Classification And Recognition Of Transmission Fault Based On Deep Belief Network

Essentially, the classification and recognition of transmission fault is a process of recognizing the fault states through machine learning approach. Generally, there are several crucial steps from the unknown original fault states to learned identifiable fault states, such as fault feature extraction, fault feature selection and optimization, classifier design and result evaluation. However, the characteristics of mechanical fault are closely associated with the properties of fault by itself. It is relatively cumbersome, burdensome and labor elapsing to acquire the appropriate features. Furthermore, it greatly increases difficulty and uncertainty of feature extraction and optimization due to human involvement, and increases the difficulty of the transmission mechanical fault recognition, which weakens the intelligence of machine learning.In this paper, the concept of deep learning is introduced, and deep belief network is applied in mechanical fault recognition as a representative. When using the traditional statistical features of vibration signal as the network input to recognize bearing’s states with different types of failure and different degrees, the method utilizing deep belief network can obtain a very high rate of correct recognition which proves deep belief network has a strong feasibility in mechanical fault intelligent recognition.Deep belief network is a typical deep learning method, which can achieve an abstract representation through the combination of low-level features to discover the distributed characteristics. We can deduce the deep belief reconstruction network base on deep belief network. By reconstructing the simulated signal analysis, there are small differences between the reconstructed signal and the original signal, which indicates deep belief network can recover the original data from higher data with low distortion. In other words, the higher features can represent the low-level signal to some extent and deep belief network has the ability to keep the key information of original data.Given the characteristics of deep belief network, this paper proposes a novel method to recognize the fault states directly from the raw data. This method can avoid the artificial feature extraction and optimization process, reducing human involvement, and enhances the intelligence of transmission fault recognition. When using the original data as the network input, the novel method will develop a relatively large amount of data, which correspondingly require a great amount of computation. The total number of the algorithm calculations and its partial derivatives can be calculated based on the time complexity. Preference to select the parameter with larger partial derivative value as the adjusting parameter, the computational cost of deep belief network can be better depressed.After bearing fault and transmission complex fault recognition experiments, it proves that directly from the raw data deep belief network can effectively recognize the fault states with high accurate rate and further enhance the fault recognition intelligence.