Fault Detection And Diagnosis For Data Stream Based On Incremental Kernel Non-negative Matrix Factorization Method

With the rapid development of industrial technology, modern industrial process control system becomes more complex and intelligent and monitoring system needs more and more sensors, thus composing a sensor network. Its corresponding forms of data have been changing, among which is the data stream form. Data stream does not only have characteristics that traditional data has, but also has its own features. The fault detection and diagnosis for data stream is just beginning, so it is necessary to research about it.Kernel non-negative matrix factorization(KNMF) is a novel matrix decomposition algorithm that has been recently developed. It shows the whole structure of information by mining the partial information of the data. Therefore the decomposed matrix has a natural sparsity, non-negative restrictions, and pure additive operation, which is more in line with the actual situation of the industrial process. These features make KNMF algorithm more explanatory. This article will apply KNMF into the field of fault detection and diagnosis, and build the fault detection and diagnosis model.Specifically, the main contents are as follows:(1) This article explains the significance of KNMF from the perspective of geometry and gives a detailed analysis of KNMF to deepen the understanding of it.(2) In order to shorten the training time and meet the requirement of real-time and dynamic requirement, this article proposes incremental kernel non-negative matrix factorization(ILKNMF) algorithm, provides three corresponding principles of ILKNMF and analysis the training model and the adaptive model. Furthermore, it proposes a method to streamline the historical data to overcome the shortcoming of high storage requirements of the data stream.I(3) This paper builds a falut detection model based on the ILKNMF algorithm, designs two corresponding statistics K2 and SPE and gives a method for the control limits. As for on-line monitoring, it is a fault data when the value of K2 and SPE exceed the control limits. At last, this article establishes a fault identification model based on contribution diagram method.(4) A new method call FKNMF is based on KNMF and Fisher discriminant analysis. FKNMF is a supervised algorithm and monotonic. FKNMF regards fault diagnosis as a classification problem from the perspective of the pattern classification. ILFKNMF is a new method based on FKNMF in order to meet the requirements of data stream. This paper establishes two fault diagnosis models based on FKNMF and ILFKNMF respectively.