Study On Motor Bearing Fault Diagnosis With Dictionary Learning

Fault diagnosis of motor bearings is very important to ensure the safety of industries.In this thesis,dictionary learning algorithm is used as a statistical machine learning fault diagnosis method for diagnose faults in motor bearings,with the aim to break through the limitations of traditional vibration signal analysis based methods and overcome the problems that fault features are difficult to extract and detected effectively.Starting from the basic principles of dictionary learning,the algorithm flows for two stages of sparse representation and dictionary update are explained in detail.And a deterministic method for initial dictionary selection,based on principal component analysis(PCA),is proposed,which serves as a dictionary learning optimization algorithm.The dataset is processed by PCA/KPCA to achieve low-redundancy samples and then used as an optimized initial dictionary for model establishment to improve the feature representation performance of dictionary and improve the fault diagnosis performance.Then,the dictionary learning fault diagnosis method is divided into two stages as offline dictionary learning and online fault diagnosis.Meanwhile,the reconstruction error limit is constructed and works as a judgment index for fault data detection,fault type judgment,and fault source location.In this thesis,the effectiveness of dictionary learning fault diagnosis with optimization methods are firstly verified on a simulation model of Tennessee-Eastman chemical process.The experiment results show that the dictionary model has dual performance to maintain the characteristics of data operation mode and the fault type at the same time.It shows not only the excellent modal specificity,distinguishability and fault detection adaptability,but also precise fault types detection without modes aliasing phenomenon.Moreover,the proposed algorithm presents an effective combination between principal component analysis and dictionary learning.It not only ensures the certainty and excellentness of dictionary model performance indicators,avoids fluctuations in the fault detection results caused by the random selection of the initial dictionary,but also improves the detection effect of random faults with unclear characteristics.By contrast,the introduction of kernel components enhances the performance of the dictionary model to distinguish faults.Finally,the fault diagnosis test is carried out on the motor bearing of a transformer oil pump under no-load condition and load condition respectively.The results of no-load case show that the fault detection process with dictionary learning algorithm is simple since a single index as reconstruction error limit is enough to distinguish the normal and fault data accurately.Besides,accurate fault location results are obtained because the reconstructed sample data contains the radial and axial fault characteristics completely.Furthermore,the effect of fault detection and type discrimination of the dictionary model established with PCA/KPCA optimized initial dictionary has been effectively improved,especially for the actual reduction of misjudgment rate between two bearing radial displacement fault samples with similar characteristics.In addition,the kernel initial dictionary is more suitable for the non-linear oil pump operation process,because the fluctuation range of data reconstruction error is relatively smaller,which shows obvious difference in distribution of reconstruction results among multiple fault types.Moreover,it can be concluded according to the results under load condition that the dictionary learning algorithm can realize cross-working condition detection.Both discrimination indexes and dictionary models reflect strong distinguishability while accurately representing a single kind of working condition feature.Here,the optimized dictionary learning algorithm improves the detection rate of working conditions effectively.In fault diagnosis test under changeable working conditions,the unified dictionary model with small sample capacity is used to obtain satisfactory fault detection and location results successfully.While,the optimized dictionary learning algorithm reduces the false detection rate to zero and complete fault source location for all samples.In addition,it can be seen from the distribution of observation indicators that voltage is a relatively important factor with greater impact on the changes of operating conditions during oil pump motor bearing working process.Therefore,the fault diagnosis algorithm is more sensitive to voltage changes,indicating that the dictionary learning algorithm has precious reference value and application on the process monitoring and fault diagnosis for an actual changeable working condition operation of oil pump.