چكيده به لاتين
Statistical Process Control (SPC) plays an important role in modern processes due to industrial revolution and the emergence of competitive space among producers. The main purpose of SPC is to detect any out-of-control condition and moreover, provide the responsible parameter for deviation of process as an important diagnostic information. This information can help practitioners to find the root cause for deviation and restore the process to its in-control condition.
SPC in its earliest stages was summarized to Shewhart control charts which could only monitor one quality charachteristic over time. As the products and processes got more complicated, the need for monitoring more than one quality characteristric was felt and lead the researchers to develop multivariate control charts. Growing of SPC techniques continued and researches found that in some cases, it is more efficient to monitor the relationship between the response variable and one or more explanatory variables which is called as profile. The process speed and time consuming measurments lead scientists to find novel data types which can be gathered without any contact to the prduct and even interfering the process. This type of data is called image data which could provide thousands to milions of data points in a fraction of a second. Various types of images such as RGB, X-Ray, Infrared, etc, can provide us wide range of data types from the product including apparent or interior features. Volume and velocity of this data type could not be handled by traditional SPC methods. In recent years, some researches proposed new statistical methods that can handle image data.
In this dissertation, we tried to propose a number of statistical methods to monitor and analyze image data. In our proposed procedures, wavelet is applied as a transform to help us exploit frequency domain feaures from the image. The first proposed procedure of this dissertation is based on one dimensional descrete wavelet transform and applying Generalized Likelihood Ratio (GLR) control chart for monitoring images. Then, a sub- procedure is proposed which applies hard and soft thresholding to help the practitioner keep useful information in detail coefficients of wavelet transformation. Results show that in some cases, this sub procedure improve the performance evaluation indicatiors.
A 2 dimensional wavelet based remedy is then proposed and a likelihood ratio statistic is developed in process monitoring. This remedy helps us not to lose useful information inherited in the image structure. Numerical studies illustrate the improvement of performance indices by applying this techniques.
All of the aforementioned methods can signal for out of control states but after each signal, the practitioner needs to find the defect type which can lead to faster process restoration. For this purpose, an Artificial Neural Network (ANN) based procedure is proposed which applies Multi Layer Perceptron (MLP) to classify the image (which leads the monitoring procedure t signal) into the defect types. This remedy can be applied as a supplementary procedure for previous proposed methods to provide defect type after any out of control signal.
