Sensor measurements/vibration data from five test rigs were used (shown below) to design a health monitoring system of a rotating machine through exracting informative features from the raw (sensor) data along with evaluating the data in frequency and time-domain.
- Fault 1: Bearing Fault
- Fault 2: Gear Mesh
- Fault 3: Imbalance
- Fault 4: Misalignment
- Fault 5: Resonance
The raw data was initially normalized and then processed further to deduce the power spectral density for each individual features with appropriate filter (Butterworth) implemented
Principal Component Analysis, PCA, method was implemented to reduce the number of dimensions/features and for visualization purposes.
1-Nearest Neighbor Algorithm was implemented using the Euclidean distance measure. KNN algorithm clasifies the new/upcoming measurement based on the class of one of its nearest neighbours. In other words, the algorithm implemented finds the similarity between the training measurements and the new/test measurements.
Following assumptions were made:
The algorithm had an accuracy of 98.66% in determining and classifying faults via frequency data
Multisensor signal estimation and health monitoring system was designed and implemented for a wind turbine manufacturing company with prior knowledge of the pitch angle being uniformly distributed in the range 0° <= w <= 30° and the sensor noise being normally distributed V ~ N(0,9).
MMSE Estimator was implemented by taking Bayes' Theorem into consideration to predict the actual measured data (without noise) of wind turbine blades. Additionally, CUSUM two-sided test was also implemented to identifiy if the measured data exceeds the desired threshold and generate an alert.
A report was generated highlighting important findings and critically evaluating the results.