Continuous Hidden Markov Model Based Gear Fault Diagnosis and Incipient Fault Detection

Wear fault is one of the failure modes most frequently occurring in gears. Identifying different wear levels, especially for early wear is a challenge in gear fault diagnosis. In this paper, a machine learning methodology for performing diagnosis is presented. It is developed based on hidden Markov model (HMM). A HMM is defined as a doubly stochastic process. The underlying stochastic process is a discrete time finite-state homogeneous Markov chain. The state sequence is not observable and so is called hidden. HMM's construction process and the condition monitoring process are same. So, the HMM can be used to model the mechanical fault progress process. There are some authors applying HMM in machine fault diagnosis. However, HMM with different observation styles applying in machine fault diagnosis will produce different results. With enough history data, this paper aims to using these different HMMs to classify the different levels of gear wears automatically and analysis the performance of these models. These observation styles are discrete integer observation which is the quantized fault features extracted from the vibration signals, continuous observation which is the feature with no quantizing extracted from the vibration signals and observation which is the preprocessed vibration signals. The gear wear experiments were conducted and the vibration signals were captured from the gears under different loads and motor speeds. HMM with three different styles observation is applied to identify the gear wears and the applied results are analyzed. The results show that the features with no quantizing as the observations of the HMM for diagnosis perfected best. But, at the case of no history data, HMM which the observations are the preprocessed vibration signals is used to do the incipient fault detection. The result shows the validity of the method.