|
|
Implementation and practical verification of methods for predictive identification of rolling bearings failures
Bár, Martin ; Havránek, Zdeněk (referee) ; Klusáček, Stanislav (advisor)
The aim of this thesis is to identify and classify rolling bearing failures. The first part of the thesis deals with the diagnosis of bearings using in-house measured data. Faults were introduced into two bearings. The first one was deformed and corroded. In the second bearing, the outer raceway was damaged. Vibration data was collected at regular intervals and processed in MATLAB. The values of the statistical features indicated faults in both bearings. Envelope analysis showed that both bearings had developed a fault on the outer raceway and a gradual fault on the cage. In the second part of this thesis, machine learning methods were used to classify defective bearings using the CWRU data set. The accelerometer data were divided into blocks in two ways. Vibration images were created from these blocks for convolutional neural networks (CNNs). The best prediction accuracy was achieved by 1D convolutional neural network (1DCNN) (99.2 %), followed by neural network (94.6 %) and SVM (94.4 %). Random Forest and SVM are the best methods when the training set is reduced, and among CNNs, MATLAB architecture and 1DCNN are the best. The most noise resistant method is Random Forest and neural network, and among CNNs, 1DCNN is the best. Methods using statistical features perform better than CNNs on extremely noisy data. Convolutional networks do not achieve good accuracy, which is probably due to the conversion of raw accelerometer data into vibration images.
|
|
|
Implementation and practical verification of methods for predictive identification of rolling bearings failures
Bár, Martin ; Havránek, Zdeněk (referee) ; Klusáček, Stanislav (advisor)
The aim of this thesis is to identify and classify rolling bearing failures. The first part of the thesis deals with the diagnosis of bearings using in-house measured data. Faults were introduced into two bearings. The first one was deformed and corroded. In the second bearing, the outer raceway was damaged. Vibration data was collected at regular intervals and processed in MATLAB. The values of the statistical features indicated faults in both bearings. Envelope analysis showed that both bearings had developed a fault on the outer raceway and a gradual fault on the cage. In the second part of this thesis, machine learning methods were used to classify defective bearings using the CWRU data set. The accelerometer data were divided into blocks in two ways. Vibration images were created from these blocks for convolutional neural networks (CNNs). The best prediction accuracy was achieved by 1D convolutional neural network (1DCNN) (99.2 %), followed by neural network (94.6 %) and SVM (94.4 %). Random Forest and SVM are the best methods when the training set is reduced, and among CNNs, MATLAB architecture and 1DCNN are the best. The most noise resistant method is Random Forest and neural network, and among CNNs, 1DCNN is the best. Methods using statistical features perform better than CNNs on extremely noisy data. Convolutional networks do not achieve good accuracy, which is probably due to the conversion of raw accelerometer data into vibration images.
|
guest ::
