Mohamad Khalil

Self adaptive growing neural network classifier for faults detection and diagnosis

Fault detection and diagnosis have gained widespread industrial interest in machine monitoring due to their potential advantage that results from reducing maintenance costs, improving productivity and increasing machine availability. This article develops an adaptive intelligent technique based on artificial neural networks combined with advanced signal processing methods for systematic detection and diagnosis of faults in industrial systems based on a classification method. It uses discrete wavelet transform and training techniques based on locating and adjusting the Gaussian neurons in activation zones of training data. The learning (1) provides minimization in the number of neurons depending on cost error function and other stopping criterions; (2) offers rapid training and testing processes; (3) provides accuracy in classification as confirmed by the results on real signals. The method is applied to classify mechanical faults of rotary elements and to detect and isolate disturbances for a chemical process. Obtained results are analyzed, explained and compared with various methods that have been widely investigated for fault diagnosis.

A new algorithm for spatiotemporal analysis of brain functional connectivity.

Specific networks of interacting neuronal assemblies distributed within and across distinct brain regions underlie brain functions. In most cognitive tasks, these interactions are dynamic and take place at the millisecond time scale. Among neuroimaging techniques, magneto/electroencephalography - M/EEG - allows for detection of very short-duration events and offers the single opportunity to follow, in time, the dynamic properties of cognitive processes (sub-millisecond temporal resolution). In this paper, we propose a new algorithm to track the functional brain connectivity dynamics. During a picture naming task, this algorithm aims at segmenting high-resolution EEG signals (hr-EEG) into functional connectivity microstates. The proposed algorithm is based on the K-means clustering of the connectivity graphs obtained from the phase locking value (PLV) method applied on hr-EEG. Results show that the analyzed evoked responses can be divided into six clusters representing distinct networks sequentially involved during the cognitive task, from the picture presentation and recognition to the motor response.

Parameter selection algorithm with self adaptive growing neural network classifier for diagnosis issues

Neural networks have been widely used in the field of intelligent information processing such as classification, clustering, prediction, and recognition. In this paper, a non-parametric supervised classifier based on neural networks is proposed for diagnosis issues. A parameter selection with self adaptive growing neural network (SAGNN) is developed for automatic fault detection and diagnosis in industrial environments. The growing and adaptive skill of SAGNN allows it to change its size and structure according to the training data. An advanced parameter selection criterion is embedded in SAGNN algorithm based on the computed performance rate of training samples. This approach (1) improves classification results in comparison to recent works, (2) achieves more optimization at both stages preprocessing and classification stage, (3) facilitates data visualization and data understanding, (4) reduces the measurement and storage requirements and (5) reduces training and time consumption. In growing stage, neurons are added to hidden subspaces of SAGNN while its competitive learning is an adaptive process in which neurons become more sensitive to different input patterns. The proposed classifier is applied to classify experimental machinery faults of rotary elements and to detect and diagnose disturbances in chemical plant. Classification results are analyzed, explained and compared with various non-parametric supervised neural networks that have been widely investigated for fault diagnosis.

Comparison of Different EHG Feature Selection Methods for the Detection of Preterm Labor

Numerous types of linear and nonlinear features have been extracted from the electrohysterogram (EHG) in order to classify labor and pregnancy contractions. As a result, the number of available features is now very large. The goal of this study is to reduce the number of features by selecting only the relevant ones which are useful for solving the classification problem. This paper presents three methods for feature subset selection that can be applied to choose the best subsets for classifying labor and pregnancy contractions: an algorithm using the Jeffrey divergence (JD) distance, a sequential forward selection (SFS) algorithm, and a binary particle swarm optimization (BPSO) algorithm. The two last methods are based on a classifier and were tested with three types of classifiers. These methods have allowed us to identify common features which are relevant for contraction classification.

Classification performance of the frequency-related parameters derived from uterine EMG signals

Frequency-related parameters derived from the uterine electromyogram (EMG) signals are widely used in many pregnancy monitoring and preterm delivery prediction studies. Although they are classical parameters, they are well suited for quantifying uterine EMG signals and have many advantages over amplitude-related parameters. The present work aims to compare various frequency-related parameters according to their classification performances (pregnancy vs. labor) using the receiver operating characteristic (ROC) curve analysis. The comparison between the parameters indicates that median frequency is the best frequency-related parameter that can be used for distinguishing between pregnancy and labor contractions. We conclude that median frequency can be the representative frequency-related parameter for classification problems of uterine EMG.

Methodology of wavelet packet selection for event detection

Uterine EMG is very useful for pregnancy and parturition monitoring but its analysis requires efficient tools for detection and identification of events contained in the recordings. The present work is based on the use of wavelet packet (WP) decomposition and direct use of WP coefficients (after delay correction) for change detection. The first step is to select the only WP of the decomposition tree that are able to highlight changes in the recordings using a training set of signals. In this way, one of the main contributions of the current work is to propose and test a criterion based on a model of the distribution of the estimated Kullback-Leibler distance. As WP decomposition produces a redundant tree, a second step proposes a best basis selection based on the suppression of WP without any specificity in terms of change detection. Results evidenced the efficiency of the method for simulated signals (detection probability > 95%, false alarm 98%, false alarm < 6%). Redundancy reduction suppressed half the number of WP selected in the first selection step without any degradation of the overall detection performance. Any application where events to be detected are characterized by their frequency content is a good candidate for such a methodology.

EEGNET: An Open Source Tool for Analyzing and Visualizing M/EEG Connectome.

The brain is a large-scale complex network often referred to as the “connectome”. Exploring the dynamic behavior of the connectome is a challenging issue as both excellent time and space resolution is required. In this context Magneto/Electroencephalography (M/EEG) are effective neuroimaging techniques allowing for analysis of the dynamics of functional brain networks at scalp level and/or at reconstructed sources. However, a tool that can cover all the processing steps of identifying brain networks from M/EEG data is still missing. In this paper, we report a novel software package, called EEGNET, running under MATLAB (Math works, inc), and allowing for analysis and visualization of functional brain networks from M/EEG recordings. EEGNET is developed to analyze networks either at the level of scalp electrodes or at the level of reconstructed cortical sources. It includes i) Basic steps in preprocessing M/EEG signals, ii) the solution of the inverse problem to localize / reconstruct the cortical sources, iii) the computation of functional connectivity among signals collected at surface electrodes or/and time courses of reconstructed sources and iv) the computation of the network measures based on graph theory analysis. EEGNET is the unique tool that combines the M/EEG functional connectivity analysis and the computation of network measures derived from the graph theory. The first version of EEGNET is easy to use, flexible and user friendly. EEGNET is an open source tool and can be freely downloaded from this webpage: https://sites.google.com/site/eegnetworks/.

Complexity analysis of the uterine electromyography

In respect to the main goal of our ongoing work for predicting preterm birth, we analyze in this paper the complexity of the uterine electromyography (EMG) by using the sample entropy (SampEn) algorithm. By considering recent methodological developments, we measure the SampEn over multiple scales using the wavelet packet decomposition method. The results obtained from the analyzed data indicate that SampEn decreases along pregnancy. Furthermore, we demonstrate that the computed SampEn parameter may discriminate between the two classes (pregnancy/labor). The results are supported by statistical analysis using t-test indicating good statistical significance with a confidence level of 95%. A surrogate data test is also performed to investigate the nature of the underlying dynamics of our experimental data. The results are very promising for monitoring pregnancy and detecting labor to help identify preterm labor.

Classification of multichannel uterine EMG signals

Classification of multichannel uterine electromyogram (EMG) signals is addressed. Signals were recorded by a matrix of 16 electrodes. First, signals corresponding to each channel were individually classified using an artificial neural network (ANN) based on radial basis functions (RBF). The results have shown that the classification performance varies from one channel to another. Then, a decision fusion method based on these classification performances was tested. After fusion, the network yielded better classification accuracy than any individual channel could provide. The high percentage of correctly classified labor/non-labor events proves the efficiency of multichannel recordings in detecting labor. These findings can be very useful for the aim of classifying antepartum versus labor patients.

Driver stress level detection using HRV analysis

This paper intends to investigate stress level detection of a driver during real world driving experiment. This detection is based on heart rate variability (HRV) analysis which is derived from ECG signal and reflects autonomic nervous system state of the human body. The alteration of autonomic nervous system predicts the stress level of drivers during driving operation and permits a safe driving by the possibility of an early warning. This stress, taking place during driving, is caused by diverse factors such as changing mood, bio rhythm, fatigue, boredom or disease which can prevent the driver from reaching inappropriate state for driving. In our study, the ECG signal of the driver is extracted and preprocessed in order to perform the HRV analysis. This analysis is accomplished using one of the domain analysis approach such as time, frequency, time-frequency or non-linear methods including Wavelet and STFT. After HRV analysis, several parameters are extracted to build a vector of features for the classification phase. Our experimentation is performed with data issued from 16 different subjects from the Stress Recognition in Automobile Driver database (DRIVEDB). Several classification techniques were investigated including support vector machine with radial basis function (SVM-RBF) kernel, K nearest neighbor (KNN), and radial basis function (RBF) classifiers. Our results indicate that stress detection could be predicted with an accuracy of 83% using SVM-RBF classifier. This also points out the robustness of ECG biometric as an accurate physiological indicator of a driver state.