Knowledge-Based AI Model for the Detection of Pinion Wear

One common fault on rack-and-pinion drives is wear of the pinion. There are typical characteristics within the recorded sensor signals that can be used as features of an AI approach to monitor this machine failure. Potential hurdles concerning the application of AI models include the lack of labeled data as well as their acceptance and trustworthiness. To address these challenges, this paper pays special attention to the inclusion of expert knowledge and the explainability of the results. The presented approach is an Adaptive Neuro-Fuzzy Inference System (ANFIS) for the detection of pinion wear. Only a small test data set from a healthy machine is used and artificially manipulated with the basic knowledge of experts. It includes the measurements of the motor current and of an accelerometer that is mounted in the gearbox housing as part of a monitoring agent. A survey of experts is used to determine the complex relationship between features and their associated fault conditions. The manipulated dataset and the converted survey results are used for the initialization of the model parameters. It is shown, that with this main usage of domain knowledge and only a small amount of data, an ANFIS model can be set up that already gives a good and explainable estimation of the possibility of a worn out pinion on a rack-and-pinion axis.