Fault detection of gas sensor arrays based on KNN rules

Fault detection techniques are critical to improving the efficiency of airborne overall equipment. It has been recognized that fault detection based on the k nearest neighbor rules (kNN) can effectively handle some of the characteristics of the gas sensor process, such as multimode bulk trajectories and nonlinearities. However, the computational complexity and storage space involved in kNN's neighbor search prevents it from being monitored online, especially for high-dimensional situations. In order to solve this problem, Before the implementation of kNN rules to fault detection, we reduce dimensions by principal component analysis (PCA) and propose a K-fold cross-verification method in the training stage of the model to learn from multiple directions, which plays a good role in improving the accuracy of fault diagnosis. This method enables the verification set to traverse all the data sets during the division of the training set and the test set, thus avoiding the problem of local minima to a certain extent. In the specified range of K values, the smaller the K value is, the fewer the nearest neighbors are, and the more useful training samples can not be included, resulting in the increase of false recognition rate; the K value is too large and the noise significantly affects the error rate; the optimal K value is obtained by cross-verification K nearest neighbor algorithm, so that the error rate of pattern recognition is the lowest. The practical application shows that this method can realize the on-line diagnosis and location of array gas sensor faults, and can be applied to other similar systems.