Decision Trees and the Effects of Feature Extraction Parameters for Robust Sensor Network Design

Abstract

Reliable sensors and information are required for reliable condition monitoring. Complex systems are commonly monitored by many sensors for health assessment and operation purposes. When one of the sensors fails, the current state of the system cannot be calculated in same reliable way or the information about the current state will not be complete. Condition monitoring can still be used with an incomplete state, but the results may not represent the true condition of the system. This is especially true if the failed sensor monitors an important system parameter. There are two possibilities to handle sensor failure. One is to make the monitoring more complex by enabling it to work better with incomplete data; the other is to introduce hard or software redundancy. Sensor reliability is a critical part of a system. Not all sensors can be made redundant because of space, cost or environmental constraints. Sensors delivering significant information about the system state need to be redundant, but an error of less important sensors is acceptable. This paper shows how to calculate the significance of the information that a sensor gives about a system by using signal processing and decision trees. It also shows how signal processing parameters influence the classification rate of a decision tree and, thus, the information. Decision trees are used to calculate and order the features based on the information gain of each feature. During the method validation, they are used for failure classification to show the influence of different features on the classification performance. The paper concludes by analysing the results of experiments showing how the method can classify different errors with a 75% probability and how different feature extraction options influence the information gain.