ER4165 Advanced Topics in Machine Intelligence
This design study is to give you an insight into selected aspects of machine intelligence (MI) applied to detection and classification of faults in a mechatronic system. You are asked to solve various tasks related to a design of a MI system, including features selection, dimensionality reduction, classification and the MI system quantitative evaluation.
This design study will test your ability to:
• Design methods and processes necessary for deployment of an artificial intelligence system for quantitative evaluations of a mechatronic system.
• Recognize software design challenges behind implementations of machine learning algorithms.
• Design and optimise software to meet specified requirements.
• Design and provide a working solution for faults detection in a transmission system.
The objective of this design study is to conceive a system for detecting defects in a transmission system. Three types of defects are considered: gear tooth, sensor drift and shaft wear faults. The transmission system was modelled using Simulink. For various configuration and severity of the faults, a vibration of the transmission casing was recorded for the further analysis. For the purpose of this assignment 1,000 simulations were executed with eight possible different combinations of the three faults, including simulations without any faults of the transmission system. Subsequently, 18 different features calculated from the time and frequency domains of the casing vibration signal were recorded. All the results have been logged in the TrainingData table. The corresponding TrainingData.mat file is available from the Blackboard. The TrainingData table consists of 1000 rows, representing each experiment, and 22 columns. The first 18 columns contain extracted vibration signal features. Subsequent three columns (19-21) provide a ground truth labels annotating presence of each of the three faults (SensorDrift, ShaftWear, ToothFault), with ‘true’ (logical 1) indicating a presence of the defect and ‘false’ (logical 0) indicting an absence of the defect. The last, 22nd column (faultCode) encodes one of the eight possible defect combinations present in the given experiment, calculated as:
The TrainingData table contains annotations of each extracted feature and fault ground truth. The additional TestData.mat file is also available from the Blackboard. That file contains the TestData table consisting of 20 additional simulation results, with recorded values of the 18 features. However, although the fault ground truth labels are known for the TestData, you will not have access to that information before the submission deadline.
You are asked to design a system for recognition of the faults present in the transmission system. In your design you should: consider a need for feature selection or feature dimensionality reduction; if necessary, apply suitable feature selection/reduction algorithm; select a machine learning (MI) algorithm suitable for the task; training of the selected MI algorithm; quantitatively evaluate the design system.