Variable Selection Method Based on Least Absolute Shrinkage and Selection Operator (LASSO) and kNN for Faults Isolation

In the multivariable fault isolation of industrial process, the effectiveness of fault isolation is affected by the “smearing” effect because of the correlation between variables. In order to isolate all the key fault variables correctly, the fault isolation problem is regarded as a variable selection problem. Firstly, LASSO is applied to the sparse contribution vector based on k-nearest neighbor. Secondly, the LASSO-KNN model based on sliding window is established, and the sparse representation of contribution vector is obtained by least angle regression (LAR) algorithm and cumulative weight contribution percent criterion, thus the main fault variables are isolated. Then an extended EWMA algorithm combined with sliding window mechanism is introduced to achieve on-line fault isolation. Finally, the effectiveness of the proposed algorithm is verified by a simulation example of TE process.