Failure Prediction for Robot Reducers by Combining Two Machine Learning Methods

There are many production robots used at car manufacturing plants, and each of them is fitted with several reducers. A breakdown of one of these reducers may cause a huge loss due to the stoppage of all production lines. Therefore, condition-based maintenance is currently being used to predict failures by predetermined thresholds for average and standard deviations. However, this method can cause many false alarms or some false negatives. There are some ways of suppressing false alarms, such as detecting a change in the probability density function. However, when false alarms are suppressed using the probability density function in the operational range, some false negatives may occur, leading to a breakdown of a reducer and huge loss. A false negative is caused by overlooking an anomaly with slight changes and it is difficult to detect using only the probability density function. Therefore, we developed the Difference Signum Method (DSM) to detect an anomaly with slight changes by focusing on such changes. Although DSM reduces false negatives, it can cause many false alarms. This paper proposes a new failure prediction method using ensemble learning of the probability density function and DSM in order to reduce both false positives and false negatives. Using this new failure prediction method, the number of alerts is now fewer than four times/week, a substantial reduction from nine times/week with the previous method. The number of false negatives reached the target value of zero times/year from two times/year using the probability density function. Therefore, the performance of this new failure prediction method makes it applicable to actual production lines.