Abstract:
This article describes an approach to learning gearbox operating conditions, defined by torque, rotational speed, and power, from acceleration data. Learning operating conditions paves the way to learning gearbox state-of-health because health indicators have to be normalized with respect to operating conditions to avoid false alarms. Moreover, because operational data is vastly larger than data associated with faults, representation learning is easier (and often only possible) from the operational data. The article compares two different solutions, one based on a multi-layer perceptron and the other on a recurrent network using the first four statistical moments as input features. The decision process, including heuristics and domain knowledge, used for selection of the network topology is described in detail. Models were found most effective in estimating the mechanical power transmitted through the gearbox and provided improvements over the second moment (RMS) alone.