kNN-Ae: A Novel Concept for Increasing The Performance of Anomaly Detection on PM_2.5 Faulty Sensors with Autoencoder

Failure to measure PM<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</inf> gas due to sensor offset failure can result in inaccurate information for users. However, faulty sensor detection using conventional machine learning techniques still has difficulty separating normal data from anomaly data. The purpose of this study is to apply the hybrid k-nearest neighbor-autoencoder (kNN-Ae) method to improve the performance of PM<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</inf> faulty sensor anomaly detection. The first step in this study is to obtain a secondary dataset on faulty sensors from the Kaggle open repository. The next step is to benchmark the kNN method with other methods, support vector machine (SVM), naïve Bayes, decision tree, and logistic regression. In the next step, we use an autoencoder to enhance the fault sensor dataset. We then compare the four initial machine learning methods with the improved dataset. We use the feature importance metric from the extra trees classifier to measure the performance of the enhanced dataset from the autoencoder. The test results using MDI show that the complexity of the autoencoder model affects the feature importance of the reconstruction error, where the less complex model has a significantly higher feature importance compared to the more complicated model. Then, the reconstruction error generated by the autoencoder is shown to improve kNN’s performance in fault sensor detection, increasing its accuracy from 0.71 to 0.92. The performance of the hybrid model is also better than other detection models, namely naïve Bayes, SVM, decision tree, and logistic regression.