Data-driven Dynamic Process Monitoring Theory And Its Industrial Applications

Industrial development is changing with each passing day,and production safety is an eternal topic.In the event of an accident,the factory will be shut down for maintenance in light cases,and environmental pollution and even human casualties in serious cases.The embankment of a thousand miles was destroyed in the ant nest.In order to avoid the discovery of malignant accidents,it is the top priority to monitor the production process and find hidden dangers in time.With the development of science and technology,process monitoring technology has become a powerful tool in ensuring industrial safety.Traditional process monitoring methods often assume that the process is static,that is,there is no correlation between samples at different times.Due to the complex operation mechanism,real industrial processes are often dynamic,that is,there is a correlation between samples at different times.This feature not only greatly deepens the difficulty of process monitoring,but also evolves small faults into large faults over time,causing great harm.In the era of big data,there is a wealth of information hidden behind massive industrial data.The data-driven process monitoring method,because it does not require too much prior knowledge and modeling mechanism,can fully mine data information and quickly extract sample features,which shows great advantages.Therefore,this thesis has done related studies on the monitoring problem of dynamic process.The main contributions of this thesis are as follows:1.Aiming at the data space division of dynamic and static components in dynamic process,a multi-step dynamic slow feature analysis method is proposed.The method can be divided into several steps: the first step is to divide the data space into dynamic components and static components by identifying the dynamic structure;the second step is to use the principal component analysis algorithm to monitor the principal components in the static components;the third step,a dynamic slow feature algorithm is used to monitor the residual space in the static component.By effectively dividing the data space,the method improves the monitoring performance,increases the interpretability of the monitoring results,and reduces false positives for non-stationary processes.2.Aiming at the problem that the amount of calculation in dynamic process is too high,based on the overall relationship between time-delay samples,a dot product feature analysis method is proposed.This method calculates the dot product feature between the current sample and the time-delayed sample,and analyzes the overall correlation between the samples.At the same time,the proofs of the detectability conditions of the method for additive faults and multiplicative faults are given.After comparison,the online computational complexity of this method is much lower than that of classical static algorithms such as principal component analysis.3.In order to further reduce the amount of computation,an autocorrelation feature analysis method is proposed based on the variable relationship between time-delayed samples.This method considers the correlation of a single variable between time-lag samples,and then obtains the autocorrelation characteristics.At the same time,the proof of the fault detectable condition on the dynamic model is given for the first time.The online computational complexity of this method is not only lower than that of static algorithms such as principal component analysis,but also lower than that of the dot product feature analysis method.4.To detect the incipient faults in a dynamic process,a recursive innovative component analysis method is proposed through combined with the sliding window function and the recursive method.The method first divides the data space to obtain innovative components,and then uses the sliding window function to focus on local information,which can detect minor faults in time.The recursive idea greatly reduces the online computational complexity while maintaining the monitoring performance.5.Furthermore,a self-attention principal component analysis method is proposed for incipient fault detection through combining the sliding window function and the self-attention mechanism.For the three key variables in the self-attention mechanism,a novel calculation method is given,which fully integrates global information and local information,and focuses attention on effective data features.This method is the first to introduce the self-attention mechanism into the study of multivariate statistical process monitoring,which is of pioneering significance.