Optimization of sEMG Classification Model Based on Correlation Analysis and Feature Selection

PurposePenetration state is one of the most important factors for judging the quality of a gas tungsten arc welding (GTAW) joint. The purpose of this paper is to identify and classify the penetration state and welding quality through the features of arc sound signal during robotic GTAW process.Design/methodology/approachThis paper tried to make a foundation work to achieve on‐line monitoring of penetration state to weld pool through arc sound signal. The statistic features of arc sound under different penetration states like partial penetration, full penetration and excessive penetration were extracted and analysed, and wavelet packet analysis was used to extract frequency energy at different frequency bands. The prediction models were established by artificial neural networks based on different features combination.FindingsThe experiment results demonstrated that each feature in time and frequency domain could react the penetration behaviour, arc sound in different frequency band had different performance at different penetration states and the prediction model established by 23 features in time domain and frequency domain got the best prediction effect to recognize different penetration states and welding quality through arc sound signal.Originality/valueThis paper tried to make a foundation work to achieve identifying penetration state and welding quality through the features of arc sound signal during robotic GTAW process. A total of 23 features in time domain and frequency domain were extracted at different penetration states. And energy at different frequency bands was proved to be an effective factor for identifying different penetration states. Finally, a prediction model built by 23 features was proved to have the best prediction effect of welding quality.

Taste sensation recognition is valuable for developing virtual taste and the early diagnosis of taste system disorders. Electromyography (EMG) can characterize the information of taste sensation and have the advantages of a high signal-to-noise ratio and easy acquisition. This paper presents feature extraction and feature combination methods for taste sensation knowledge based on facial EMG. The time domain features, frequency domain features, time-frequency domain features, and entropy features are extracted as individual feature sets after data pre-processing. Combinations of different features are performed to acquire multivariate feature representations. Machine learning methods are used to predict taste sensation categories. The combination of the frequency domain, time-frequency domain, and entropy features as feature set and Random Forest method as classifier obtain the highest recognition accuracy. The results provide a reference for feature extraction and selection for basic taste sensation recognition.

During the long-term operation of wind turbines, due to environmental factors and equipment aging, the health and reliability of each component will gradually decline, leading to failure. To assess the health status of wind turbines, timely grasp the subsequent changes and development trends, it is necessary to extract degradation characteristics, including time domain, frequency domain, and time-frequency domain characteristics. These degradation characteristics can reflect the operating status of the equipment, help build health indicator curves, and evaluate the health status of high-speed shaft bearings of wind turbines. Selecting reasonable degradation characteristics is an important prerequisite for constructing a health index curve, and using evaluation indicators to construct a comprehensive evaluation function to screen degradation characteristics. The feature fusion method based on a self-organizing feature mapping network is used to fuse multiple selected degradation features and fuse the selected multiple degradation features into a curve that can reflect the bearing degradation process. Finally, a quantitative analysis is performed on the health index curve to scientifically assess the health status of bearings. Bearings are one of the key components of wind turbines. Based on the health index curve constructed in this article, an appropriate prediction model is selected to predict the health index trend of bearings. A timely and effective grasp of the health trends of wind turbine bearings is of great practical significance for formulating scientific and reasonable maintenance measures for wind farms. The work of this article will be divided into the following four parts: (1) Extracting the time domain, frequency domain, and time-frequency domain degradation characteristics of high-speed shaft bearing vibration signals of wind turbines; (2) Comprehensive evaluation using monotonicity, correlation, and robustness constructs function to screen degenerate features; (3) Use self-organizing feature mapping network. The network integrates the selected degradation features and constructs a health index curve; (4) Based on the constructed health index curve, optimize the BiLSTM network hyperparameters through Bayesian and establish a BO-BiLSTM network model to achieve a more accurate and scientific prediction of the health index trend.

Roller bearings are commonly used components in rotating machinery and are pruned to be failure, which may cause huge economic losses and even casualties. Vibration signals of rolling bearings are signs of rotating machinery. Therefore, they can be used to describe the process of performance degradation. Unfortunately, the signals exhibit non-stationary and non-linear as well as highly non-monotonic behaviors as the bearing condition degrades especially under complex working conditions. In this paper, a method to construct a constitution health index based on vibration signals is presented. Firstly, 40 features in time domain, frequency domain, time-frequency domain and entropy domain are extracted individually. Then, the combined evaluation index is constructed based on the above 40 features in terms of monotonicity, robustness, predictability and correlation. Finally, by comparing the degradation fusion indexes of rolling bearings under different operating conditions with the traditional characteristics, the remaining useful life prediction method based on grey correlation analysis is proposed. The results suggest that the proposed method can effectively predict the RUL of rolling bearings with an acceptable degree of accuracy up to 90%, which outperforms the traditional features.

In the study of detection of an epileptic seizure using Electroencephalogram (EEG), pattern recognition has been recognized as a valued tool. In this pattern recognition study, the first time the authors have attempted to use time domain (TD) features such as waveform length (WL), number of zero-crossings (ZC) and number of slope sign changes (SSC) derived directly from filtered EEG data and from discrete wavelet transform (DWT) of filtered EEG data for the detection of an epileptic seizure. Further, the authors attempted to study the performance of other time domain features such as mean absolute value (MAV), standard deviation (SD), average power (AVP), which had been attempted by other researchers. The performance of the TD features is studied using naive Bayes (NB) and support vector machines (SVM) classifiers for the university of Bonn database with fourteen different combinations of set E with set A to D. The proposed scheme was also compared with other existing scheme in the literature. The implementation results showed that the proposed scheme could attain the highest accuracy of 100% for normal eyes open and epileptic data set with direct as well as DWT based TD features. For other data sets, the highest accuracy is obtained with DWT based TD features using SVM. However, no significant difference in the classification of 14 data sets with TD features filtered EEG data and from DWT of filtered EEG data.

The contemporary diagnosis of epileptic seizures is dominated by non-invasive EEG signal analysis and classification. In this paper, we propose a patient-specific seizure detection technique, which selects the optimal feature subsets and trains a dedicated classifier for each patient in order to maximize the classification performance. Our method exploits time domain, frequency domain, time-frequency domain and non-linear feature sets. Then, by using Conditional Mutual Information Maximization (CMIM) as the feature selection method the optimal feature subset is chosen over which the Support Vector Machine is trained as the classifier. In this study, both train and test sets contain 50% of seizure and non-seizure segments of the EEG signal. From the CHB-MIT Scalp benchmark EEG dataset, we used the EEG data from four subjects with overall 21 hours of recording. Support Vector Machine (SVM) with linear kernel is used as the classifier. The experimental results show a delicate classification performance over the test set: i.e., an average of 90.62% sensitivity and 99.32% specificity are acquired when all channels and recordings are used to form a composite feature vector. In addition, an average of 93.78% sensitivity and a specificity of 99.05% are obtained using CMIM.

This paper investigates the applicability of time and time-frequency features\nbased classifiers to distinguish internal faults and other transients -\nmagnetizing inrush, sympathetic inrush, external faults with current\ntransformer saturation, and overexcitation - for Indirect Symmetrical Phase\nAngle Regulating Transformers (ISPAR). Then the faulty transformer unit\n(series/exciting) of the ISPAR is located, or else the transient disturbance is\nidentified. An event detector detects variation in differential currents and\nregisters one-cycle of 3-phase post transient samples which are used to extract\nthe time and time-frequency features for training seven classifiers. Three\ndifferent sets of features - wavelet coefficients, time-domain features, and\ncombination of time and wavelet energy - obtained from exhaustive search using\nDecision Tree, random forest feature selection, and maximum Relevance Minimum\nRedundancy are used. The internal fault is detected with a balanced accuracy of\n99.9%, the faulty unit is localized with balanced accuracy of 98.7% and the\nno-fault transients are classified with balanced accuracy of 99.5%. The results\nshow potential for accurate internal fault detection and localization, and\ntransient identification. The proposed scheme can supervise the operation of\nexisting microprocessor-based differential relays resulting in higher stability\nand dependability. The ISPAR is modeled and the transients are simulated in\nPSCAD/EMTDC by varying several parameters.\n

Atrophy is one of the most common consequences of muscle disorder. This could be a result of both myopathy and neuropathy. Muscle atrophy becomes more possible as people age. As a result of this disorder, the amount and size of muscle fibers decrease, therefore a person cannot produce high amount of force in his/her muscles, leading to difficulties in handling daily activities. The main purpose of this research is to find a way to predict this disorder. In this study the force classification was used for the atrophy disorder detection. The results show that different classifiers and features from the proposed ones, work for this purpose. To approach this goal, data were collected by recording surface EMG (sEMG) signals. Processing the recorded signals, best features with respect to more accuracy and less calculation complexity were selected and reported. After extracting the features from each patient with using different types of classifiers including LDA (Linear Discriminant Analysis), QDA (Quadratic Discriminant Analysis) and SVM (Support Vector Machine), the best approach to separate normal and atrophic people was investigated. It is found that unlike the proposed features such as MAV (Mean Absolute Value), SSC (Slope Sign Change) and WL (Waveform Length) in upper limb movement classification, three features WL, WAMP (Wilson Amplitude) (time domain features) and MNP (Mean Power) (frequency domain feature) show better performance for atrophy characterization. The results show that these features well predict the detection of biceps atrophy.

To improve human-machine cooperation and enhance the accuracy of gait recognition in wearable lower limb exoskeletons, an enhancement method of gait recognition based on adaptive feature selection and an optimized machine learning algorithm was proposed. In this study, surface electromyography (sEMG) signals of rectus femoris, medialis femoris, lateralis femoris, semitendinosus and biceps femoris were recorded during level-ground walking. Then, time domain, frequency domain, time-frequency domain and nonlinear features were extracted. The integrated value of electromyography, variance, root mean square and wavelength were selected as the time domain features, and the mean power frequency and median frequency were selected as the frequency domain features. Wavelet packet energy was selected as the time-frequency domain feature. Nonlinear features including approximate entropy, sample entropy and fuzzy entropy of sEMG were extracted. To reduce feature dimension and improve the calculation efficiency, adaptive feature selection was performed by particle swarm optimization combined with sigmoid function. Then, the feature matrix was determined as the input for a random forest classifier to recognize different gait phases. An adaptive optimization mechanism based on Bayesian optimization was applied to random forest. Compared with random forest, the overall performance of the optimized model was improved. Its accuracy was increased by 3.57%. The feature selection and adaptive optimization mechanisms in gait recognition not only enhance the accuracy of random forest algorithms applied to sEMG for gait prediction, but also facilitate the flexibility and consistency required for lower limb exoskeleton gait control.

B-factor is highly correlated with protein internal motion, which is used to measure the uncertainty in the position of an atom within a crystal structure. Although the rapid progress of structural biology in recent years makes more accurate protein structures available than ever, with the avalanche of new protein sequences emerging during the post-genomic Era, the gap between the known protein sequences and the known protein structures becomes wider and wider. It is urgent to develop automated methods to predict B-factor profile from the amino acid sequences directly, so as to be able to timely utilize them for basic research. In this article, we propose a novel approach, called PredBF, to predict the real value of B-factor. We firstly extract both global and local features from the protein sequences as well as their evolution information, then the random forests feature selection is applied to rank their importance and the most important features are inputted to a two-stage support vector regression (SVR) for prediction, where the initial predicted outputs from the 1(st) SVR are further inputted to the 2nd layer SVR for final refinement. Our results have revealed that a systematic analysis of the importance of different features makes us have deep insights into the different contributions of features and is very necessary for developing effective B-factor prediction tools. The two-layer SVR prediction model designed in this study further enhanced the robustness of predicting the B-factor profile. As a web server, PredBF is freely available at: http://www.csbio.sjtu.edu.cn/bioinf/PredBF for academic use.

In this pattern recognition study of detecting epilepsy, the first time the authors have attempted to use time domain (TD) features such as waveform length (WL), number of zero-crossings (ZC) and number of slope sign changes (SSC) which are extracted from the discrete wavelet transform (DWT) for the detecting the epilepsy for University of Bonn datasets and real-time clinical data. The performance of these TD features is studied along with mean absolute value (MAV) which has been attempted by other researchers. The performance of the TD features derived from DWT is studied using naive Bayes (NB) and support vector machines (SVM) for five different datasets from University of Bonn with 14 different combinations datasets and 24 patients datasets from Christian Medical College and Hospital (CMCH), India database. Using feature selection and feature ranking based on the estimation of mutual information (MI), the significant features required for the classifier to get higher accuracy is obtained. Further, NB achieves 100% classification accuracy (CA) in distinguishing normal eyes open and epileptic dataset with top 4 ranked features and it gives 100% accuracy with top-ranked two features in using CMCH data.

In the context of high proportion of renewable energy integration, the Jamming fault of steam turbine inlet valves (main steam valve, main regulating valve, etc.) is difficult to be identified early due to the deep peak shaving operation of coal-fired units. This paper proposes a force-displacement collaborative perception early warning method that integrates the valve stem-liner contact dynamics mechanism and 1D-CNN data-driven model. By non-intrusively deploying high-precision force and displacement sensors on the valve stem surface, the synchronous online monitoring of "axial force-valve stem displacement" throughout the entire process of the main valve operation of the steam turbine is achieved. Based on the multi-physical field coupling mechanism of heat, force and chemistry, a valve stem-liner contact dynamics model is established to extract a four-dimensional sensitive feature set covering force-displacement time domain, frequency domain, time-frequency domain and coupling features. A lightweight 1D-CNN network is designed to achieve precise early fault identification. The engineering verification results on a 660 MW supercritical unit show that this method can issue high-accuracy Jamming early warnings, which is 3.2 times more effective than the traditional opening deviation method, providing a new technical path for valve health management in the context of deep peak shaving units

Epileptic seizure detection is commonly implemented by expert clinicians with visual observation of electroencephalography (EEG) signals, which tends to be time consuming and sensitive to bias. The epileptic detection in most previous research suffers from low power and unsuitability for processing large datasets. Therefore, a computerized epileptic seizure detection method is highly required to eradicate the aforementioned problems, expedite epilepsy research and aid medical professionals. In this work, we propose an automatic epilepsy diagnosis framework based on the combination of multi-domain feature extraction and nonlinear analysis of EEG signals. Firstly, EEG signals are pre-processed by using the wavelet threshold method to remove the artifacts. We then extract representative features in the time domain, frequency domain, time-frequency domain and nonlinear analysis features based on the information theory. These features are further extracted in five frequency sub-bands based on the clinical interest, and the dimension of the original feature space is then reduced by using both a principal component analysis and an analysis of variance. Furthermore, the optimal combination of the extracted features is identified and evaluated via different classifiers for the epileptic seizure detection of EEG signals. Finally, the performance of the proposed method is investigated by using a public EEG database at the University Hospital Bonn, Germany. Experimental results demonstrate that the proposed epileptic seizure detection method can achieve a high average accuracy of 99.25%, indicating a powerful method in the detection and classification of epileptic seizures. The proposed seizure detection scheme is thus hoped to eliminate the burden of expert clinicians when they are processing a large number of data by visual observation and to speed-up the epilepsy diagnosis.

Cancer remains the most lethal condition in the world, accounting for 10 million deaths worldwide i.e. 1/6 of all deaths. Currently, cancer detection is primar-ily through symptomatic routes wherein patients present with serious symptoms and undergo imaging and subsequently biopsy of suspected cancer growths for histopathological confirmation of cancer. A diagnostic tool through serum anal-ysis could revolutionise current cancer pathways. Raman spectroscopy offers the ability to measure a complex biochemical fingerprint of a sample through vibra-tional energy shifts. Numerous studies with Raman of biological samples (serum, plasma, tissue, cellular) exist showing promising results for the detection of can-cer. However, these studies are typically limited to proof-of-concept and halt at larger scale studies or forward looking to how these methods might inter-rupt current clinical pathways. Presented in this thesis are a methodology for the optimisation of cancer detection from Raman spectroscopy of human blood serum. The first results chapters show tools in R for the pre-processing of Raman spectra using a variety of techniques. This is followed by an application for qual-ity control of Raman spectra with a view of the necessary safety nets required for tools to integrate into a clinical setting. These tools are then utilised for the task of optimising pre-processing specifically for colorectal cancer detection with human blood serum spectra using high-performance computing (HPC). 2.4 million different pre-processing permutations are trialled in total. This method-ology saw an improvement in diagnostic abilities, with sensitivity increasing by 14.6%, specificity increasing by 6.9%, positive predictive value (PPV) increasing by 3.4%, and negative predictive value increasing by 2.4% when compared to a standard pre-processing optimisation. A similar methodology using HPC is then utilised in chapter 7 to optimise machine learning algorithm selection and feature reduction for colorectal cancer detection of serum spectra. Feature reduction methods principle component analysis (PCA), factor analysis (FA), ElasticNet (EN), and random forest feature selection (RFFS) combined with model types k nearest-neighbours (KNN), logistic regression (LR), support vector machines (SVM), and random forest (RF) are trialled. Traditional feature reduction meth-ods such as PCA and FA were found to perform poorly compared to techniques EN and RFFS. In addition, model types SVM and RF outperform methods LR and KNN. This chapter also shows results from applying artificial neural net-work architectures for colorectal cancer detection and finding that linear machine learning methods (SVM/RF) outperform neural networks. Finally, the last re-sults chapter presents the culmination of these optimised methods applied to building machine learning models for the detection of other cancer types; breast, pancreatic, lung, colorectal cancer. The models achieved 90.9% sensitivity and 77.3% specificity for pancreatic cancer vs. controls, 90.3% sensitivity and 64.5%specificity for lung cancer vs. controls, 92.1% sensitivity and 65.8% specificity for breast cancer and controls, and finally 91.3% sensitivity and 44.0% specificity for colorectal cancer vs. controls where models are thresholded to achieve a minimum of 90% sensitivity. This chapter also focuses on differences in metabolite profile from the Raman spectra between different cancer types and aims to elucidate biochemical linkages.

This paper attempts to explore the feasibility of classifying relaxed and stressful mental states based on two-channel prefrontal EEG signal from 35 healthy human subjects. Specific objective of this paper is to explore the best choice of features and compare the performance of various feature classification algorithms suitable for this purpose. Here, we included different bivariate features in time domain and frequency domain and compared the classification performance of artificial neural network, linear discriminant analysis, quadratic discriminant analysis (QDA), K nearest neighbour and support vector machine algorithms. Common spatial patterns (CSP) algorithm was used successfully for feature reduction. Best classification performance (99.69%) was observed with the QDA classifier taking cross-correlation estimate as feature. We also explored the effect of combining different kinds of features, effect of varying the number of features on classifier performance, robustness of the chosen methods against in inter-individual variability and the feasibility of developing subject-independent classifiers.