Fault diagnosis method based on modified random forests

To solve the problem of inefficient determining fault location in unidentified fault diagnosis of traditional machine-learning technologies, a fault diagnosis method based on modified random forests was proposed. Firstly, random decision trees were created via modified algorithm of bagging and unbiased split selection based on conditional probability index so as to construct random forests. Secondly, weighted voting was applied to combine the prediction of the decision trees. Then, fault prototypes were computed through the measurement of variable-importance in random forests, which assisted in determining the fault location. Finally, the proposed method was illustrated and documented thoroughly in an application of standard dataset and Tennessee Eastman Process (TEP) fault diagnosis. The results verified the presented approach's feasibility and effectiveness.