The performanceof previous machine learning models for gait phase is only satisfactory under limited conditions. First, they produce accurate estimations only when the ground truth of the gait phase (of the target subject) is known. In contrast, when the ground truth of a target subject is not used to train an algorithm, the estimation error noticeably increases. Expensive equipment is required to precisely measure the ground truth of the gait phase. Thus, previous methods have practical shortcoming when they are optimized for individual users. To address this problem, this study introduces an unsupervised domain adaptation technique for estimation without the true gait phase of the target subject. Specifically, a domain-adversarial neural network was modified to perform regression on continuous gait phases. Second, the accuracy of previous models can be degraded by variations in stride time. To address this problem, this study developed an adaptive window method that actively considers changes in stride time. This model considerably reduces estimation errors for walking and running motions. Finally, this study proposed a new method to select the optimal source subject (among several subjects) by defining the similarity between sequential embedding features.

For gait phase estimation, time-series data of lower-limb motion can be segmented according to time windows. Time-domain features can then be calculated from the signal enclosed in a time window. A set of time-domain features is used for gait phase estimation. In this approach, the components of the feature set and the length of the time window are influential parameters for gait phase estimation. However, optimal parameter values, which determine a feature set and its values, can vary across subjects. Previously, these parameters were determined empirically, which led to a degraded estimation performance. To address this problem, this paper proposes a new feature extraction approach. Specifically, the components of the feature set are selected using a binary genetic algorithm, and the length of the time window is determined through Bayesian optimization. In this approach, the two optimization techniques are integrated to conduct a dual optimization task. The proposed method is validated using data from five walking and five running motions. For walking, the proposed approach reduced the gait phase estimation error from 1.284% to 0.910%, while for running, the error decreased from 1.997% to 1.484%.

A brain–machine interface (BMI) is a technology that controls machines via brainwaves. In BMI, the performance of brainwave analysis is very important for achieving machine control that reflects the user’s intention. One of the main obstacles in this analysis is an insufficient amount of data points because long-term brain signal experiments tend to reduce data quality. Data augmentation methods can be used to overcome this limitation. Recently, several neural network-based data augmentation methods have been developed. However, those methods have several limitations; first, they require considerable computation time because a very large number of parameters must be obtained. Moreover, the neural network based method can suffer from unstable training, which results in quality degradation of artificial data. To address these problems, this paper introduces a method that generates an artificial dataset which has correlation of feature similar to the original dataset. Specifically, after decomposing the covariance matrix for the features into a lower triangular matrix, an artificial dataset can be generated by multiplying the lower triangular matrix by random variables. This method is computationally fast, and the augmentation is stable. When the brainwave data were augmented using this method, classification performance was improved by 1.08%–6.72%. This method focuses on mean, correlation, and not taking into account the other statistical parameters. Since it rapidly generates a large dataset, it can also be useful in other applications.

The acquisition of a large-volume brainwave database is challenging because of the stressful experiments that are required; however, data synthesis techniques can be used to address this limitation. Covariance matrix decomposition (CMD), a widely used data synthesis approach, generates artificial data using the correlation between features and random noise. However, previous CMD methods constrain the stochastic characteristics of artificial datasets because the random noise used follows a standard distribution. Therefore, this study has improved the performance of CMD by releasing such constraints. Specifically, a generalized normal distribution (GND) was used as it can alter the kurtosis and skewness of the random noise, affecting the distribution of the artificial data. For the validation of GND performance, a motor imagery brainwave classification was conducted on the artificial dataset generated by GND. The GND-based data synthesis increased the classification accuracy obtained with the original data by approximately 8%.