Tutorial: Using Statistical Process Control to monitor your signal

Table of Contents

• Introduction
  • Available methods
• Explanations about the tutorial
  • How to run the tutorial
  • Detailed explanations about the SPC tutorial

Introduction

Statistical Process Control (SPC) is defined as the use of statistical methods to monitor if a signal is "in control".

In this tutorial, we will use a Statistical Process Control method to monitor if a random signal following a normal distribution is "in control".

Available methods

The different methods available in ViSP aim at detecting if the mean of a signal is changing, either due to an abrupt jump or due to a slow drift.

The different methods that are available are the following:

• Exponentially Weighted Moving Average (EWMA), implementent in the vpStatisticalTestEWMA class.
• Hinkley's test, implemented in the vpStatisticalTestHinkley class.
• Mean Adjusted Cumulative Sum (mean adjusted CUSUM), implemented in the vpStatisticalTestMeanAdjustedCUSUM class.
• Shewhart's test, implemented in the vpStatisticalTestShewhart class.
• Sigma test, implemented in the vpStatisticalTestSigma class.

We refer the reader to the documentation of each class to have more detailed information on each method.

Explanations about the tutorial

How to run the tutorial

To see the different options of the tutorial, please run the following commands:

$ cd $VISP_WS/visp-build/tutorial/mean-drift
$ ./tutorial-meandrift -h

If you run the program without argument, you should see something similar to the following image:

A Gaussian signal of mean equal to 6 and of standard deviation equal to 2 is generated, without any mean drift. The program tells you which method has been chosen in the console, and which are its parameters. A monitoring loop stops once an alarm is raised. When the alarm is raised, some information about the alarm and the test signal(s) + limits of the SPC method are given. Press Return to leave the program.

Detailed explanations about the SPC tutorial

For this tutorial, we use the main program tutorial-meandrift.cpp .

It uses the following enumeration to let the user choose which SPC method to use to monitor the signal:

 
typedef enum TypeTest
{
  HINLKEY_TYPE_TEST = 0, 
  EWMA_TYPE_TEST = 1, 
  MEAN_ADJUSTED_CUSUM_TYPE_TEST = 2, 
  SHEWHART_TYPE_TEST = 3, 
  SIGMA_TYPE_TEST = 4, 
  COUNT_TYPE_TEST = 5, 
  UNKOWN_TYPE_TEST = COUNT_TYPE_TEST 
}TypeTest;

The program arguments are parsed and fill the following structure that stores the SPC methods parameters:

 
typedef struct ParametersForAlgo
{
  unsigned int m_test_nbsamples; 
  bool m_test_activatedalarms[vpStatisticalTestAbstract::MEAN_DRIFT_COUNT]; 
  unsigned int m_test_nbactivatedalarms; 
  float m_cusum_h; 
  float m_cusum_k; 
  float m_ewma_alpha; 
  float m_hinkley_alpha; 
  float m_hinkley_delta; 
  bool m_hinkley_computealphadelta; 
  float m_hinkley_h; 
  float m_hinkley_k; 
  bool m_shewhart_useWECO; 
  bool m_shewhart_rules[vpStatisticalTestShewhart::COUNT_WECO - 1]; 
  float m_sigma_h; 
  ParametersForAlgo()
    : m_test_nbsamples(30)
    , m_cusum_h(4.76f)
    , m_cusum_k(1.f)
    , m_ewma_alpha(0.1f)
    , m_hinkley_alpha(4.76f)
    , m_hinkley_delta(1.f)
    , m_hinkley_computealphadelta(false)
    , m_hinkley_h(4.76f)
    , m_hinkley_k(1.f)
    , m_shewhart_useWECO(false)
    , m_sigma_h(3.f)
  {
    std::memcpy(m_shewhart_rules, vpStatisticalTestShewhart::CONST_ALL_WECO_ACTIVATED, (vpStatisticalTestShewhart::COUNT_WECO - 1) * sizeof(bool));
    memcpy(m_test_activatedalarms, CONST_ALL_ALARM_ON, vpStatisticalTestAbstract::MEAN_DRIFT_COUNT * sizeof(bool));
    m_test_activatedalarms[vpStatisticalTestAbstract::MEAN_DRIFT_NONE] = false;
    m_test_nbactivatedalarms = meanDriftArrayToNbActivated(m_test_activatedalarms);
  }
}ParametersForAlgo;

It is possible to choose to monitor only upward mean drifts, only downward mean drifts or both. To do so, use the --alarms option with the name of the alarm(s) you want to monitor.

First, the plot that will show the data is created in the following section of code:

  vpPlot plotter(1);
  plotter.initGraph(0, 1);
  plotter.setTitle(0, "Evolution of the signal");
  plotter.setUnitX(0, "Frame ID");
  plotter.setUnitY(0, "No units");

Then, the desired method is created in the following section of code, with the desired parameters:

  unsigned int idFrame = 0;
  vpStatisticalTestAbstract *p_test = nullptr;
  switch (type) {
  case TutorialMeanDrift::EWMA_TYPE_TEST:
    p_test = new vpStatisticalTestEWMA(parameters.m_ewma_alpha);
    break;
  case TutorialMeanDrift::HINLKEY_TYPE_TEST:
    p_test = new vpStatisticalTestHinkley(parameters.m_hinkley_alpha, parameters.m_hinkley_delta, parameters.m_test_nbsamples);
    break;
  case TutorialMeanDrift::MEAN_ADJUSTED_CUSUM_TYPE_TEST:
    p_test = new vpStatisticalTestMeanAdjustedCUSUM(parameters.m_cusum_h, parameters.m_cusum_k, parameters.m_test_nbsamples);
    break;
  case TutorialMeanDrift::SHEWHART_TYPE_TEST:
    p_test = new vpStatisticalTestShewhart(parameters.m_shewhart_useWECO, parameters.m_shewhart_rules, parameters.m_test_nbsamples);
    break;
  case TutorialMeanDrift::SIGMA_TYPE_TEST:
    p_test = new vpStatisticalTestSigma(parameters.m_sigma_h, parameters.m_test_nbsamples);
    break;
  default:
    throw(vpException(vpException::badValue, TutorialMeanDrift::typeTestToString(type) + " is not handled."));
    break;
  }
 
  if ((type == TutorialMeanDrift::HINLKEY_TYPE_TEST) && parameters.m_hinkley_computealphadelta) {
    // Initialization of Hinkley's test in automatic mode
    delete p_test;
    p_test = new vpStatisticalTestHinkley(parameters.m_hinkley_h, parameters.m_hinkley_k, true, parameters.m_test_nbsamples);
  }

Then, the method is filled with "in control" signals to initialize the expected mean and the standard deviation:

  // Initial computation of the mean and stdev of the input signal
  for (unsigned int i = 0; i < parameters.m_test_nbsamples; ++i) {
    vpGaussRand rndGen(stdev, mean, static_cast<long>(idFrame * dt));
    signal = static_cast<float>(rndGen());
    p_test->testDownUpwardMeanDrift(signal);
    ++idFrame;
  }

Then, the monitoring loop is run, where the signal is randomly generated with a potential mean drift (if desired). This random signal is then tested, and the loop is stopped if an alarm we want to monitor is raised:

  float mean_eff = mean;
  bool hasToRun = true;
  vpStatisticalTestAbstract::vpMeanDriftType drift_type = vpStatisticalTestAbstract::MEAN_DRIFT_NONE;
  while (hasToRun) {
    vpGaussRand rndGen(stdev, mean_eff, static_cast<long>(idFrame * dt));
    signal = static_cast<float>(rndGen());
    plotter.plot(0, 0, idFrame - parameters.m_test_nbsamples, signal);
    drift_type = p_test->testDownUpwardMeanDrift(signal);
    if ((drift_type != vpStatisticalTestAbstract::MEAN_DRIFT_NONE) && (parameters.m_test_activatedalarms[drift_type])) {
      hasToRun = false;
    }
    else {
      mean_eff += mean_drift;
      ++idFrame;
    }
  }

Finally, some information about why the loop was stopped is displayed in the console:

  std::cout << "Test failed at frame: " << idFrame - parameters.m_test_nbsamples << std::endl;
  std::cout << "Type of mean drift: " << vpStatisticalTestAbstract::vpMeanDriftTypeToString(drift_type) << std::endl;
  std::cout << "Last signal value: " << signal << std::endl;
  if (type == TutorialMeanDrift::EWMA_TYPE_TEST) {
    vpStatisticalTestEWMA *p_testEwma = dynamic_cast<vpStatisticalTestEWMA *>(p_test);
    std::cout << "\tw(t) = " << p_testEwma->getWt() << std::endl;
  }
  else if (type == TutorialMeanDrift::MEAN_ADJUSTED_CUSUM_TYPE_TEST) {
    vpStatisticalTestMeanAdjustedCUSUM *p_testCusum = dynamic_cast<vpStatisticalTestMeanAdjustedCUSUM *>(p_test);
    std::cout << "\tLower cusum = " << p_testCusum->getTestSignalMinus() << std::endl;
    std::cout << "\tUpper cusum = " << p_testCusum->getTestSignalPlus() << std::endl;
  }
  else if (type==TutorialMeanDrift::SHEWHART_TYPE_TEST) {
    vpStatisticalTestShewhart *p_testShewhart = dynamic_cast<vpStatisticalTestShewhart *>(p_test);
    std::vector<float> signal = p_testShewhart->getSignals();
    size_t nbSignal = signal.size();
    std::cout << "Signal history = [ ";
    for (size_t i = 0; i < nbSignal; ++i) {
      std::cout << signal[i] << " ";
    }
    std::cout << "]" << std::endl;
    std::cout << "\tWECO alarm type = " << vpStatisticalTestShewhart::vpWecoRulesAlarmToString(p_testShewhart->getAlarm()) << std::endl;
  }
  else if (type == TutorialMeanDrift::HINLKEY_TYPE_TEST) {
    vpStatisticalTestHinkley *p_hinkley = dynamic_cast<vpStatisticalTestHinkley *>(p_test);
    float Mk = p_hinkley->getMk();
    float Nk = p_hinkley->getNk();
    float Sk = p_hinkley->getSk();
    float Tk = p_hinkley->getTk();
    std::cout << "S+(t) = " << Tk - Nk <<std::endl;
    std::cout << "S-(t) = " << Mk - Sk <<std::endl;
  }
  float limitDown = 0.f, limitUp = 0.f;
  p_test->getLimits(limitDown, limitUp);
  std::cout << "\tLimit down = " << limitDown << std::endl;
  std::cout << "\tLimit up = " << limitUp << std::endl;

The program stops once the Return key is pressed.