Abstract:
The idea of smart factories has aroused with the emergence of Industry 4.0. A way to contribute to the construction of smart factories is to develop intelligent main tenance strategies which enable factories to reduce maintenance costs and keep them away from catastrophic damages. Predictive maintenance warns users regarding the need for maintenance of assets at a specific moment. In this thesis, predictive mainte nance strategies for Flexible Impeller Pumps (FIP) are developed by using supervised and unsupervised learning methods. After observation of real textile dye houses, an emulation setup has been designed to collect data from real use cases of FIPs. Data consisting of three conditions, healthy, looseness, and cavitation, have been collected via an acoustic sensor involving 3 mi crophones. Obtained time-series signals are converted to informative features which are utilized for training a supervised model, convolutional neural network (CNN). The trained model is able to generate a warning for a specific failure state. In addition to the supervised model, an unsupervised model has also been trained for anomaly de tection of FIPs via extracted features of healthy data due to the applicability issues of unhealthy state data acquisition. The model can produce a warning when an anomaly is detected but it cannot specify the failure mode. Both models give satisfactory re sults for classifying failures and detecting anomaly situations of FIPs. Mel Frequency Cepstrum Coefficients (MFCC) have shown superiority among the features extracted to confirm human hearing based modeling can be applied successfully in maintenance methods.
