Hidden Markov Model-Based Fault Detection Approach for a Multimode Process

Abstract

Many industrial processes possess multiple operational modes and transitions because of various production factors, which pose a challenge to conventional fault detection methods. In this article, a novel fault detection scheme based on a hidden Markov model (HMM) is presented for multimode processes with transitions. To begin with, measurement data of stable modes and transitional modes are separated. Then, hidden state probability integration strategy is developed to combine local monitoring results into two global indices in a probabilistic manner. These two indications work together for stable mode process monitoring. Further a new HMM is built for transition process modeling. The Bayesian information criterion (BIC) is responsible for model evaluation. After an appropriate model is acquired, an index named negative log likelihood probability is employed for transition process fault detection. In the end, a numerical simulation example and the Tennessee Eastman chemical process is utilized to show that our proposed approach is effective.

Multimode complex process monitoring using double‐level local information based local outlier factor method

“…Traditional LOF, NSLOF, and the proposed DLI‐LOF method are applied to the numerical simulation. To reduce the influence of the variable scales, z ‐score technique is applied in LOF method suggested by Ma et al In LNS strategy, 50 neighbors are determined for data reprocessing according to the literature . And the LNS result is shown in Figure , from which the multimodality can be efficiently eliminated.…”

“…Based on Equation 23, a total of 400 samples under normal condition are generated as training dataset which contain 200 samples from each mode. For testing, 2 fault cases are designed as follows, and each fault dataset includes 400 samples.…”

“…Deng et al employed the PCA similarity factor between the current data window and its local neighborhood data window to monitor process status. Several different multimode monitoring approaches are also proposed and obtain satisfactory monitoring performances …”

“…Several different multimode monitoring approaches are also proposed and obtain satisfactory monitoring performances. [22][23][24][25][26] However, most aforementioned multimode process monitoring algorithms are based on the assumption that the process variables within each individual operating mode should follow a multivariate Gaussian distribution. In fact, the process variables cannot conform to the strict Gaussian distribution but follow the non-Gaussian distribution or the mixture distribution of Gaussian and non-Gaussian.…”

Stochastic Modeling Technology for Grain Crops Storage Application : Review

“…Hidden Markov (HMMs), Artificial Neural Networks (ANNs), and Markov Chains (MC) models are popular tools for modelling dependent random variables in diverse areas [5] such as speech processing and enhancement [6], audio segmentation [7], DNA recognition [8] , fault [9], and rainfall occurrence [10]. These are based on a stochastic process [11] in which a chain produces an unobservable state that can be inferred only through another set of stochastic process.…”