A Physically Segmented Hidden Markov Model Approach for Continuous Tool Condition Monitoring: Diagnostics and Prognostics

Abstract

In this paper, a temporal probabilistic approach based on the hidden Markov model (HMM), named physically segmented HMM with continuous output, is introduced for continuous tool condition monitoring in machinery systems. The proposed approach has the advantage of providing an explicit relationship between the actual health states and the hidden state values. The provided relationship is further exploited for formulation and parameter estimation in the proposed approach. The introduced approach is tested for continuous tool wear prediction in a computer numerical control milling machine and compared with two well-established neural network (NN) approaches, namely, multilayer perceptron and Elman network. In the experimental study, the prediction results are provided and compared after adopting appropriate hyper-parameter values for all the approaches by cross-validation. Based on the experimental results, physically segmented HMM approach outperforms the NN approaches. Moreover, the prognosis ability of the proposed approach is studied.

Remaining useful life estimation methods for predictive maintenance models: defining intervals and strategies for incomplete data

“…Moreover, hidden Markov models (HMM) take into account the fact that it is not possible to measure directly the degradation of a component. Besides, these models have proven to be effective in various domains and have been used for predictive maintenance purposes in several studies (Sikorska, Hodkiewicz, & Ma, 2011) (Tobon-Mejia, Medjaher, Zerhouni, & Tripot, 2011 (Geramifard, Xu, Zhou, & Li, 2012) (Tang, Makis, Jafari, & Yu, 2015). They are based on Markov chains that represent the health state of the studied component.…”

Multi-scale hybrid HMM for tool wear condition monitoring

“…In this work, since cutting tool wear conditions (hidden states values) and complete cutting force signal in the whole tool life have been measured for the training dataset beforehand, then the transition probability, a ii , with no changes in HHMM, can be computed by explicit method [32] directly as a ii =1−1/c i , where c i is the number of samples belonging to the ith wear state. As has been mentioned, the tool wear is an irreversible left-right Markov process, and the wear value increases progressively with machining time.…”

“…All these three forces contain the information of tool wear while their sensitivities to the tool wear condition are different when applied in different methods. In [11], the thrust cutting force is found to be most sensitive when applying the singularity analysis, while Penedo et al [33] chose the twisting force Fz to monitor the tool wear by a hybrid incremental model, and in [32], all the forces in three directions were applied. In this work, since the cutting force fluctuates a lot in cutting the Inconel 718 material, its amplitude would not be applied.…”

“…As a result, each health class is, in general, multimodal with different probabilistic structure. Since the membership degree distribution over all states has been calculated, the current tool wear [32] can be estimated based on the real-valued wear state output of HHMM as…”

Multi-Sensor Condition Monitoring of Bearings Using Support Vector Machines

“…• Applicable to nonlinear dynamic systems with satisfactory performance [144,145] • Simple and efficient filtering process [62,144] • Applicable to real-time applications [144] • Requires measurement data [62,144] • Higher computation cost than the basic Kalman filtering technique [62,144] • Needs to define equations dictating system dynamic and measurement model [137,140] Hidden • Can perform fault and degradation diagnosis on non-stationary signals and random dynamic systems [115,118,137,146] • Suitable for multi-failure modes [137] • Can distinguish different types and states of bearing faults by training [65,115] • Can model different stages of degradation so failure trend does not need to be monotonic [62] • Can model spatial and temporal data [62,118] • Does not require specific knowledge of failure mechanism progression [62] • Quick management of incomplete data sets [62,118] • Provides confidence limits as part of their RUL prediction [118,147] • Strong mathematical structure and forms a solid theoretical foundation for use [147] • Ease of model interpretation [147] • Had many successful practical applications [147] • Not applicable in cases of observable failure state [137] • Requires large amount of training data, proportional to the number of hidden states, for accurate modelling [62,137,147] • Computationally intensive, particularly for a large number of hidden states [62,146] • Prognosis projection relies on a failure threshold …”