Omid Geramifard

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

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.

Data-driven approaches in health condition monitoring — A comparative study

In this paper, four data-driven classification approaches, that is, K-nearest neighbors (K-NN), self-organizing map (SOM), multi-layer perceptron (MLP), and Bayesian Network classifier (BNC), are applied to a health condition monitoring problem for the wearing cutter. The dataset is produced from a cutting machine using force sensing. A genetic algorithm (GA) based search is performed to select 3 dominant features from a 16-dimensional feature space, which is computed directly from the real dataset. Subsequently K-NN, SOM, MLP, and BNC algorithms are trained to predict the wearing status of the cutter, respectively. The suitability of the four data-driven approaches for the health condition monitoring are investigated and compared.

Fault detection and diagnosis in synchronous motors using hidden Markov model-based semi-nonparametric approach

Early detection and diagnosis of faults in industrial machines would reduce the maintenance cost and also increase the overall equipment effectiveness by increasing the availability of the machinery systems. In this paper, a semi-nonparametric approach based on hidden Markov model is introduced for fault detection and diagnosis in synchronous motors. In this approach, after training the hidden Markov model classifiers (parametric stage), two matrices named probabilistic transition frequency profile and average probabilistic emission are computed based on the hidden Markov models for each signature (nonparametric stage) using probabilistic inference. These matrices are later used in forming a similarity scoring function, which is the basis of the classification in this approach. Moreover, a preprocessing method, named squeezing and stretching is proposed which rectifies the difficulty of dealing with various operating speeds in the classification process. Finally, the experimental results are provided and compared. Further investigations are carried out, providing sensitivity analysis on the length of signatures, the number of hidden state values, as well as statistical performance evaluation and comparison with conventional hidden Markov model-based fault diagnosis approach. Results indicate that implementation of the proposed preprocessing, which unifies the signatures from various operating speeds, increases the classification accuracy by nearly 21% and moreover utilization of the proposed semi-nonparametric approach improves the accuracy further by nearly 6%.

Continuous health condition monitoring: A single Hidden Semi-Markov Model approach

In this paper, a single Hidden Semi-Markov Model (HSMM) approach is introduced for continuous health condition monitoring in machinery systems. Contrary to previous attempts in using hidden Markov models in this area which have not provided the relationship between the hidden state values and the physical states, this method provides the aforementioned relationship. In this paper, HSMM is applied as the core model being used in the method in order to increase flexibility of our previously used HMM-based method and consequently its generalization capability. The newly introduced method is compared with our initial HMM-based method which previously outperformed the conventional Artificial Neural Networks approach. Results show that the additional flexibility provided in the new method has improved the performance. As an example, the proposed method is used for tool wear prediction in a CNC-milling machine and results of the study is provided. 482 features are extracted from 7 signals (three force signals, three vibration signals and Acoustic Emission) acquired for each experiment of our dataset. These features include, 48 statistical features extracted from force signals in three directions (16 from each force signal) and 434 averaged wavelet coefficients from all seven signals (62 from each signal). After feature extraction phase, Fisher Discriminant Ratio is applied to find the most discriminant features to construct the prediction model. 38 features out of 482 extracted features are selected to be used in the prediction models. The prediction results are provided for three different cases i.e. cross-validation, diagnostics and prognostics.

Continuous health assessment using a single hidden Markov model

In this paper, two temporal models, Hidden Markov Model and Auto Regressive Moving Average model with exogenous inputs (ARMAX), are used for health condition monitoring of the cutter in a milling machine. Dataset is acquired through real time force signal sensing. A heuristic statistical approach is used to select dominant features, leading to the selection of 3 dominant features from the 16-dimensional feature space. Subsequently Hidden Markov Model and ARMAX model have been trained to predict the wearing status of the cutter in the milling machine. Suitability of these approaches are investigated and compared.

An HMM-based semi-nonparametric approach for fault diagnostics in rotary electric motors

In this paper1, a semi-nonparametric approach based on hidden Markov model (HMM) is introduced for fault diagnostics in the rotary electric motors. The introduced approach uses multiple HMMs to capture various underlying trends for each probable fault in the electric motors. In this work, only two major faults in the rotary motors, namely, bearing faults and unbalanced rotor are tried to be distinguished from the health condition. The experimental results are provided for single HMM for each fault, multi HMMs for each fault and multi-HMMs using semi-non parametric approach to recognize the faults.

Feature selection for tool wear monitoring: A comparative study

One of the challenging tasks in the domain of Tool Condition Monitoring (TCM) is feature selection. Feature selection is crucial as extracting all possible features and creating a model based on those features results in two major disadvantages, i.e. high computational cost and inefficient complexity of the model, which leads to overfitting. In this paper, four statistical feature selection methods are applied to the TCM problem in a CNC-milling machine. These methods are Ridge Regression (RR), Principal Component Regression (PCR), Least Absolute Shrinkage and Selection Operator (LASSO), and Fishers Discriminant Ratio (FDR). Applicability of these methods are compared based on their diagnostic results in two cases using a single Hidden Markov Model (HMM) approach.

A multi-modal hidden Markov model based approach for continuous health assessment in machinery systems

In this paper1, a multi-modal approach based on the single hidden Markov model (HMM) with continuous output is introduced for continuous health condition monitoring in machinery systems. Comparing with existing approaches such as single HMM-based approach, artificial neural networks (ANN) approach, auto-regressive moving average with exogenous inputs (ARMAX), the proposed approach improves the performance of health condition monitoring (HCM) by using multiple HMM models in parallel. Each model emphasizes on different regiments, and outputs of all models are integrated as the ultimate output. The integration of HMM outputs are conducted by either a parametric or a semi-nonparametric hindsight method. The proposed approach is applied to tool wear prediction of a CNC-milling machine, and results are compared with an existing HMM-based approach.

Power-signature-based Bayesian multi-classifier for operation mode identification

In this paper, a power-signature-based Bayesian multi-classifier is proposed to identify various operational modes of a complex machinery system that can help determine the energy contribution of different operation modes, identify potential energy hot-spots and provide basis for more accurate energy consumption calculation. This technology can also help process experts and managers to perform the process optimization from an energy saving point of view, and benchmark the energy efficiency of the processes. Based on our experimental results on an Engel injection molding machine, our proposed approach can successfully classify its operation modes to an acceptable extent based on its electrical power signatures.

Multi-model diagnostics for various machining conditions: A similarity-based approach

In this paper, a similarity-based multi-model approach based on a pre-existing physically segmented hidden Markov model with continuous output (PSHMCO) is proposed for diagnostics and tool wear monitoring. The proposed approach helps to improve the estimation accuracy in cases that the machinery system undertakes different operating (machining) conditions. The proposed multi-model approach is compared with its single model variant on a tool wear monitoring dataset with various machining conditions. The results indicate that using a similarity function to identify and apply the most similar model out of multiple models can significantly improve the prediction performance compared to blindly utilizing a single general model while adequate training data is available.