Fault detection and diagnosis of dynamic processes using weighted dynamic decentralized PCA approach
9. Decentralized and Dynamic Fault Detection Using PCA and Bayesian Inference
As industrial plants tend to increase their level of automation the number of sensors and, therefore, the amount of collected data is huge, this implies that the requested computational power to process this amount of data has grown signiﬁcantly. For this reason, the fault detection have became highly time and resources consuming. In small plants fault detection can be developed using a central processor which collects all the information from the sensors, but in complex plants, with lot of sensors, this central processor must be able to deal with this vast amount of data, which is not always possible. One way to afford this problem is dividing the plant in blocks, each one containing a speciﬁc number of measured variables in it and assigning one small processor in every block, this is called decentralized approach. Here, each local processor does the fault detection task with its own sensors and send the result to a central processor which only has to fuse the local results to obtain a global detection result.
The most important step in this technique is to establish the criteria used to divide the plant. Some authors opted for one-variable blocks , while other authors constructed blocks that group more than one variable ( , , , ).
Once the local fault detection models obtain a result, it is necessary to fuse all these results in order to get an unique decision for the whole plant. There are different ways to do that as  that proposed Maximum Entropy, Multicriteria decision making (MCDM) methods as Ordered Weighted Average (OWA) operators  are another option, also ( , ) used Bayesian Inference Criterion (BIC) for their Fault Detection and Diagnosis (FDD) method.
The ﬁrst step is to decompose the plant into different blocks and then to perform a DPCA method to detect faults locally in each block, and, ﬁnally, a central processor fuse all the local detection results using the BIC index to take a global decision.
Sparse PLS, ANN
8. Knowledge-data-integrated sparse modeling for batch process monitoring
#sparse-modeling #variable-group #multiblock
According to process flow diagram (PFD) and piping and instrumentation diagram (P&ID) of the industrial process, the actual variable correlations are classiﬁed into four categories: control, reaction, type and location correlations.
- Using variable correlations based on process knowledge or similarities
- control correlation
- reaction and location correlation
- type correlation
- similarity driven variable grouping technique
- Variable correlation tendency analysis to analyze the variables in the same group are positively or negatively correlated
- Knowledge based sparse projection matrix
Construct a projection matrix based on the grouping and tendency result.
- Two level contribution plot
An industrial scale fed-batch fermentation process carried out in a 100000L bioreactors. link