Predictive Maintenance Mechanism with Vibration Measurements and Support Vector Machines
Koivu, Aki (2017-08-15)
Predictive Maintenance Mechanism with Vibration Measurements and Support Vector Machines
Koivu, Aki
(15.08.2017)
Tätä artikkelia/julkaisua ei ole tallennettu UTUPubiin. Julkaisun tiedoissa voi kuitenkin olla linkki toisaalle tallennettuun artikkeliin / julkaisuun.
Turun yliopisto
Kuvaus
Siirretty Doriasta
Tiivistelmä
When a process in an industry environment is implemented as complex machinery, there is a need for their maintenance. Availability and reliability of an online system are the major areas of interest, when we consider the health of a machine. For machinery whose operation involves movement, resonance and vibration analysis is a commonly used method to monitor health. The analysis is conducted with values collected from a vibration sensor.
Machine learning algorithms have been developed for detecting abnormality within data. This detection can be extended to vibration measurement data when properly prepared. With preprocessing methods such as the Fourier transform, vibration can be transformed into features, that can be mapped into classes of abnormal and normal state of the machine. Commonly if there has been no reason to collect vibration data from a device before, it does not exists. It can also be that data can also be available in increments, a device can be measured once every couple hours for example. These types of cases are far too common in industry, and to accommodate this, specific machine learning methods have to utilized in order to learn characteristics of a vibration signal without a priori data, and new data becoming available over time. To answer these real life problems, one-class online support vector machine algorithm is utilized.
This thesis was done in collaboration with the PerkinElmer Turku site, by applying the research question to their GSP® instrument's HV-pipette module. The GSP® is a diagnostic analyzer instrument for newborn screening, and the HV-pipette is a crucial part of the overall system. By constructing a vibration measurement setup prototype and taking measurements with an accelerometer of the HV-pipette while in operation, measurements were recorded that were experimented upon to create a predictive support vector machine algorithm. The results were evaluated with PerkinElmer engineers, and the model was taken in for future experimentation and validation.
Machine learning algorithms have been developed for detecting abnormality within data. This detection can be extended to vibration measurement data when properly prepared. With preprocessing methods such as the Fourier transform, vibration can be transformed into features, that can be mapped into classes of abnormal and normal state of the machine. Commonly if there has been no reason to collect vibration data from a device before, it does not exists. It can also be that data can also be available in increments, a device can be measured once every couple hours for example. These types of cases are far too common in industry, and to accommodate this, specific machine learning methods have to utilized in order to learn characteristics of a vibration signal without a priori data, and new data becoming available over time. To answer these real life problems, one-class online support vector machine algorithm is utilized.
This thesis was done in collaboration with the PerkinElmer Turku site, by applying the research question to their GSP® instrument's HV-pipette module. The GSP® is a diagnostic analyzer instrument for newborn screening, and the HV-pipette is a crucial part of the overall system. By constructing a vibration measurement setup prototype and taking measurements with an accelerometer of the HV-pipette while in operation, measurements were recorded that were experimented upon to create a predictive support vector machine algorithm. The results were evaluated with PerkinElmer engineers, and the model was taken in for future experimentation and validation.