Research On Weak Signal Processing Methodology Of Incipient Fault In The Machining Process

Manufacturing industry is the foundation of national development and a pillar industry of national economy, as well as an important embodiment of national comprehensive strength. High reliability and near-zero dowtime are the guaranty of machining process. Incipient fault identification can provide efficient maintenance decision with fault location, fault type and severity degree by identifying key information from the weak signal, which is of great significance for achieving an efficient and reliable mechanical system. However, weak signal in the incipient fault stage has characteristics of small amplitude, low signal-to-noise ratio(SNR), nonlinear, high redundancy and dimensionality, etc. Thus, weak signal processing, dominant feature selecting and nonliear feature dimensionality reducing are urgent issues needed to be solved in the incipient fault diagnosis field.For weak signal processing with low SNR in incipient fault identification, a denoising embeded sifting process was developed to remove noise during the sifting process of EMD method, as well as smooth pseudo-IMFs in iteration process, which would eventually reduce the end effect resulting from noise interferences and iteration errors. Meanwhile, to improve the stopping criterion of EMD method and eliminate redundant IMFs, correlation analysis was introduced to develop a novel stopping criterion, with the consideration of local relationship between continuous modes, as well as the correlation between pseudo-IMFs and the investigated signal. The analysis of weak signal from incipient rotor degradation system, and comparative studies with singal extension based EMD methods, demonstrated that the developed method can effectively remove end effect and eliminate redundant IMFs of EMD method, which provided theoretical and technical support for processing weak signal from incipient fault with low SNR.Weak signal with small amplitude was analyzed in the spatial-frequency domain, components in different frequency domain were extracted by the power spectral density(PSD), bidimensional wavelet transform(BWT), bidimensional empirical mode decomposition(BEMD). And a BWT denosing embeded sifting process was developed to remove errors of the envelope interpolation. The analysis of microtopography demonstrated that the developed method efficiently identified texture features of the cutting path and feeding direction, as well as surface defects and impacts on surface quality by texture features, which promisingly provided technical support for improving surface quality from the mechanism of machining process.A novel dominant feature selection(DFS) method was developed using a genetic algorithm with dynamic searching strategy, which was employed to heuristicly search the most repsentative features obtained from multiscale analysis method. The dynamic encoding strategy was developed to represent the non-linear and high dimensional features, and the feature space can be updated by a n elitist mode. Then the optimal feature sub-sets can be selected on each feature dimension. Sensitivity and specificity were calculated to evaluate classification models of different fault states by the receiver operating characteristic(ROC) method, and dominant features can be identified by a ROC indicator and cumulative frequency. The analysis of incipient faults in rotor degradation process and bearing faults, as well as comparative analysis with five other feature selection methods, demonstrated that the developed dominant feature selection method can achieve improvements in identification accuracy with a lower feature dimensionality.For the research on non-linear feature dimensionality reduction(FDR), a novel supervised locally linear embedding(LLE) method was developed based on mutual information weighted features and optimized manifolds with the consideration of overall feature information. Neighborhoods of investigated features were selected from weighted features, which reserved all the feature information and highlighted features with main contributions. Then, the maximum likelihood estimation(MLE) was employed to calculate manifold dimensionality in a low dimensional space, and the partial least squares method was used to project testing samples. Instead of employing all the manifolds as input to implement the classification task, mutual information based manifold ranking technique was proposed to select the most representative manifolds and remove redundant ones. Comparative analysis with other feature dimensionality reduction methods, and different techinique combinations, demonstrated that the improved manifold learning(ML) method can efficiently reduce feature dimensionlity, as well as improve the identification accuracy of incipient cutter failure.