**1. Introduction**

The exoskeleton robot could be a revolutionary technology in human limb rehabilitation [1] and power enhancement [2]. However, the motion intention of its wearer has limited the development of this technology, because traditional sensors for the exoskeleton are unable to detect motion tendency ahead of time. Since the surface electromyography (sEMG) signal is noninvasive, and has the potential to predict people's movement intentions 30–100 ms in advance [1], the sEMG has been favored by many researchers, and with the help of sEMG sensors, the performance of wearable devices would be improved [1,3–5]. Thus, in addition to the exoskeleton technology, there are a wide range of applications for using sEMG, such as wearable devices [6], prosthetic limbs [7], and other such myoelectric control systems [8]. The existing studies focus on how the sEMG signals relate to human movement for better control of the wearable devices and similar products.

Some of the existing research has investigated the relationship between sEMG and human biomechanics. Chen et al. [9] proposed a musculoskeletal biomechanical model connecting sEMG and knee joint torque, based on the underlying physiological mechanism facilitating the study of neural control. Tagliapietra et al. [10] used a subject-specific EMG-driven Neuro MusculoSkeletal (NMS) model to estimate ankle torque and muscle forces expressed by the subject. Zhuang et al. [11] proposed

an sEMG-based admittance controller that could enable a more synchronized human–robot interaction, as compared to the torque-sensing-based admittance controller.

However, to avoid building a complicated biomechanical model, some researchers have tried to use a data training method to do the job. Anwar et al. [12] proposed an adaptive neuro fuzzy inference system (ANFIS), such as a neuro-fuzzy type knowledge-based adaptive network that contained a non-parametric model, with an EMG signal of two muscles used as the input to estimate torque. Gui et al. [13] used radial basis function (RBF) neural networks to approximate the active joint torque of subjects during the swing phase.

Besides the biomechanical method, there is a new intuitive myoelectric control strategy for assistive devices, which relies on the sEMG-based intention estimation of human motion. These predictions can be broadly categorized as classification and regression models [14]. For the classification, Toledo-Pérez et al. [15] used a support vector machine (SVM) based on sEMG to classify the intention of right lower limb movement. Morbidoni et al. [16] proposed a deep learning (DL) approach for sEMG-based classification of stance/swing phases and the prediction of the foot–floor-contact signal in more natural walking conditions. Nazmi et al. [17] proposed a classification system for both stance and swing phases, by extracting the patterns of electromyography signals from time domain features and feeding them into an artificial neural network (ANN) classifier.

For the continuous estimation of the joint angle, there are various methods. Bao et al. [18] presented a single stream convolutional neural network (CNN) for mapping sEMG to wrist angles within three degrees-of-freedoms. Xiao et al. [19] used the mean absolute value, waveform length, zero crossing, slope signs changes, and the difference in absolute standard deviation value of sEMG, in order to estimate continuous elbow motion by random forest (RF). Lei Z. [20] used the back propagation (BP) neural network to establish a model of the relationship between elbow angles and sEMG signals features, through which they estimated the angles of the elbow joint and achieved continuous motion control of the exoskeleton. Huang et al. [21] presented deep-recurrent neural networks (RNNs) for predicting the knee joint angle in real-time, based on a fusion of sEMG and kinematics signals.

It can be concluded that most of the existing studies are model-free approaches for the estimation of joint angles from sEMG, based on machine learning (ML). Furthermore, the majority of the existing research used a single method of ML, and most of them seldom considered the influence of the previous sEMG as the input on the accuracy of their methods, even though both the joint movement and sEMG signals are in a continuous time-sequence. Also, the size of the training sample for ML is also a debatable issue in terms of the accuracy of estimation, since a large training sample would lead to overfitting and be time consuming.

Thus, a novel double-ML-method, based on random forest combined with principal component analysis (RFPCA) has been proposed in this work to estimate the movement of the knee joint from sEMG, with the expectation of achieving high accuracy and efficiency. This method was also utilized to analyze how the input of the previous sEMG and the sample size for model validation affect the estimation of the knee joint movement. Moreover, a BPPCA was constructed to compare with the RFPCA, and the RFPCA presented a better performance in this work.

The remainder of the paper is organized as follows. Section 2 introduces the experiment and proposes the knee angle estimation method. Section 3 presents the estimation results using RFPCA and BPPCA, followed by the discussion in Section 4. Finally, the conclusions are drawn in Section 5.

### **2. Materials and Methods**
