Understanding labeling - Amazon Lookout for Equipment

Amazon Lookout for Equipment is no longer open to new customers. Existing customers can continue to use the service as normal. For capabilities similar to Amazon Lookout for Equipment see our blog post.

Understanding labeling

Amazon Lookout for Equipment takes an input dataset, which it assumes is under normal operating conditions, and trains a model to detect deviations from this baseline, normal operation. However, if there are known periods of abnormal behavior in the input dataset, then that abnormal behavior can lead to less accurate models. To address this, we recommend that you use labels to identify the abnormal behavior in the input dataset and Lookout for Equipment can exclude that labeled data from model training. For example, if it is known that historical data for a machine contains data for planned or unplanned downtime states, you can use labels to identify and exclude the downtime state data from model training.

By using the labels as inputs to the model, Lookout for Equipment can use additional modeling techniques that can improve the accuracy of the model.

As an example, the following image shows the time intervals of known healthy equipment behavior and the time intervals of abnormal equipment behavior (that is, the width of the bars in the image).

Time intervals where abnormal behavior occured.

In your labeling data, you define the abnormal time interval (bar width in image) from the actual failure point (for example, Failure 1). You provide the labeled data as a CSV file to model training. Each line of the CSV indicates the time intervals when your equipment did not function properly. For more information, see Labeling your data.

By consulting with Subject Matter Experts (SMEs) and understanding the various failure modes of the equipment you can provide a “lookahead” window indicating the amount of time the onset of the problem could have been detected.

You typically get information for labeling abnormal behavior data from two sources:

  • Work orders which have been reported on the equipment. Work orders are notoriously subjective and inconsistent. That is why with Lookout for Equipment you only need to provide the approximate time of the failure and the approximate lookahead window in the labeling data you provide to Lookout for Equipment for model training.

  • SMEs who work and maintain the equipment often have in depth knowledge about when machinery was in an erroneous state.

For information on how to apply labels when training a model, and the format of the label file, see Labeling your data.