Multivariate statistics and neural networks in process fault detection

The objective of statistical process control (SPC) is to monitor the performance of a process over time in order to verify that it remains in a state of statistical control. Until recently, the application of SPC to manufacturing processes has been based upon the charting of a small number of variables using univariate tools. This approach is appropriate if the variables an known to exhibit independent behaviour. However, examining a limited number of preselected variables as if they were independent makes identification and interpretation of potential process malfunctions difficult, possibly misleading and dangerous. Also, malfunctions may not be identified. The multivariate techniques of principal component analysis and projection to latent structures overcome these difficulties. This is achieved by projecting the multivariate data down on to a lower dimensional space which contains all the relevant information in possibly as few as two or three latent variables, which are a linear combination of the original variables. In the fuzzy neural network fault detection approach, a preprocessing fuzzification layer is added to a conventional feedforward artificial neural net topology. The fuzzification layer converts increments in online measurements and controller outputs (deviations from steady state) into three fuzzy sets (increase/steady/decrease) with corresponding membership functions. Each output of the network represents a particular fault calibrated to the range (0,1) with a fault being indicated by an output close to 1