Study On Fault Signal Extraction Based On Neural Networks And Local Linear Embed Model

Sound information based engine fault analysis has been being one of the most important diagnosis methods. But it doesn't develop fast because it hasn't combined with the modern signal processing technology very well. The main idea is getting the engine's signal by the untouched sensors, extracting the fault signals of measurable engine's signals using reasonable methods. This provides much useful knowledge for characters extraction and pattern recognition. Thus, studying on extraction of engine's fault signals is very important and significant to promote fault diagnosis schemes with untouched equipments for engine.After discovery, working sound of engine possesses the chaos characteristics. Therefore fault sound of the engine extraction methods are proposed in this paper based on the chaos background. As we know, the engine working sound signal is determined by complex and rule component under common movement interior, therefore the engine working sound signal is characterized by random, internal discipline determinism, which complies with the chaos signal. Fault sound will appear if engine breakdown, then sound obtained in the engine includes working sound (chaos signal), the noise signal and the fault signal. Because working sound and the noise signal belong to background signals, this provides the low signal-to-noise ratio environment for the fault signal extraction. Main idea in this paper is predicting the signals using the non-linear model, and then withdrawing the engine chaos signal, finally removing the noise signal with normal denosing laws from the surplus signal to obtain the pure fault sound signal. As a result, fault sound signal extraction is achieved. In this paper, all the algorithms are based on Takens Theorem. In the theorem, there are interior relationships among variables during the evolution of the dynamic system, and arbitrary element is determined by reciprocity between itself and others. Meanwhile, each one also concludes the system evolution information. Thus, in order to reconfigure an equivalent state space, only an element's change should be considered. As usual, delayed fixed point data are picked out to form a one-dimensional scale with delay method, which means single variable is mapped into multi-dimension as a vector, hence a phase space is reconfigured by single variables. However, vectors in the reconfiguration still possess the characters of original phase space.Dynamic system and chaos theory are introduced and many signal processing methods are discussed deeply. In this paper, engine sound signals without faults are analyzed carefully combined with kinds of signal processing algorithms. Then, engine sound is proved to be chaos through max Lyapunov exponent, interconnected dimension and embed dimension. This character provides proof for this paper to predict the chaos model of engine and abstract the fault sound.Reconfigured phase space method and RBF neural networks are combined to build the predict model of chaos time series. Then it is used to predict the working noise time series and isolate the fault information of the engines. Simulations between the proposed method and traditional RBF neural networks are conducted for prediction, as a result, the proposed combination model is more effective than the latter for short period prediction. Based on the estimated model, fault sounds are isolated successfully. Anyway, working time series and noise with faults are assumed to match each other approximately. In fact, it is impossible, even worse matched. Therefore, how to obtain the matched points is very significant, this is the further research in future.Local linear embed model applied into isolating fault sound of engines is studied based on manifold learning theory. In simulations, prediction and fitting ability of the above model are verified by using typical Duffing chaos signals. After then, proposed model are used to isolate the fault signals. The results show that this model is very satisfactory on Duffing signals and it can abstract the faults with the signal-noise rate–48.1702 db successfully; however, due to the measured signals with the fault, this model is not better than that of Duffing, but it still can isolate fault signals with the signal-noise rate–24.7456 db. We found that there is much measuring noise in the isolated fault signals, but it can be eliminated well with adaptively filtering method, so the pure fault signals are left. The proposed model is very useful for further fault signal identification.