how to fix oscillations of analog input read by STM32F107

Question

I have to read the input value from an external source that is a balance using the processor STM32F107. This balance is external to the board that contains the processor and communicates with it via PA4.

Here is my first attempt to read the input from the balance.

I use this function to setup the ADC:

void ADC_Configuration(void) {

    ADC_InitTypeDef ADC_InitStructure;
   /* PCLK2 is the APB2 clock */
   /* ADCCLK = PCLK2/6 = 72/6 = 12MHz*/
   RCC_ADCCLKConfig(RCC_PCLK2_Div6);
   /* Enable ADC1 clock so that we can talk to it */
   RCC_APB2PeriphClockCmd(RCC_APB2Periph_ADC1, ENABLE);
   /* Put everything back to power-on defaults */
   ADC_DeInit(ADC1);

   /* ADC1 Configuration ------------------------------------------------------*/
   /* ADC1 and ADC2 operate independently */
   ADC_InitStructure.ADC_Mode = ADC_Mode_Independent;
   /* Disable the scan conversion so we do one at a time */
   ADC_InitStructure.ADC_ScanConvMode = DISABLE;
   /* Don't do contimuous conversions - do them on demand */
   ADC_InitStructure.ADC_ContinuousConvMode = DISABLE;
   /* Start conversin by software, not an external trigger */
   ADC_InitStructure.ADC_ExternalTrigConv = ADC_ExternalTrigConv_None;
   /* Conversions are 12 bit - put them in the lower 12 bits of the result */
   ADC_InitStructure.ADC_DataAlign = ADC_DataAlign_Right;
   /* Say how many channels would be used by the sequencer */
   ADC_InitStructure.ADC_NbrOfChannel = 1;

   /* Now do the setup */
    ADC_Init(ADC1, &ADC_InitStructure);
   /* Enable ADC1 */
   ADC_Cmd(ADC1, ENABLE);
   /* Enable ADC1 reset calibaration register */
   ADC_ResetCalibration(ADC1);
   /* Check the end of ADC1 reset calibration register */
   while(ADC_GetResetCalibrationStatus(ADC1));
   /* Start ADC1 calibaration */
   ADC_StartCalibration(ADC1);
   /* Check the end of ADC1 calibration */
   while(ADC_GetCalibrationStatus(ADC1));
}

And I use this function to get the input:

u16 readADC1(u8 channel) {

   ADC_RegularChannelConfig(ADC1, channel, 1, ADC_SampleTime_1Cycles5);

   // Start the conversion
   ADC_SoftwareStartConvCmd(ADC1, ENABLE);
   // Wait until conversion completion
   while(ADC_GetFlagStatus(ADC1, ADC_FLAG_EOC) == RESET);
   // Get the conversion value

   return ADC_GetConversionValue(ADC1);
}

The problem is that in N measurements of the same weight, I get N different results. For example, the weight is 70kg and the output of the readADC1(ADC_Channel_4) is 715,760,748,711,759 etc.

What am I doing wrong?

Edit. I have added this function (that simulates a lp filter) to stabilize the input and it works fine. The problem is how to convert the value returned form this function to kilograms. Using a costant coefficient (determined by measuring a known object.) gives a growing error proportional to increasing weight of the input. Any suggest to have a better conversion?

double fix_oscillations(){
    int i;
    double LPOUT=0,LPACC=0;
    int K = 5000;

    for(i=0;i<5000;i++){
       LPACC = LPACC + readADC1(ADC_Channel_4) - LPOUT;
       LPOUT = LPACC / K;
    }
    return LPOUT;
}

Answer 1

It is possible either that the input is inherently noisy or that noise is introduced by other equipment. It is worth observing the the signal with an oscilloscope to determine the kind of interference or noise on the signal since that may affect the best solution.

If possible you should eradicate any source of external interference, then external analogue signal condition should be applied, ideally by a low-pass filter with a cut-off frequency of half your intended sample rate or less. Then you are in a position to apply filtering in the digital domain if necessary. For static signals (voltage levels), a simple block average will suffice, for faster moving signals a moving average (boxcar filter) would be better. For complex signals from which you need to extract certain frequencies, more complex filters are needed, but that is probably not the case in this instance.

Answer 2

Edit. I have added this function (that simulates a lp filter) to stabilize the input and it works fine. The problem is how to convert the value returned form this function to kilograms. Using a costant coefficient (determined by measuring a known object.) gives a growing error proportional to increasing weight of the input. Any suggest to have a better conversion?

This was probably worth posting a separate question on since it is not clearly related to the question title.

If your output is non linear, you could use multiple calibration points with linear interpolation between points, however it is likely that the curve can be characterised by an equation. Take a number of measurements over the range of interest, then plot these points in a spreadsheet tool such as Excel or OpenOffice.org Calc. The charting tools include "trend-line" plotting of various types, and can display the equation of the curve. Select the simplest curve type with adequate fit. If you need to resort of a polynomial with more than two terms, makes sure that you display the equation terms with sufficient decimal places since these can be critical. You can test to see if you have enough precision by using the equation to generate a curve and see how well it meets the trendline. Plotting the equation is probably a good idea for any of the curve as a test for sufficient precision. When you write the code, make sure you use a data type with sufficient precision too.

A note about the line:

LPOUT = LPACC / K;

by taking 5000 samples you effectively increased your ADC resolution by about 12 bits (at the expense of sample time), but by dividing by K, you have lost that precision unnecessarily, and its a truncating divide as well. It would be better to use the undivided sum directly in the conversion to Kg. I realise that dividing makes the value appear "stable" but it is about signal:noise ratio not absolute noise magnitude. The conversion to Kg and display to the required real-world precision will have the same effect of "stabilising" the result but with a smaller cumulative error.

Answer 3

Have you seen the STM32 application notes for improving ADC accuracy? There is also an application note for increasing resolution, and a few others as well.

You should have a look at this link here. You can then click "Documentation" and select "Application Note" and search for "Accuracy".

STMicro's application notes are often helpful, as they like to get their products out sooner than later, and then append documentation/support afterwords.

Answer 4

You might find this a useful way of averaging:

float FilteredValue;
#define TIME_CONSTANT 100
FilteredValue += ((float)ADCreading - FilteredValue)/TIME_CONSTANT;

This implements a true low pass filter with a time constant of TIME_CONSTANT x Sample Frequency. If there is a step change in ADCreading, FilteredValue will gradually change to the new value. In theory it will never reach it, since it implements an inverse exponential filter. The larger the value of TIME_CONSTANT the better the noise rejection but the longer it will take to stabilise.