Feature extraction of wavelet transform coefficients for sEMG classification

As a powerful tool, the wavelet transform method has been widely used in feature extraction for hyperspectral data, while few studies are focused on soil moisture retrieving. The objective of this study was to evaluate the effectiveness of discrete wavelet transform (DWT) for soil moisture retrieving. In this study, a total of 78 measurements of soil moisture and hyperspectral data were collected through soil sampling as well as laboratory quantitative control. There were 13 different mother wavelets capable of decomposing hyperspectral data that were recommended, along with six decomposition levels from 5 to 10. The performances of two feature extraction methods namely band selection and DWT were compared, using three indexes, i.e., R2, Radj and root mean square error (RMSE) introduced to validate soil moisture retrieving results. The experimental results indicated that the wavelet transform method could significantly reduce the dimensionality of hyperspectral data, resulting in a much more effective performance. Among the 78 estimation models using the wavelet transform, there were 42 models superior to band selection, with 24 models yielding good correlations between the predicted soil moisture and the measured ones (R2 ≥ 0.7, RMSE ≤ 0.050, p ≤ 0.05). Furthermore, when the wavelet decomposition level was 9 and the mother wavelets were Daubechies 2, Daubechies 4, Reserve Biorthogonal 3.3 and Biorthogonal 6.8, the retrieving results were optimum. Additionally, the experimental results proved that the wavelet analysis technique was capable of preserving high- and low-frequency spectral information at different decomposition scales, and could correctly reflect the variation of soil moisture. Thus, it would be helpful in further environmental monitoring.

BackgroundMost of the objective and quantitative methods proposed for spasticity measurement are not suitable for clinical application, and methods for surface electromyography (sEMG) signal processing are mainly limited to the time-domain. This study aims to quantify muscle activity in the time–frequency domain, and develop a practical clinical method for the objective and reliable evaluation of the spasticity based on the Hilbert–Huang transform marginal spectrum entropy (HMSEN) and the root mean square (RMS) of sEMG signals.MethodsTwenty-six stroke patients with elbow flexor spasticity participated in the study. The subjects were tested at sitting position with the upper limb stretched towards the ground. The HMSEN of the sEMG signals obtained from the biceps brachii was employed to facilitate the stretch reflex onset (SRO) detection. Then, the difference between the RMS of a fixed-length sEMG signal obtained after the SRO and the RMS of a baseline sEMG signal, denoted as the RMS difference (RMSD), was employed to evaluate the spasticity level. The relations between Modified Ashworth Scale (MAS) scores and RMSD were investigated by Ordinal Logistic Regression (OLR). Goodness-of-fit of the OLR was obtained with Hosmer–Lemeshow test.ResultsThe HMSEN based method can precisely detect the SRO, and the RMSD scores and the MAS scores were fairly well related (test: χ2 = 8.8060, p = 0.2669; retest: χ2 = 1.9094, p = 0.9647). The prediction accuracies were 85% (test) and 77% (retest) when using RMSD for predicting MAS scores. In addition, the test–retest reliability was high, with an interclass correlation coefficient of 0.914 and a standard error of measurement of 1.137. Bland–Altman plots also indicated a small bias.ConclusionsThe proposed method is manually operated and easy to use, and the HMSEN based method is robust in detecting SRO in clinical settings. Hence, the method is applicable to clinical practice. The RMSD can assess spasticity in a quantitative way and provide greater resolution of spasticity levels compared to the MAS in clinical settings. These results demonstrate that the proposed method could be clinically more useful for the accurate and reliable assessment of spasticity and may be an alternative clinical measure to the MAS.

Comparative analysis of SVM and Naive Bayes classifier for the SEMG signal classification

<p>Accurate water level (WL) forecasting is important for water resources management and planning purposes in the Great Lakes. The objectives of this research are two-fold.  The first objective is to apply machine learning (ML) (i.e., random forest (RF) and support vector regression (SVR)) and hybrid convolutional neural network(CNN)-long-short term memory (LSTM) deep learning (DL) models for multi-step (i.e., one-, two- and three-monthly step ahead) WL forecasting in the Great Lakes (Michigan and Ontario). The second objective is to integrate the boundary corrected (BC) maximal overlap discrete wavelet transform (MODWT) with SVR, RF, and CNN-LSTM models to improve the performance of the individual models. By employing a BC-wavelet decomposition method, the ‘future data’ issue (i.e., data from the future that is not available), often overlooked in the literature and a major barrier to achieving realistic forecasting performance is overcome. </p><p>For Lakes Michigan and Ontario, 1212 monthly WL (m) records (spanning Jan 1918–Dec 2018) were used to develop the models. For the non-wavelet-based models (SVR, RF, and CNN-LSTM), candidate model inputs included the WL recorded over the previous 12 months.  For the BC-MODWT-based models (BC-MODWT-SVR, BC-MODWT-RF, and BC-MODWT-CNN-LSTM), the lagged input time series were decomposed into BC-wavelet and scaling coefficients by using different mother wavelets (Haar, Daubechies, Symlets, Fejer-Korovkin and Coiflets), filter lengths (from two up to 12) and decomposition levels (from one up to seven).  For each method (SVR, RF, and CNN-LSTM), mother wavelet, and decomposition level a model was generated.  For both wavelet- and non-wavelet-based models, the particle swarm optimization (PSO) method was used to select the most appropriate inputs to include in the proposed multi-step WL forecasting models.</p><p>The datasets were partitioned into calibration and validation subsets. After calibrating the models, various performance evaluation metrics, e.g., coefficient of determination (R<sup>2</sup>), root mean square error (RMSE), mean absolute error (MAE), root mean square percentage error (RMSPE), mean absolute percentage error (MAPE) and the Nash-Sutcliffe efficiency coefficient (NSC) were used to assess model accuracy.</p><p>Of the ML models, the SVR outperformed RF while the DL models outperformed the ML models for each forecast lead time (one-, two-, and three-step(s) ahead). Results from this case study indicate that not all wavelet families and decomposition levels perform equally and in some cases, the wavelet-based models do not improve performance over the non-wavelet-based models. However, the BC-MODWT-CNN-LSTM using suitable mother wavelets (e.g., Haar) outperforms the individual ML and BC-MODWT-ML-based models. More accurate forecasts were obtained for Lake Michigan although the performance in both Great Lakes was accurate. The outcomes of this research indicate that the BC-MODWT-CNN-LSTM model is a promising tool for generating accurate WL forecasts.</p>

In this article, we make a comparative study for a new approach compression between discrete cosine transform (DCT) and discrete wavelet transform (DWT). We seek the transform proper to vector quantization to compress the EMG signals. To do this, we initially associated vector quantization and DCT, then vector quantization and DWT. The coding phase is made by the SPIHT coding (set partitioning in hierarchical trees coding) associated with the arithmetic coding. The method is demonstrated and evaluated on actual EMG data. Objective performance evaluations metrics are presented: compression factor, percentage root mean square difference and signal to noise ratio. The results show that method based on the DWT is more efficient than the method based on the DCT.

Surface electromyogram (sEMG) signals are widely used to control the myoelectric prosthetic arm for amputees. In this study, the authors investigated the usefulness of discrete wavelet transform (DWT) features from multiple levels of approximation and detail coefficients obtained from sEMG signals. DWT is used for de-noising as well as feature extraction in this study and further tested using Support Vector Machine (SVM) classifier and Artificial Neural Network (ANN) Classifier. The performance of SVM classifier is compared with ANN classifier. The classification accuracy of SVM classifiers was found better as compared to ANN in term of speed and robustness. In the first section of paper the authors presented the introduction related to the study. The experimental set-up for recording surface EMG signal is presented in section 2. In section 3 the experimental results are depicted and in section 4 the results are concluded.

The aim of this paper is to find the optimal mother wavelet function and wavelet decomposition level when denoising the Doppler cardiogram (DCG), the heart signal obtained by the Doppler radar sensor system. To select the best suited mother wavelet function and wavelet decomposition level, this paper presents the quantitative analysis results. Both the optimal mother wavelet and decomposition level are selected by evaluating signal-to-noise-ratio (SNR) efficiency of the denoised signals obtained by using the wavelet thresholding method. A total of 115 potential functions from six wavelet families were examined for the selection of the optimal mother wavelet function and 10 levels (1 to 10) were evaluated for the choice of the best decomposition level. According to the experimental results, the most efficient selections of the mother wavelet function are “db9” and “sym9” from Daubechies and Symlets families, and the most suitable decomposition level for the used signal is seven. As the evaluation criterion in this study rates the efficiency of the denoising process, it was found that a mother wavelet function longer than 22 is excessive. The experiment also revealed that the decomposition level can be predictable based on the frequency features of the DCG signal. The proposed selection of the mother wavelet function and the decomposition level could reduce noise effectively so as to improve the quality of the DCG signal in information field.

The measurement of the rectal pressure is considered to be the 'gold standard' for the assessment of the abdominal pressure. However, conventional rectal catheters can cause erroneous results and are not comfortable for the patients. To reduce these problems, we devised a non-invasive technique for the measurement of the abdominal pressure using the parametric curve fitting method, based on linear, polynomial, exponential or sine equation modeling, between the rectal pressure and electromyographic (sEMG) signals recorded simultaneously from the abdomen. The sEMG signals and rectal pressure were obtained simultaneously from 12 patients with neurogenic bladders due to spinal cord injury (age = 53.2 ± 11.9 years, BMI = 24.4 ± 2.7, ASIA classification: D). Using our algorithm, the correlation coefficient and root mean square error (RMSE) between the measured and estimated abdominal pressure obtained by the quartic polynomial modeling were 0.86 ± 0.05 and 4.70 ± 1.56, respectively. The results obtained herein suggest that the sEMG signals can be used reliably for the indirect measurement of the abdominal pressure in ambulatory urodynamics monitoring systems.

The new generation of prosthetic devices are based on simultaneous and proportional estimation of kinematics from recorded surface electromyographic (sEMG) signals of the desire limb. In this paper we applied Generalized Regression Neural Network (GRNN), a non-linear system identification approach, to estimate fingers kinematics (15 Degrees of Freedom) from sEMG signals. The parameters were optimized based on training data of 40 subjects during 9 hand’s principal movements. In order to reduce the input parameters of the model in a feature selection, suitable features such as auto regressive coefficients, zero crossing, slope sign change, waveform length, root mean square, and discrete wavelet transform were computed from sEMG signal. The performance of the estimation was assessed based on Pearson correlation coefficient or R-value index. The average overall Rvalue for 15 DoFs in all the subjects was 87.84±5.02%, comparable with the state of the art approaches in the literature. As the proposed method and set-up use dataglove to record kinematic information, thus has more realistic data acquisition protocol which has potential to be used in clinical setting to provide fast, accurate, and intuitive simultaneous and proportional control strategy for myoelectric hand prostheses.

Compressive sensing (CS) is a new technique that give an approach to reconstruct the signal with few numbers of observations or measurements. The CS is based on L1 -norm minimizations to find the sparse solutions and it is known as basis pursuit. In this investigation, CS scheme with different transforms is proposed. The three utilized transform techniques are: Discrete Fourier Transform(DFT), Discrete Cosine Transform(DCT) and Discrete Wavelet Transform(DWT) with daubechies1(DB1) and coiflets1(coif1) basis. The proposed system is tested by employing the following signals: Blocks, Heavy Sine, ‘Bumps’ and ‘Doppler’ which cover wide range of applications. The four testing signals are represented in sparse domain using different transforms. In order to threshold the coefficients of the signals in sparse domain, the universal threshold is utilized in the case of CS with FFT and DWT whereas, the universal threshold is modified to prune the DCT coefficients. The main aim of this study is to investigate the differences among CS with DFT, CS with DCT and CS with DWT, and consequently a suitable transform domain used with CS to be selected. The comparative study is established by assessing the performance of the proposed system using Root Mean Square Error (RMSE), output SNR, and the time required to reconstruct approximated signals. Simulation results have shown that the CS with DWT outperforms the CS with FFT and DCT. CS with DWT has achieved good RMSE values about (0.0014 to 3.359e-8) even when half of the signal elements are removed. CS with FFT and DCT enhanced the noisy Blocks and Bumps signals by 3dB and 1dB respectively, while it is failed to enhance noisy Heavy Sine and Doppler signals. CS with DWT of two basis and for single decomposition level have improved the noisy Blocks, Bumps, Heavy Sine and Doppler signals by 5dB, 4dB, 3dB, and 3dB respectively.

The efficacy of using surface electromyography (sEMG) signals to understand the biomechanics of movement has been an active area of research. The need for a reliable standard to monitor/predict rehabilitative outcomes and an efficient control to drive orthotics and exoskeletons has motivated this work to use sEMG as an estimate for joint angle prediction. This work involves acquisition of 8 channel sEMG signals from lower extremity during knee articulation, involving flexion and extension for 20 able bodied subjects. Simultaneous knee joint angle measurement from goniometer is used to train a Back Propagation Neural Network (BPNN) to estimate knee Range of motion (ROM) with Root Mean Square values of 8 sEMG signals as input. Knee joint angles at full ROM for flexion $(90.9\pm 2.4)$ and extension $(19.1 \pm 2.6)$ were obtained. Estimation accuracy based on average Mean square error for flexion (0.146 $\pm 0.197$ ) and extension $(0.098\pm 0.129)$ ) with correlation factor (r) between actual and estimated values were found to be 0.93 and 0.95 respectively. Estimated flexion and extension angles were used to actuate the knee joint of a Virtual Human model. The results suggest that a high degree of correlation can be achieved between sEMG and kinematic measurements.

Continuous prediction of knee and ankle joint angles based on sEMG and hip joint angle fusion signals by NRBO-LSTM network

Geometric correction is used to correct the registration errors in remotely sensed images. These images are often compared to ground control points (GCPs) either by using an accurate map (image to map) or using another geo-referenced image (image to image) and then resampled. Accordingly, the exact locations and the appropriate pixel values can be calculated in more accurate, time-wise and effortless manner. In the traditional methods, the GCPs are manually selected and then the transformation models are applied which yield time consuming and less accurate processes. The objective of this work is to develop an automatic approach for image registration based on another geo-referenced image using five feature extraction models. They are Scale Invariant Feature Transform (SIFT), Speeded Up Robust Features (SURF), Discrete Wavelet Transforms (DWT), (SIFT & DWT), and (SURF & DWT). The GCPs were selected based on the least-squares adjustments as the basis for improving the spatial accuracy of all the linking points in both images. The obtained results showed that models have higher accuracy in image registration with Root Mean Square Error (RMSE) less than 0.5. The developed automated image registration method provides more accurate results and saves time, money and effort.

Amyotrophic Lateral Sclerosis (ALS) is one of the most common neuromuscular diseases which affects both lower and upper motor neurons. In this paper, a dilated one dimensional convolutional neural network, named ALSNet, is proposed for identifying ALS from raw EMG signal. No hand-crafted feature extraction is required, rather, ALSNet is able to take raw EMG signal as input and detect EMG signals of ALS subjects. This makes the method more feasible for practical implementation by reducing the computational cost required for extracting features. To our best knowledge, no research work for identification of ALS from raw EMG signal has been conducted yet. The performance of the ALSNet was evaluated using popular metrics such as overall accuracy, sensitivity, specificity and balanced accuracy and compared with other existing methods. The proposed method showed a better performance than the other existing methods with an overall accuracy of 97.74%.

A neural network-based model for lower limb continuous estimation against the disturbance of uncertainty