A ROBUST METHOD TO IDENTIFY FAULTS IN CORRELATED SENSORS IN MACHINE CONDITION MONITORING

We present a robust method to identify and isolate faulty sensors among a set of correlated sensors. For each sen- sor, we estimate the sensor a number of times, using each of the other correlated sensors separately. We use the me- dian of these estimates as the estimate for the sensor. When up to less than half of the sensors are faulty, this method identifies the faulty sensors accurately. Since the median is used and since estimates for the same sensor as opposed to different original sensor values are used in the median, this method is very robust. The method gives much better spill-over and error recognition rates, com- pared to the traditional method of using the mean of the actual sensor measurements. 1. INTRODUCTION In automated monitoring of plants and machines, a model is trained based on sensor data collected during the nor- mal operation of the machine or plant. The new test sen- sor data are used as input to the trained model and it is checked if the test data are in agreement with the training data. If the residuals (actual data - estimates) are higher than some thresholds for a sensor, then a fault is reported. Sometimes, a large number of sensors are correlated with each other. They measure the same physical entity (such as temperature or pressure) at the same or similar machine parts. If a sensor behaves unlike others (a drift or a step), it is important to identify that sensor as soon as possible. If the estimate for a sensor is based on other sensors, as in