A Bearing Fault Diagnosis Method Based On Generalized S Transform And SSTD-2DPCA

Rotating machinery condition is usually monitored and diagnosed based on vibrationsignal parameters. Due to unstable speed, load change and lots of impact and friction causedby fault, vibration signals often exhibit strong nonlinearity and non stationarity. Thetime-frequency analysis can convert one-dimensional vibration signal in time domain totwo-dimensional spectrum in time frequency domain. Since time-frequency images containrich state information, the status of equipment can be interpretated through these images. Butthere exisits two problems: First, because of the complexity of vibration spectrum, thetraditional diagnostic methods usually need technical personnel with rich experience inengineering, and the diagnosis efficiency is low. Besides, the diagnosis results of differentexperts also differ from man to man; Second, since image is stored in the form of matrix witha huge dimension, image recognition technology based on artificial intelligence often needs tovectorize the image matrix, which will not only destory image information, but also cause thecurse of dimensionality.This paper applies two-dimensional principal component analysis to time-frequencyfeature extraction of image and puts forward a supervised and semi-supervised bidirectionaltwo-dimensional principal component analysis method based on integrated matrix distance.The supervised method does not need to vectorize the image matrix, and its feature extractionefficiency is higher than the two-dimensional non-negative matrix factorization method, thediagnostic performance is higher than the two-dimensional linear discriminant analysismethod; with a few known image samples, semi-supervised method can achieve a largenumber of unknown image recognition, and its recognition performance is better than thesupervised method, but the diagnosis efficiency becomes low.According to the experiment, this paper applies bidirectional two-dimensional principalcomponent analysis to bearing state recognition under different failure, different fault degreemodes. The time-frequency image is obtained by generalized S transform and the Stankovicmethod and Djurovic method is compared to improve the energy concentration of thegeneralized S transform. Then a fusion of the above method, which optimizes the parameterin frequency segmentation, is presented. Simulation is carried out to demonstrate the proposedmethod. Then, the optimization method is applied to analyse the actual bearing signal. Resultsshow that the proposed method has certain applicability in practical signal. Finally, theinfluence of different matrix distance metric, time-frequency image source, feature extraction, dimension and other factors on the supervised algorithm is analyzed; the influence of thenumber of training samples and iteration on semi-supervised algorithm is analyzed.