
int currentState = 0;
int tunningState = 0;
boolean inRange = false;
float P = 0;
float I = 0.0;
float D = 0.0;
float drive = 0;
int pwmDrive = 0;
const int refInput = 8;  //reference pin
const int currentInput = 9;  //final pin
const int ForMotorControl = 10;  //pump control pin PWM
const int BacMotorControl = 11;  //pump control pin PWM
const int tunning = 12;  //tuning control pin
int bin = 0;
int frequency = 0;
int resolution = 9615 / 256;
int maxLogOut = 0;


void setup() {
  Serial.begin(9600); // use the serial port
  pinMode(refInput, INPUT);
  pinMode(currentInput, INPUT);
  pinMode(tunning, INPUT);
  pinMode(ForMotorControl, OUTPUT);
  pinMode(BacMotorControl, OUTPUT);
  ADCSRA |= (1<<ADPS2)|(1<<ADPS1)|(1<<ADPS0); //set adc prescaler=128
  ADMUX = 0x40; // use adc0
  ADCSRA |= (1<<ADATE);  //enable auto trigger
  ADCSRA |= (1<<ADEN);  //enable ADC
  ADCSRA |= (1<<ADSC);  //start ADC measurements
}

void loop() {
  while(1) { // reduces jitter
    cli();  // UDRE interrupt slows this way down on arduino1.0
    //sampling
    for (int i = 0 ; i < 512 ; i += 2) { // save 256 samples
      while(!(ADCSRA & 0x10)); // wait for adc to be ready
      ADCSRA |=(1<<ADSC); //start ADC measurements
      byte m = ADCL; // fetch adc data
      byte j = ADCH;
      int k = (j << 8) | m; // form into an int
      k -= 0x0200; // form into a signed int
      k <<= 6; // form into a 16b signed int
      fft_input[i] = k; // put real data into even bins
      fft_input[i+1] = 0; // set odd bins to 0
    }//end of for
    
    //FFT
    fft_window(); // window the data for better frequency response
    fft_reorder(); // reorder the data before doing the fft
    fft_run(); // process the data in the fft
    fft_mag_log(); // take the output of the fft
    sei();
    maxLogOut = 0;
    for(int i = 1; i < 128; i++){
      if(fft_log_out[i] > maxLogOut && fft_log_out[i] > 150){
          maxLogOut = fft_log_out[i];
          bin = i;
      }//end of if
    }//end of for
    frequency = bin * resolution;
    if(frequency > 1500 && frequency < 2700){
      inRange = true;
    } else {
      inRange = false;
    }
    
    //if digital pin 8 is high, set the reference frequency
    //if digital pin 9 is high, set the current frequency
    //if digital pin 12 is high, set tunning state = high
    refState = digitalRead(refInput);
    currentState = digitalRead(currentInput);
    tunningState = digitalRead(tunning);
    if(refState == HIGH && inRange == true){
      refFrequency = frequency;
    } else if(currentState == HIGH && inRange == true){
      currentFrequency = frequency;
    } //end of else if
    
    //error tracking
    //highest frequency = (2405)2479 hz  lowest frequency = (1517) 1628 hz
    errorFrequency = refFrequency - currentFrequency;
    P = errorFrequency;  //proportional term
   // I = Integral * 0;  //intergral term
   // D = (previousFrequency - currentFrequency) * 0;  //derivative term
   // previousFrequency = currentFrequency;  
    drive = P * (255.0/888.0); //scaleFactor = 255 * 888;

    //start tuning if tunning state = high
    if(tunningState == HIGH){
    //motor control
      if(errorFrequency > 30){      //pump liquid out 
        //digitalWrite();  //pump in one direction
        digitalWrite(BacMotorControl, HIGH);
        analogWrite(ForMotorControl, 0);
      } else if (errorFrequency < -30){  //pump liquid in
        //digitalWrite();  //pump in another direction
        digitalWrite(BacMotorControl, LOW);
        pwmDrive = (int)abs(drive);
        if(pwmDrive < 80){ pwmDrive = pwmDrive + 80;}
        analogWrite(ForMotorControl, pwmDrive);
      } else {  //stop pumping
        analogWrite(ForMotorControl, 0);  //stop the motor
        digitalWrite(BacMotorControl, LOW);
      }
    } else {
      analogWrite(ForMotorControl, 0);  //stop the motor
      digitalWrite(BacMotorControl, LOW);
    }
    
    //debug
    Serial.print("frequency: ");
    Serial.print(frequency);
    Serial.print(" | ");
    Serial.print("reference frequency: ");
    Serial.print(refFrequency);
    Serial.print(" | ");
    Serial.print("Current frequency: ");
    Serial.print(currentFrequency);
    Serial.print(" | ");
    Serial.print("error frequency: ");
    Serial.println(errorFrequency);
  }//end of while
}//end of loop

hongp, thank you brother.

hongp:
Can I do the same to fft library in arduino? I tried it. but a lot of the outputs aren't zero. Not sure if it's correct. Please point out any possible error.
Here is the code I used to test fft library:

#define LOG_OUT 1 // use the log output function

#define FFT_N 256 // set to 256 point fft
#include <FFT.h> // include the library

void setup() {
  Serial.begin(9600); // use the serial port
}

void loop() {
  // put your main code here, to run repeatedly:
  int k = 0;
  int t = 0;
    for (int i = 0 ; i < 512 ; i += 2) { // save 256 samples
      k = 127sin(23.14 * 100 * t );
      t++;
      fft_input[i] = k; // put real data into even bins
      fft_input[i+1] = 0; // set odd bins to 0
    }
    fft_window(); // window the data for better frequency response
    fft_reorder(); // reorder the data before doing the fft
    fft_run(); // process the data in the fft
    fft_mag_log(); // take the output of the fft
    for(int i = 0; i < 128; i++){
          Serial.println(fft_log_out[i]);
    }
    delay(1000);
}

I used your codes based on fft library but where is it mentioned that the input is from A0.

I tried them with my hydrophone, pre-amp setup and I get the same answer:

32 43 39 43 35 39 39 39 30 30 30 35 59 82 58 35 19 19 24 27 27 27 30 19 27 24 30 27 24 24 27 27 
32 43 39 43 35 39 39 39 30 30 30 35 59 82 58 35 19 19 24 27 27 27 30 19 27 24 30 27 24 24 27 27 
32 43 39 43 35 39 39 39 30 30 30 35 59 82 58 35 19 19 24 27 27 27 30 19 27 24 30 27 24 24 27 27

where am I going wrong

in setup ( ADMUX = 0x40; // use adc0 ) this is using assembly language, a very basic level code. I took a class on it and its not easy or fun.

whats a hydrophone?

rexhex:
in setup ( ADMUX = 0x40; // use adc0 ) this is using assembly language, a very basic level code. I took a class on it and its not easy or fun.
Assembly language - Wikipedia

ADC - Wikipedia

whats a hydrophone?

Hydrophone is an underwater mic

rexhex:
in setup ( ADMUX = 0x40; // use adc0 ) this is using assembly language, a very basic level code. I took a class on it and its not easy or fun.
Assembly language - Wikipedia

ADC - Wikipedia

whats a hydrophone?

AWOL lol

Hi

I managed to get a sine curve using the codes.

void setup() {
 Serial.begin(9600);
}

void loop() { 
  int a[250];  //250 refers sample number
  for (int i=0; i<250; i++) {
  a[i] = analogRead(0);
  int k = 250*sin(2*3.14*0.006*i);
  Serial.println(k);
  }
Serial.println();
}

Now, if the sine curve varies with output from A0, I will obtain a FFT curve.
How do i accomplish this ?