Statistical Methods for Improving Continuous Production
The project started on 1 January 2014 funded by the Swedish Research Council and will run for five years, i.e. to the end of December 2018. Besides the project manager and researcher Murat Kulahci also postgraduate student Francesca Capaci and researchers Erik Vanhatalo and Bjarne Bergquist work in the project.
Continuous production processes are common in the process industries, e.g. in the mining, steel, chemical, pharmaceutical and paper industries. Since process industries are an important part of Swedish industry, these processes are quite vital for the Swedish economy. A continuous production means that products are produced around the clock, all year round in an unbroken flow. The products can for example be in the form of powder, liquid, particles or solid. This makes it difficult to track the product along the production chain and assure the final quality.
Process improvement efforts normally involves production being continuously monitored to detect systematic variation and problems that should be eliminated. In that respect it also becomes very important to better understand the production process by studying the impact of the certain inputs on the expected outcome. This often requires data gathering through controlled experiments that will bring forth more relevant information than the observational data. Statistical Process Control and Design of Experiments are therefore indispensable tools for achieving desired quality and improved productivity.
Literature describing methods in the areas of statistical process control and design of experiments is extensive. Unfortunately, these methods are often poorly adapted to the particular circumstances that continuous production entails. Moreover today’s automated data collection schemes allows for high-frequency, continuous sampling which increases the amount of data and the serial dependence amongst observations. From experimentation perspective, this leads to responses and/or input variables to be best described as time series. This is relatively unchartered territory in the classical applications of design of experiments where both the response and the inputs are usually in the form of scalars. As for the statistical process control, the modern data collection schemes result in (mega)multivariate data with high cross (among variables) and auto (in time) correlation. Many standard statistical methods assume time independent observations. If ignored or not properly handled, autocorrelation will lead to erroneous conclusions in statistical process control application, as in excessive false alarms or delays in detection. What happens when the autocorrelation and cross correlation occurs at the same time as in many continuous production applications is especially problematic.
This project has two specific goals.
Goal 1: Develop statistical methods based on statistical process control and control charts that can address the problems of cross- and autocorrelation. First, we aim to study how the existing methods are affected by the autocorrelation. We will then develop methods in the field of multivariate statistical process control that can handle both cross- and autocorrelated multivariate data.
Goal 2: Develop statistical methods for the analysis of experiments that take into account that the responses in process industry often generate time series instead of individual observations from the individual experimental settings. By combining the statistical fields of time series analysis and statistical design of experiments, we intend to create more powerful analytical tools for practitioners. We will also explore the case where the experimental inputs are also in the form of times series. Both the design and analysis aspects of such experiments will be studied.
Updated: