Improving your results

Did unrecorded maintenance events, system inefficiencies, or a new normal operating mode happen during the time of flagged anomalies in the test set? If so, the results indicate those situations. Change your train-evaluation splits so that each normal mode is captured during model training.

Are the sensor inputs relevant to the failure labels? In other words, is it possible that the labels are related to one component of the equipment but the sensors are monitoring a different component? If so, consider building a new model where the sensor inputs and labels are relevant to each other and drop any irrelevant sensors. Alternatively, drop the labels you're using and train the model only on the sensor data.

Is the label time zone the same as the sensor data time zone? If not, consider adjusting the time zone of your label data to align with sensor data time zone.

Is the failure label range inadequate? In other words, could there be anomalous behavior outside of the label range? This can happen for a variety of reasons, such as when the anomalous behavior was observed much earlier than the actual repair work. If so, consider adjusting the range accordingly.

Are there data integrity issues with your sensor data? For example, do some of the sensors become nonfunctional during the training or evaluation data? In that case, consider dropping those sensors when you run the model. Alternatively, use a training-evaluation split that filters out the non-functional part of the sensor data.

Does the sensor data include uninteresting normal-operating modes, such as off-periods or ramp-up or ramp-down periods? Consider filtering those out of the sensor data.

We recommend that you avoid using data that contains monotonically increasing values, such as operating hours or mileage.