Process Monitoring Based On Gaussian- Bernoulli Restricted Boltzmann Mechine

Process monitoring technologies play an important part in modern industry integrated automation, which make the operation of processes and products quality stable, and the research work on process monitoring means a lot in both academia and application. With the rapid development of information technologies and network technologies, the costs of information collection, storage as well as transmission have reduced significantly. Industrial enterprises could get a mass of process data from sensors and distributed control systems in industrial process. Therefore, data-driven techniques for process monitoring have become a popular research subject and got a lot of progresses and applications.However, the traditional process monitoring methods haven’t considered the characteristics in complex industrial process, such as big data, nonlinear, data label unbalance etc, which has some limitations in practical application. This thesis summarized previous work and proposed several novel process monitoring mothods based on Gaussian-Bernoulli Restricted Boltzmann Machine.(1) Described the background, basic conceptions and research value of process monitoring. The main research and development were discussed in detail. Introduced data characteristics in complex industrail process and analysed advantages of Restricted Boltzmann Mechine in process monitoring.(2) Gaussian-Bernoulli Restricted Boltzmann Machine (GRBM) was constructed to model nonlinear process and a novel fault detection algorithm based on GRBM was proposed to monitor nonlinear process, which could extract the nonlinear feature in its hidden layer and get reconstruction data in its visible layer. Two types of monitoring statistics were constructed in both feature space and residual space.(3) A semi-supervised GRBM was proposed to solve the overfitting problem which traditional supervised learning methods couldn’t avoid, when there were a few labeled samples available. This novel fault classification algorithm could learn process feature from unlabeled data and use the information from labeled data. It could perform well even with insufficient labeled samples.