To improve the efficiency of feature extraction from mechanical vibration signal of circuit breaker

and to avoid the problems that no-training samples of unknown type are recognized as the normal samples or wrong types of known fault

we propose a new method for monitoring and diagnosing the mechanical state fault of circuit breaker. Firstly

the vibration signal of circuit breaker’s mechanism operating system is temporal segmented

and 7 features are extracted from each part of the divided signal to construct different kinds of feature vector. The Tsallis entropy is used to analyze the mechanical fault of circuit breaker by scatter matrix. Then

the feature vector is input to the hierarchical hybrid classifier based on one-class support vector machine（OCSVM） and extreme learning machine（ELM） for fault diagnosis. In the hybrid classifier

the state of normal and fault is recognized by OCSVM at first. If the sample is identified as fault state

the fault types of training samples are identified using ELM classifier. And then

the recognition results are corrected by OCSVM. The experiments carried out on the real circuit breaker prove that the scatter matrix can effectively analyze the class separability of features and the temporal segmented method is efficient to extract features. Hierarchical hybrid classifier can not only identify the mechanical state and fault type of circuit breaker accurately

but also identify the unknown fault type without training samples effectively.