Hybrid BF–PSO and fuzzy support vector machine for diagnosis of fatigue status using EMG signal features

Abstract

In this study, a novel BF–PSO–FSVCM model has been proposed to identify the fatigue status of the electromyography (EMG) signal. To improve the classifier accuracy of fuzzy support vector classification machine (FSVCM), a hybrid Bacterial Foraging (BF) and particle swarm optimization (PSO) is proposed to optimize the unknown parameters of the classifier. In the proposed method, the EMG signals are firstly decomposed by discrete wavelet transform (DWT), Fast Fourier Transformation (FFT) and Ensemble Empirical Mode Decomposition (EEMD)–Hilbert transform (HT), and then a set of combined features were extracted from different types of fatigue or normal EMG signals. The optimal fatigue vectors of static, local and dynamic fatigue are also provided in this study. The obtained results obviously indicate that further significant enhancements in terms of classification accuracy can be achieved by the proposed BF–PSO–FSVCM classification system. BF–PSO–FSVCM is developed as an efficient tool so that various support vector classification machines (SVCMs) can be used conveniently as the core of BF–PSO–FSVCM for diagnosis of fatigue status.

Introduction

Muscular fatigue [1] is an exercise-induced reduction in maximal voluntary muscle force. It may arise not only because of peripheral changes at the level of the muscle, but also because the central nervous system fails to drive the motoneurons adequately. The measurement of the muscular fatigue is according to the decline of the system maximal function. The adaptation of the neuromuscular system to heavy resistance exercise is a very complex result of many factors. This implies central and peripheral neural adaptations as well as chemical and morphological modifications of muscle tissue [2].

Surface electromyographic signal (sEMG) is an one-dimensional time series signal of neuromuscular system that recorded for skin surface; the time–frequency domain and nonlinear dynamical characters of it are sensitive to the intensity and state for muscular activities and therefore it is a valuable method for muscle functional evaluation [9]. For many years, the analysis of sEMG concentrated on two main fields, time domain and the frequency domain [3]. The common use of traditional time domain analysis is to regard the sEMG signal as the function of time, using some indices like Integrated EMG (IEMG) or some statistical features such as Root Mean Square (RMS). In aspect of spectrum analysis, the typical method is the Fast Fourier Transform short as FFT. The spectrum we acquire from the FFT can show the distribution of the signals in different frequency components quantitatively. In order to portray the characteristic of the spectrum, researchers often use the following two indicators: Median Frequency (MF) and Mean Power Frequency (MPF) [4]. The mean frequency of sEMG signal is an important index of local muscle fatigue. Most research works have found that the indicators of frequency analysis of the muscles in limbs and waist have good regularity in the condition of the static burthen. The main manifestation is the decline of the MPF or MF and the increase of power spectrum or the ratio of the low/high frequency during the process of fatiguing. However, under the dynamic condition, the alteration of the MPF and MF during the fatiguing process exists considerable difference from which can hardly get a universal conclusion [17].

Recently, Georgakis et al. [5] proposed to analyze the instantaneous frequency (IF) of fatigue EMG directly using Hilbert transform. They reported that the reliability and accuracy of the IF was better than the conventional spectral variables, i.e. mean frequency and median frequency. However, contrary to the suggestion given by [5], several authors [6], [7] argued that one should not just take any data to perform a Hilbert transform, find the phase function, and define the instantaneous frequency as the derivative of this phase function. They pointed out, if one follows this path, one would obtain a finite number of points where the frequency becomes very high and even assumes negative values that bear no relationship to the real oscillation of the data. It is to say that the instantaneous frequency derived from this method is typically oscillatory and often extends beyond the spectral range of the signal. To overcome the shortage of instantaneous frequency, a novel fatigue feature is proposed by the weighted frequency energy signals in this study.

Obviously, the feature extraction methods are also very important in the recognition of the fatigue status of EMG signals. From the above review of literatures [39], [40], [41], [42], [43], [44], features can be time and frequency domain information, such as Fourier coefficients, Hilbert transform coefficients [8], wavelet coefficients, and a wide range of quantities derived from other signal processing techniques. In the published literatures, the changes of features such as IEMG, RMS, MPF and MF were usually used to find the static or local fatigue phenomena, while the dynamic fatigue, feature IF has no enough capability to clue on the occurrence of fatigue. Features extracted from FuzzyEn, approximate entropy (ApEn), sample entropy (SampEn), and cumulative residual entropy (CrEn) have been applied to the classification of EMG signal [45], [50], [51], while wavelet entropy and wavelet energy techniques have been rarely applied as feature of EMG signal reported in [50]. By the analysis of the above, some statistical features of fatigue status such as IEMG, RMS, MPF, MF, wavelet entropy, wavelet high-low frequency energy ratio and improved IF are used to characterize the EMG interference pattern in the diagnosis of different types of muscular fatigue status in this study. These features can then be used as input data for classifiers such as NNs and SVMs, which can classify or detect muscular fatigue status.

However, design of practical and accurate automated recognition system remains challenging. The challenge in this area is to develop a robust and efficient classification technique that preserves important discriminatory information so as to provide a better accuracy for classification. Consequently, for an effective automated EMG signal classification, a systematic treatment of EMG signals must be carried out. For this reason, a number of computer-based quantitative EMG analysis algorithms have been developed [38].

More recently, Hassoun et al. [10] proposed the NNERVE algorithm in 1994 in an attempt to computerize the extraction of individual EMG. Schizas et al. [11] used artificial neural network (ANN) to classify action potentials of a large group of muscles. Their work was later extended by Schizas et al. [12] to compare K-means, multi-layer perceptron (MLP)–neural network (NN), self-organizing maps (SOMs), and genetic-based classifiers. It was shown that the simple K-means algorithm was not suitable as it gave the lowest classification accuracy, but both ANN and genetic-based models produced promising results. Pattichis et al. [13] extended the work by Schizas et al. [12] by applying ANN to motor unit action potential (MUAP) signals collected from the biceps brachii muscle. They compared the MLP network against the K-means clustering and Kohonen’s SOM. Comparisons with unsupervised learning methods were undertaken by Pattichis and Elia [14] using SOM, learning vector quantization (LVQ), and statistical methods. Pattichis and Pattichis [15] have investigated the usefulness of the wavelet transform (WT), that provides a linear time-scale representation for describing MUAP morphology and three different neural networks, the back propagation (BP), the radial basis function network (RBF), and the self-organizing feature map(SOFM). Subasi et al. [16] investigated the practicality of using an autoregressive model with wavelet neural network to classify EMG signals. Subasi and Kiymik [17] developed the EMG decomposition system consisting of time–frequency methods and independent component analysis (ICA).

SVM is a promising data classification technique proposed by Vapnik [18]. SVM is generated from the training process with the training data. Later on, classification is implemented based on the trained model [19], [20], [21], [22], [23]. Due to the fact that it is hard to obtain sufficient fatigue samples in practice, SVM is introduced into fatigue recognition of EMG signal due to its high accuracy and good generalization for a smaller number of samples. Moreover, some potential influencing factors are fuzzy and uncertain in EMG data collection [23], [42]. The standard SVM is hard to handle the fuzzy sample set. Therefore, a new fuzzy SVM should be explored. Yan et al. [46] constructed a fuzzy support vector machine model, in which a fuzzy membership function was utilized to transfer the output of a SVM discriminant function into a fuzzy class score. They made a comparison between the FSVM scheme and a BP neural network. The results indicated that four hand/wrist motions could be identified by FSVM with about 5% higher accuracy than BP network [50].

However, the fuzzy support vector machines mentioned in the above literatures [47], [48], [49], [50] are not suitable for the input and output variables described as triangular fuzzy numbers from the triangular fuzzy number space of which the input variables in classification problem of SVM may come from. The triangular fuzzy SVCM has the ability to establish optimal problem of uncertain information. The results will be obtained precisely. Moreover, the main problems encountered in setting up the FSVCM model are how to decide on the kernel function and its parameter values. Improper parameter settings and feature-selection may lead to deprived classification results [21]. In order to improve the FSVCM classification accuracy, the parameters that should be optimized including the penalty parameter s and the kernel function parameter for the radial basis function (RBF) kernel must be tuned. So, the tuned model parameters have an effect on the classification accuracy. This study attempts to increase the EMG signal classification accuracy rate by utilizing an approach based on a novel hybrid evolutionary computation algorithm based on bacterial foraging (BF) and particle swarm optimization (PSO) in FSVCM. BF algorithm is proposed by Passino [24] and a population-based search algorithm that is initialized with a population of random solutions, called bacteria. Recently, several researches on BF theories and applications have been reported [25], [26], [27], [28], [29], [30], [31], while BF algorithm has been not applied in the classification of EMG signal [50]. In BF algorithm, a unit length direction of tumble behavior is randomly generated which may lead to delay in reaching the global solution. To solve the shortcoming, BF and PSO were combined to use advantages of both techniques. The aim is to make use of PSO ability to exchange social information and BF ability in finding a new solution by elimination and dispersal. In the combined BF–PSO, the unit length random direction of tumble behavior can be obtained by the global best position and the best position of each bacterium by PSO algorithm.

In this study, the BF–PSO is proposed to use as an optimization technique to optimize the FSVCM parameters in fatigue classification of EMG signal. To recognize the fatigue status of EMG signal, a methodology has been developed for EMG signal classification which consists of three steps, which is shown in Fig. 1. In the first step, the EMG signals are decomposed using different signal process techniques such as wavelet, FFT and EEMD–Hilbert transform. In the second step, statistical features are extracted from the decomposed EMG signals using some technique such as entropy and energy ratio, and then provide better accuracy for diagnosis of muscular fatigue. In the last step, an unknown part of EMG signal is classified as normal and fatigue by different classification methods.

Obviously, in recognition applications of EMG or electroencephaloGram (EEG) signals, the observed input data cannot be measured precisely and affected by noise. Traditional support vector machine method cannot cope with fuzzy and uncertain information. It is well known that fuzzy logic is a powerful tool to deal with fuzzy and uncertain data. The conventional fuzzy support vector machine (FSVM) introduced a fuzzy membership to each training sample of SVM, so that fuzzy training samples could be trained and tested by standard SVM. In our proposed method, fuzzy information of sample data is represented accurately by triangular fuzzy number space, and then tree optimization problems on the left, middle and right of triangular fuzzy number are modeled and represented in the unified SVM model. The established SVM model develops SVM literatures in the field of fuzzy training samples.

And then, the main novelty of this paper is the proposed FSVCM-based approach and its parameter optimization algorithm, which aim at the performances shortage of SVCM classifier by fuzzy modeling of SVCM. The flowing results verify that the proposed BF–PSO–FSVCM classification method improves significantly the generalization capability of the classifier. Another advantage of the proposed BF–PSO–FSVCM approach can be found in its high sparseness, which is explained by the fact that the adopted optimization algorithm is based on hybrid BF and PSO. Another contribution of this study is how to select the optimal feature vectors for different fatigues such as static fatigue, local fatigue and dynamic fatigue. Our conclusion strengthens the published literatures of fatigue recognition.

The remaining paper is organized as follows: in the next section, we give information about the fatigue feature extracted using different signal process methods. The proposed classifier system is arranged in Section 3. Section 4 provides a complete experimental study for the BF–PSO–FSVCM model-based fatigue classification scheme of EMG signal, in which the impact of feature set and algorithmic issues are examined with respect to classification performance. The advantage of our proposed method is arranged in Section 5. Finally, the conclusions are summarized in Section 6.