ABSTRACT
It is a hot problem in epilepsy research to detect and predict seizures by EEG signals. Clinically, it is generally observed that there are only sudden abnormal signals during the ictal stage, but there is no significant difference in the EEG signal between the interictal and preictal stages. To solve the problem that preictal signals are difficult to recognize clinically, and then effectively improve the recognition efficiency of epileptic seizures, so, in this paper, some nonlinear methods are comprehensively used to extract the hidden information in the EEG signals in different stages, namely, phase space reconstruction (PSR), Poincaré section (PS), synchroextracting transform (SET), and machine learning for EEG signal analysis. First, PSR based on C-C method is used, and the results show that there are different diffuse attractor trajectories of the signals in different stages. Second, the confidence ellipse (CE) is constructed by using the scatter diagram of the corresponding trajectory on PS, and the aspect ratio and area of the ellipse are calculated. The results show that there is an interesting transitional phenomenon in preictal stage. To recognize ictal and preictal signals, time-frequency (TF) spectrums, which are processed by SET, are fed into the convolutional neural network (CNN) classifier. The accuracy of recognizing ictal and preictal signals reaches 99.7% and 93.7%, respectively. To summarize, our results based on nonlinear method provide new research ideas for seizure detection and prediction.NEW & NOTEWORTHY Our results based on nonlinear method have better practical significance and clinical application value and improved the prediction efficiency of epileptic EEG signals effectively. This work provides direct insight into the application of these biomarkers for seizure detection and prediction.
Subject(s)
Electroencephalography , Epilepsy , Nonlinear Dynamics , Signal Processing, Computer-Assisted , Humans , Electroencephalography/methods , Epilepsy/physiopathology , Epilepsy/diagnosis , Machine Learning , Neural Networks, Computer , Seizures/physiopathology , Seizures/diagnosisABSTRACT
BACKGROUND AND OBJECTIVE: Epilepsy is a serious brain disorder affecting more than 50 million people worldwide. If epileptic seizures can be predicted in advance, patients can take measures to avoid unfortunate consequences. Important approaches for epileptic seizure predictions are often signal transformation and classification using electroencephalography (EEG) signals. A time-frequency (TF) transformation, such as the short-term Fourier transform (STFT), has been widely used over many years but curtailed by the Heisenberg uncertainty principle. This research focuses on decomposing epileptic EEG signals with a higher resolution so that an epileptic seizure can be predicted accurately before its episodes. METHODS: This study applies a synchroextracting transformation (SET) and singular value decomposition (SET-SVD) to improve the time-frequency resolution. The SET is a more energy-concentrated TF representation than classical TF analysis methods. RESULTS: The pre-seizure classification method employing a 1-dimensional convolutional neural network (1D-CNN) reached an accuracy of 99.71% (the CHB-MIT database) and 100% (the Bonn University database). The experiments on the CHB-MIT show that the accuracy, sensitivity and specificity from the SET-SVD method, compared with the results of the STFT, are increased by 8.12%, 6.24% and 13.91%, respectively. In addition, a multi-layer perceptron (MLP) was also used as a classifier. Its experimental results also show that the SET-SVD generates a higher accuracy, sensitivity and specificity by 5.0%, 2.41% and 11.42% than the STFT, respectively. CONCLUSIONS: The results of two classification methods (the MLP and 1D-CNN) show that the SET-SVD has the capacity to extract more accurate information than the STFT. The 1D-CNN model is suitable for a fast and accurate patient-specific EEG classification.
Subject(s)
Epilepsy , Humans , Epilepsy/diagnosis , Seizures/diagnosis , Neural Networks, Computer , Sensitivity and Specificity , Electroencephalography/methods , Signal Processing, Computer-Assisted , AlgorithmsABSTRACT
Time-frequency analysis methods can be used to characterize the time-varying characteristics of a signal. The postprocessing algorithm further enhances this ability. The synchroextracting transform is a typical postprocessing algorithm that has the advantage of energy aggregation. However, based on a short-time Fourier transform, shortcomings such as a fixed window length and amplitude distortion when processing frequency modulation signals are unavoidable. This paper proposes a time-frequency postprocessing algorithm with high adaptability, which is called the adaptive synchroextracting transform (ASET). The filter window width for the ASET is adaptive and is determined by the instantaneous frequency change rate for the signal. On this basis, the improved extraction operator can be used to achieve a high-resolution time-frequencyâ representation. This algorithm can be used to better deal with strong frequency modulation signals and has better noise robustness while allowing for signal reconstruction. The effectiveness and practicability of the proposed algorithm are demonstrated by simulation signals and faulty bearing signals.
ABSTRACT
Deep learning has helped achieve breakthroughs in a variety of applications; however, the lack of data from faulty states hinders the development of effective and robust diagnostic strategies using deep learning models. This work introduces a transfer learning framework for the autonomous detection, isolation, and quantification of delamination in laminated composites based on scarce low-frequency structural vibration data. Limited response data from an electromechanically coupled simulation model and from experimental testing of laminated composite coupons were encoded into high-resolution time-frequency images using SynchroExtracting Transforms (SETs). The simulated and experimental data were processed through different layers of pretrained deep learning models based on AlexNet, GoogleNet, SqueezeNet, ResNet-18, and VGG-16 to extract low- and high-level autonomous features. The support vector machine (SVM) machine learning algorithm was employed to assess how the identified autonomous features were able to assist in the detection, isolation, and quantification of delamination in laminated composites. The results obtained using these autonomous features were also compared with those obtained using handcrafted statistical features. The obtained results are encouraging and provide a new direction that will allow us to progress in the autonomous damage assessment of laminated composites despite being limited to using raw scarce structural vibration data.
Subject(s)
Support Vector Machine , Vibration , AlgorithmsABSTRACT
BACKGROUND: Visual evoked potential (VEP) offers a promising research strategy in the effort to characterise brain disorders. Pertinent signal processing techniques enable the development of potential applications of VEP. A joint time-frequency (TF) representation provides more comprehensive information about the underlying complex structures of these signals than individual time or frequency analysis. However, this representation comes at the expense of low TF resolution, increased data volume, poor energy concentration and increased computational time. Owing to the high non-stationarity and low signal-to-noise ratio of VEP, a TF representation that retains only the pertinent components is indispensable. METHOD: The objective of this study is to investigate and demonstrate the ability of various TF approaches to provide an energy-concentrated and sparse TF representation of VEP. The performance of each method has been assessed for its energy concentration and reconstruction ability on both simulated and real VEPs. Renyi entropy, computation time and correlation coefficient are chosen as the performance measures for the assessment. RESULTS: In comparison with the other state-of-the-art approaches, Synchroextracting transform (SET) exhibits the lowest Renyi entropy and the highest correlation coefficient, thereby ensuring a compact TF representation for the better characterisation of VEP signals. These results are also statistically verified through the Friedman test (p<0.001). CONCLUSION: SET assures a powerful TF framework with improved energy concentration at a faster pace while remaining invertible and preserving vital information.
Subject(s)
Electroencephalography , Evoked Potentials, Visual , Entropy , Signal Processing, Computer-Assisted , Signal-To-Noise RatioABSTRACT
The raw vibration signal of a faulty mechanical component carries a large amount of information reflecting its health condition, and the faulty information is typically carried by the high-frequency term in the vibration signal. However, the high-frequency term can easily by overwhelmed by the interference from the low-frequency term and noise. Considering the elimination of interference of the low-frequency term a novel preprocessing technique is presented, one-level kernel regression residual decomposition (KRRD), which can be used to extract the high-frequency term from the raw vibration signal to track the fault information. Combined with the synchroextracting transform (SET) technique, a one-level KRRD-based SET method is proposed. First, the high-frequency term in the raw vibration signal, which contains the faulty information, is extracted using one-level KRRD. Then, the high-frequency term is purified using SET, and the signal-to-noise ratio (SNR) is increased. Finally, a Hilbert envelope analysis is applied to the purified signal to demodulate the faulty feature frequency. To validate the performance and necessity of the proposed method, numerical simulations and experimental investigations are conducted. By introducing two commonly used methods, i.e., empirical mode decomposition (EMD) and variational mode decomposition (VMD), four comparisons (KRRD & EMD, KRRD & VMD, EMD, VMD) are conducted, and the superiority of the proposed method is verified. The analysis results show the effectiveness of the one-level KRRD-based SET method for the detection of mechanical component faults.