Gait Phase Recognition of Hip Exoskeleton System Based on CNN and HHO-SVM Model

Abstract

Gait phase recognition is crucial for developing wearable lower-limb exoskeleton robots and is a prerequisite for the compliance control of lower-limb exoskeleton robots. Accurately estimating the gait phase is still a key challenge in exoskeleton control. To address these challenges, this study proposes a hybrid model that combines Convolutional Neural Networks (CNN) and Harris Hawks Optimization (HHO)—based Support Vector Machine (SVM). First, the collected sensor signals are preprocessed by normalization to reduce the differences in the data of the subjects. Then, a simplified CNN is used to automatically extract more discriminative features from the dataset. These features are classified using SVM instead of the softmax layer in CNN. In addition, an improved Harris hawk optimization (HHO) algorithm is used to optimize the SVM classification process. This model can accurately identify the heel strike (HS), flat foot (FF), heel off (HO), and swing (SW) phases of the gait cycle. The experimental results show that the CNN-HHO-SVM algorithm can achieve an average phase recognition accuracy of 96.03% for seven subjects in the self-built dataset, which is superior to the traditional method that relies on manually extracting time-frequency features. In addition, the F1-score and macro-recall of the CNN-HHO-SVM algorithm are better than those of other algorithms, which verifies the superiority of the algorithm.

1. Introduction

Given how quickly lower limb exoskeleton robots are developing, their potential for application in military, industrial, medical, and other fields has been enormous [1]. Depending on the intended purpose, the primary focus of research on lower limb exoskeleton robots may be categorized into two distinct groups: gait rehabilitation for patients with motor disorders and strength enhancement for healthy people. Rehabilitation robots aid in the recovery of patients who have had surgery and assist patients with impaired lower limb motor function in regaining their capacity to walk, e.g., the LOKOMAT [2] and ALEX [3] used on a treadmill. There are also lower limb exoskeleton robots for augmentation of the human body, which can increase the speed, strength, and endurance of the user for specific scenarios of use, e.g., ReWalk [4] HAL [5] BLEEX [6].

As technology for human–computer interaction for lower limb exoskeleton robots advances, these devices have gradually changed from passive control to an active understanding of human motion intent. Hence, a correct understanding of human motion intent is a prerequisite for achieving coordinated human-computer control. Human motion intention mainly includes motion patterns and the gait phase. Gait phase detection is essential for the lower limb exoskeleton robot’s gait evaluation, control, and trajectory generation. Developing reliable gait detection algorithms can improve the control accuracy, stability, and safety of lower limb exoskeleton robots and other assistive devices. However, inaccurate information about the gait phase may result in the motor’s inappropriate application of joint torque, which may hinder the user’s movement [7]. For example, Miguel Sánchez-Manchola et al. [8,9] proposed a Hidden Markov Model (HHM) algorithm for detecting gait phases, which uses angular velocity and acceleration signals on the instep to identify four gait events and varies the system impedance of the impedance controller in response to the detected gait phases, enabling on-demand assisted control strategies for lower limb exoskeleton robots.

To accurately identify the phase of human motion, relying on the combined effect of sensors and classifiers is necessary. A lot of work has been conducted in the past many years. Currently, the commonly used gait phase recognition methods are mainly based on optical tracking recognition [10,11,12] and wearable sensor-based recognition [13]. While optical tracking-based recognition has higher accuracy and better repeatability, it also requires expensive equipment and experiments and is limited to the laboratory [14]. Gait recognition based on wearable sensors mainly consists of inertial measurement units IMUs [15], pressure sensors [16], surface electromyography EMG [17], and electroencephalography EEG, etc. [18]. This technique is not only low-cost, compact in size, simple, and easy to install, but it is also not readily impacted by the outside environment. Therefore, gait phase segmentation using wearable devices has received more and more attention from scholars [19].

In classification using classifiers such as Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Random Forest, etc., it is usually necessary to extract some temporal domain attributes, spectral domain attributes, or combined temporal and spectral features from the original signals, such as mean, standard deviation, variance, wavelet decomposition coefficients, spectral density, etc., which require a lot of time for experienced researchers to test and validate to select the features with high classification performance [20,21]. Pingping Lv et al. [22] introduced a particle swarm optimization approach that utilizes the radius edge of the new support vector machine PSO-FSVM for phase identification. The standard deviation and mean of the pressure in the four regions of the foot and the sum of the pressure on the foot are selected as the model’s characteristic inputs. They experimentally verified that the method at four different speeds achieved 96% accuracy. Johnny D. Farah et al. [23] verified the performance of Decision Tree, Random Forest, Multilayer Perceptron, and SVM classifiers in phase recognition using velocity and gyroscope signals at the thighs as inputs. They compared the effect of whether or not to use the knee angle as a parameter on the recognition effect, and the experimental results showed that the J-Decision Tree with the knee angle has a better overall performance. Jakob Ziegler et al. [24] proposed a novel EMG feature calculated from the bilateral EMG signals of muscle pairs. They used the SVM algorithm to identify the standing and swinging phases with 96% recognition accuracy. Although these approaches have demonstrated efficacy, they frequently entail laborious feature extraction and selection procedures. Consequently, there is an increasing interest in automated methodologies that might optimize this process.

Convolutional neural network (CNN), as a kind of feed-forward deep learning algorithm, has powerful feature extraction and non-linear arithmetic ability, which can directly extract its obvious features from the original data and thus effectively avoid the problem that traditional deep learning methods need to take complex mathematical methods for data feature extraction. It has thus been widely researched in the fields of computer vision, natural language processing, and speech recognition [25,26]. As a result of the successful performance of CNN architectures in several domains, researchers have started to widely employ them in time series classification issues [27,28]. The CNN, or convolutional neural network, incorporates an automated feature extractor that includes a convolutional layer and a subsampling layer. This addition to the conventional neural network allows for the implicit and automatic extraction of highly discriminative features from the training data, eliminating the need for explicit feature extraction [29]. While CNNs offer several advantages over more conventional neural networks, their extensive use of convolutional, pooling, and fully connected layers—designed to capture complex features—makes them prone to overfitting, particularly when trained on small or imbalanced datasets. This overfitting can significantly reduce the model’s ability to generalize new, unseen data, a critical issue in real-world applications such as human action recognition and time series classification. As a result, CNNs often struggle to adapt to individual differences and environmental variability. Therefore, to improve the generalization ability of CNN, people have studied alternatives to the softmax function in handwriting recognition [30], human action recognition [31], and time series classification problems [32]. SVM [33], grounded in statistical learning theory, seeks to minimize structural risk.

Another approach to overcome CNN limitations, especially for sequence-based tasks, is combining CNNs with long short-term memory (LSTM) networks. LSTM captures temporal dependencies, making the hybrid CNN-LSTM model effective for gait recognition. Studies like Shi et al. [34] used 1D-CNN for feature extraction from sensor data followed by bidirectional LSTM for time–series modeling, while Kreuzer and Munz [35] integrated 2D-CNN and LSTM to predict gait phases with over 92% accuracy. Despite these advances, this paper focuses on the CNN-SVM framework. CNN excels at feature extraction, and SVM reduces overfitting with limited data, offering a simpler yet accurate model for gait recognition.

It is worth noting that in the practical application of lower limb exoskeleton robot’s motion phase recognition, collecting a large number of human gait features consumes a lot of manpower and time, which is impractical. Therefore, developing a lower limb motion phase recognition model for a small sample dataset is of great significance.

To summarize the above considerations, the use of deep learning models and traditional feature extraction methods can ensure high recognition rates. However, this approach comes at the cost of high computational demands and a strong reliance on large amounts of labeled data. Additionally, traditional feature extraction methods are time-consuming and labor-intensive, potentially leading to insufficient model generalization and difficulty adapting to individual differences among subjects. Furthermore, many previous studies have primarily conducted experiments in controlled laboratory settings, often on treadmills, which may not accurately represent natural gait patterns in real-world environments [36,37,38,39]. This limitation can affect the adaptability and practicality of gait phase recognition algorithms outside the laboratory.

To address these challenges, this paper proposes a model that combines CNN with HHO-SVM to process time series signals for human motion phase recognition. CNN is used to automatically extract complex, high-level features from the data. At the same time, HHO-SVM leverages its strong generalization ability to overcome overfitting and perform well with small, noisy datasets. This combination enhances the model’s accuracy and significantly improves robustness, particularly in the presence of sensor noise, a common issue in real-world applications. To validate the proposed approach, we established a hip exoskeleton experimental platform and collected data in natural environments, moving beyond controlled laboratory settings. The goal of this study is to develop a more robust and accurate model for human motion phase recognition in real-world conditions, addressing the limitations of traditional methods and improving the generalization of gait phase recognition models.

2. Methods

2.1. Experimental Procedures

The experimental protocol was designed to quantify the accuracy of phase estimation during natural gait in human subjects in an outdoor environment. To provide a relevant dataset for the gait phase recognition model, seven subjects were recruited from the laboratory for the gait data collection experiments. The subjects had an average age of 25.0 ± 2.1 years, height of 1.76 ± 0.08 m, and weight of 72.2 ± 16.3 kg. They were guided to walk on a flat outdoor surface, and the datasets collected from these subjects were divided into a training dataset and a test dataset, which were used to train, validate, and optimize the proposed algorithm offline. All subjects were informed of the scope and objectives of the experiment and provided written consent prior to their participation. The study protocol was approved by the Medical Ethics Committee of the Guangdong Jiangmen Maternal and Child Health Care Hospital, China.

2.2. Data Collection

2.2.1. Wearable Device Overview

As shown in Figure 1, we developed a hip exoskeleton robot for rehabilitation in this work. In this paper, we have researched the motion phase recognition approach based on this hip exoskeleton robot. The following constituents are the primary components of the hip exoskeleton robot: (1) The mechanical support structure can provide lumbar support for the human body, achieve a balanced transfer of torque from the hip motor, and improve the wearer’s carrying capacity. (2) The motion sensing system consists of IMU and force-sensitive resistor (FSR) sensors, which can obtain information about the motion attitude of the thigh and plantar pressure during walking. (3) The drive system selects the RMD series servo motor developed by Myactuator (Tokyo, Japan) for the hip joint drive, which has high power, lightweight, and miniaturization characteristics and can provide a peak torque of up to 18N-M for the hip joint. (4) The control system is a PCB control board designed according to the pin size of the STM32 microcontroller, which integrates all the modules into the PCB to realize the data acquisition and sending work of the sensors and to realize the auxiliary function of the exoskeleton robot hip joint according to the result of the recognition of the motion intention of the upper computer and the subsequent control technology. (5) The upper computer system comprises a laptop computer that offers a flexible and adaptable computing and communication processing platform. The laptop incorporates a built-in gait phase recognition model capable of analyzing gait data and producing accurate gait phase identification results. (6) The hip exoskeleton robot relies on a high-capacity lithium polymer battery to give electricity to its power supply system.

In this paper, FlexiForce A301 sensors were selected to acquire the plantar pressure signals, which are tiny polymer thick-film devices with a thickness of only about 0.6 mm and were arranged in the first metatarsal, fifth metatarsal, and heel regions of the insole to ensure that they fully reflect the pressure characteristics of the foot when walking. The IMU chose the LP-RESEARCH LPMS-CURS3 sensors from Japan with an accelerometer measuring range of ±16 g and a gyroscope measuring range of ±4000°/s. It also included a set of three-axis magnetometers and barometers. Previous research [40,41,42] verified that the leg and foot acceleration and angular velocity signals are highly periodic, which is the basis for gait phase recognition. Therefore, in this study, a miniature IMU is placed in the anterior thigh at 10 cm above the knee joint and the anterior tibia, avoiding the leg muscle groups. The IMU data are acquired using the CAN protocol in the integrated PCB circuit. The FSR signals are amplified by a signal amplifier. A built-in analog-to-digital converter circuit of the STM32 is then employed to acquire the FSR digital signal. The sampling frequencies are all 200 Hz.

2.2.2. Data Collection and Preprocessing

Firstly, subjects wore the hip exoskeleton robot described in Section 2.1, and the harness was tied around the leg using an elastic band to prevent the line from swinging and interfering with normal walking. Subjects were guided through walking experiments on a 50 m experimental field, where subjects were asked to walk according to their locomotor habits and to enter the locomotor state from a standing position for each experiment. The data were collected through the PCB control board and sent to the host PC for saving.

Prior to the data collection procedure, the sensors are calibrated to mitigate the impact of errors on the accuracy of motion phase recognition, as the manufacturing process inherently introduces bias in the IMU data. In this study, only the gyroscope and accelerometer data from IMU sensors are used, as magnetometers are affected by the Earth’s magnetic field during regular use. Each IMU is capable of providing signals for three-axis acceleration and three-axis rotation, resulting in a total of 12 dimensions that may be measured. The collected information is defined as follows:

(1)$\begin{array}{l}{G}_x=[g_{x1},g_{x2},\dots ,g_{xn}]\\ G_y=[g_{y1},g_{y2},\dots ,g_{yn}]\\ G_z=[g_{z1},g_{z2},\dots ,g_{zn}]\\ A_x=[a_{x1},g_{x2},\dots ,g_{xn}]\\ A_y=[a_{y1},g_{y2},\dots ,g_{yn}]\\ A_z=[a_{z1},g_{z2},\dots ,g_{zn}]\\ X=\left(G_{x,}{G}_y,G_{z,}{A}_{x,}{A}_y,A_z\right)\end{array}$

where $G_x,G_y,G_z$ is the angular velocity in the x, y, and z directions, respectively, $A_x,A_y,A_z$ is the acceleration in the x, y, and z directions, and n is the length of the input sequence. Based on the above data acquisition method, the three-axis acceleration and angular velocity acquired by the IMU sensor can be obtained. The curves plotted from the data based on the gait cycles, including the mean and standard deviation of acceleration and angular velocity for all gait cycles of the seven subjects, are shown in Figure 2.

Data normalization is a widely used technique in data preprocessing. It involves adjusting the input features of a dataset to a specified range, which helps improve the stability of optimization during the training of artificial neural networks [43]. This process ensures that data with varying magnitudes are transformed into a normalized and dimensionless form. In addition, data normalization also reduces the differences between individual subjects, simplifies the complexity of the computation, and helps to improve the accuracy of motion phase recognition. By using the following equation [44], it is possible to restrict the data to [−1, 1]:

(2)$\tilde{x}={\displaystyle \frac{(x_i-x_{\min })}{(x_{\max }-x_{\min })}}\cdot (y_{\max }-y_{\min })+y_{\min }$

where $x_i$ is the element in the normalized money input vector $i$. $x_{\max }$$x_{\min }$ are the maximum and minimum values of $x_i$, respectively. $y_{\max }$$y_{\min }$ are the upper and lower bounds of the normalized data, here 1 and −1, respectively.

Finally, the data were segmented using a sliding window segmentation method, moving from one sampling point to another at 30-ms intervals and assigning labels to each sliding window to ensure rationality and accuracy. When a sliding window contained multiple labels due to overlapping events, we resolved this conflict by assigning the most common label within the window. This approach ensures consistency and accurately represents the dominant gait phase within each window, reducing ambiguity in the labeling process.

2.2.3. Segmentation and Labeling

The most obvious aspect of how humans move while walking is the gait cycle. The typical human gait cycle consists of two successive ground landings by a specific area on the heel of one leg. As shown in Figure 3, redundant phase delineation can instead impede human motion. During human walking, several gait events are used to distinguish distinct phases of human motion, such as heel strike (HS), toe strike (TS), commonly referred to as flat foot (FF), heel off the ground (HO), and toe off the ground (TO). The approach generally used for partitioning the human motion cycle is based on the ankle’s rotational angle trend. The process is categorized into four distinct phases: the plantar flexion phase (CP), the controlled dorsiflexion phase (CD), the powered plantar flexion phase (PP), and the swing phase (SW). These four stages may alternatively be distinguished by four gait events: HS, FF, HO, and TO. This split of phases enables us to categorize gait phases using FSR sensors to identify when the heel and toe make contact with the ground.

Plantar pressure signals and footswitches are often considered the gold standard in wearable sensors [45]. In this paper, however, FSR signals are used exclusively as ground truth labels to create labels for the input data due to the severe limitations on the mounting location of plantar pressure sensors and their lifespan constraints. The IMU detects just the motion signals of the lower limbs as inputs to the motion phase detection model. This approach eliminates the issues of wearing discomfort caused by excessive sensors and the reduced accuracy of the model identification when utilizing only a single signal.

The data in Figure 4 are derived from Volunteer 1, showing the sample points from 200 to 400, which are time-sequenced samples collected from the foot pressure sensor. The figure illustrates the trend of the plantar pressure signal data during the gait cycle, as the FSR signal can be affected by several factors. For example, the FSR is mounted on the insole and will detect forces exerted due to flexing or tightening of the insole and shoe, especially when the foot is in the swing phase of human movement, which flexes the toes and results in increased pressure at the ball of the foot. This also explains the fact that the FSR raw signal is not in the 0 position when the foot is in the swing phase, and the pressure signal at the ball of the foot is particularly pronounced. In addition, during walking, the pressure values may not maintain balance, especially when individuals are focused on the movement of the swinging leg. While the foot arch can influence plantar pressure distribution, this factor is not considered in this study; the focus is primarily on developing and validating the proposed gait recognition algorithm. This approach helps streamline the analysis and minimize potential confounding factors. Future research may explore the impact of foot arch variations to enhance an understanding of how anatomical differences affect gait recognition outcomes.

In this study, the labeling procedure for each gait phase was achieved by setting the trigger threshold for the pressure signal detected by FSR. For the above reasons, setting appropriate thresholds according to different subjects is particularly important. In this study, the plantar pressure signals were first normalized to the data and converted to the 0–1 interval, which can effectively reduce the differences between individual subjects. Then, threshold1 and threshold2 were set for the heel and forefoot pressure signals, respectively, to complete the labeling of the key events. These thresholds were determined by observing the pressure characteristics of each subject, with distinct values designed for each individual to ensure accurate event labeling. The specific rules are as follows:

1. Heel strike (HS): initial ipsilateral heel contact: FSR Heel > threshold1, FSR Thumb < threshold2, n above > 20

2. Flatfoot (FF): Initial ipsilateral foot contact: FSR Heel > threshold2, FSR Thumb > threshold2, n above > 20

3. Heel off the ground (HO): ipsilateral heel raise: FSR Heel < threshold2, FSR Thumb > threshold2, n above > 20

4. Swing (SW): ipsilateral foot lift: FSR Heel < threshold2, FSR Thumb < threshold2, n above > 20

FSR Heel and FSR Thumb are the heel and foot pressure, respectively. Whether the length of consecutive data segments meets the requirements is judged by ‘n above’—set to 20 data points in this paper—to avoid misidentification for mutant data. Figure 5 shows the phase markers of plantar pressure data during seven consecutive walking gait cycles of one subject.

2.3. Gait Phase Estimator Model

This research presents a combinatorial model based on CNN, which utilizes raw acceleration and angular velocity data. The model is meant to analyze human activity data collected by the exoskeleton robot system mentioned in Section 2.1. The IMU is used to measure human motion data, which is then input into the model after extracting features using a sliding window. The FSR is used to measure the ground truth gait phase data, and a threshold algorithm is used to label the key events. This labeled data are then used as the ground truth gait phase for training and validating the network. Figure 6 illustrates the sequence of steps in the method.

2.3.1. SVM Classifier

SVM is a supervised learning model used for classification and regression tasks [46]. The key idea behind SVM is to find the optimal hyperplane that maximally separates data points of different classes in a feature space. SVMs are particularly useful for high-dimensional data and can handle non-linear relationships using kernel functions, such as the Radial Basis Function (RBF). The RBF kernel enables SVM to project data into a higher-dimensional space, making it easier to separate data that are not linearly separable in the original feature space. In this study, SVM is applied for its robust performance in small sample datasets and high-dimensional data classification.

2.3.2. CNN Classifier

CNN is a neural network that automatically extracts features from raw data through convolutional layers, followed by classification using a fully connected layer. The feature extraction process includes convolutional filtering and subsampling (max pooling), which reduces dimensionality and mitigates overfitting. In this study, the batch size is chosen to be 64 in order to accelerate the convergence of the network. We present a reduced CNN structure whereby the data are preprocessed and subsequently handled using a sliding window with a window size set to 6-time steps, obtaining an input dimension of the data of 64 × 6 × 12. The size of the convolution kernel is set to 1 × 3 following two feature extraction layers ($L_i$ containing one layer of one-dimensional convolution and maximum pooling $i=1,2$). The parameter stride of the convolutional layer is set to 1, the padding to 1, and the number of parameters is lowered by the sparse connectivity of the convolutional layer and the parameter sharing. This ensures that the input and output dimensions of the convolutional layer are constant and will help somewhat lower overfitting and speed up the learning pace by adding dropout as a normalization method in the fully connected layer [47]. Setting dropout $=0.5$ in each training batch ignores half of the feature detectors as a normalizing method. Hence, half of the hidden value nodes are 0, and the specific data flow is shown in Table 1, which can greatly lower the overfitting phenomenon and improve the generalizing ability of the model.

The combination of CNN and SVM leverages each individual’s capabilities to enhance classification performance [48]. The convolutional and pooling layers of CNN automatically collect features from the input signal. At the same time, the SVM substitutes the fully connected layers of the CNN network to enhance classification performance. Figure 7 displays the architectural arrangement of the CNN-SVM model.

2.3.3. Improved Algorithm

Currently, researchers have integrated numerous intelligent optimization algorithms into SVMs for hyperparameter search, such as Particle Swarm Optimization (PSO), Genetic Algorithm (GA), and Whale Optimization Algorithm (WOA) [49], which are commonly used to perform this task. While heuristic methods like GA and PSO do not inherently face the same level of risk of falling into local optima as gradient-based methods, they may still encounter difficulties in maintaining the balance between exploration and exploitation during the search process, leading to suboptimal solutions in some cases. Furthermore, adjusting the optimization algorithm’s parameters to enhance global search capability using prior knowledge remains challenging, and secondary parameter tuning may be required.

In contrast, the HHO algorithm, as a newer optimization method, incorporates several exploratory and exploitative mechanisms and consequently has shown a higher potential in jumping out of local optimum solutions [50]. Therefore, based on the demonstrated potential of HHO in other domains, this paper introduces it as a promising approach to address the challenges of blind SVM parameter selection [51,52].

HHO is a biological optimization algorithm simulating the hunting behavior of Harris Hawks, divided into global exploration and local exploitation based on prey escape energy. This paper proposes two improvement strategies to enhance the balance between local and global searches, which will be detailed in later sections.

For each position update: prey jump strength set to a random number J from 0 to 2 and initialized prey escape energy set to a random number $E_0$ from 0 to 1, then by updating the prey escape energy $E$:

(3)$E=2E_0(1-{\displaystyle \frac{t}{T}})$

t is the current iteration count, while T is the total iteration count.

• (1). Global search phase when $E\ge 1$: The Harris Hawk randomly perches and searches for prey using two strategies:

  (4)$X(t+1)=\left\{\begin{array}{l}{X}_{rand}(t)-r_1\left|X_{rand}(t)-2r_{2}X(t)\right| q\ge 0.5\\ \left[X_{rabbit}(t)-X_m(t)\right]-r_3\left[lb+r_4(ub-lb)\right] q<0.5\end{array}\right.$

  where $X(t)$ and $X(t+1)$ are the positions of the individuals at the current and next iterations, respectively. $X_{rand}(t)$ is the position of the randomly selected individual. $X_{rabbit}(t)$ is the position of the prey, i.e., the position of the individual with optimal fitness. $r_1,r_2,r_3,r_4,q$ are random numbers between [0, 1]. $q$ is used to randomly select the strategy to be employed. $X_m(t)$ is the average position of the individuals. $ub,lb$ are the upper and lower bounds of the Harris hawk population search space.

• (2). Local exploration phase when $E<1$:

  • (a). Development phase: define a random number $r$ between 0 and 1. When $0.5\le \left|E\right|<1$ and $r\ge 0.5$, a soft siege strategy is adopted for position update:

    (5)$X(t+1)=X_{rabbit}(t)-X(t)-E\left|J{X}_{rabbit}(t)-X(t)\right|$

  • (b). When $\left|E\right|<0.5$ and $r\ge 0.5$, a hard siege is taken for a positional follow-up:

    (6)$X(t+1)=X_{rabbit}(t)-E\left|X_{rabbit}(t)-X(t)\right|$

  • (c). When $\left|E\right|<0.5$ and $r<0.5$, an asymptotically fast swooping hard-envelope strategy is adopted for position updating:

    (7)$X(t+1)=\left\{\begin{array}{l}Y=X_{rabbit}(t)-E\left|J{X}_{rabbit}(t)-X_m(t)\right|\hspace{1em}f(Y)<f(X(t))\\ Z=Y+S\ast LF(D)\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em} f(Z)<f(X(t))\end{array}\right.$

  • (d). When $1>\left|E\right|\ge 0.5$ and $r<0.5$, the soft envelopment strategy of asymptotic fast dive is adopted for position update:

    (8)$X(t+1)=\left\{\begin{array}{l}Y=X_{rabbit}(t)-E\left|J{X}_{rabbit}(t)-X(t)\right|\hspace{1em}f(Y)<f(X(t))\\ Z=Y+S\ast LF(D)\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}f(Z)<f(X\left(\mathrm{t}\right))\end{array}\right.$

    where $f$ is the cross-validation function of the support vector machine. $S$ is a 2-dimensional random vector, elements of which are random numbers between [0, 1], and LF is the mathematical expression for the Levy Flight.

Improvement 1: Inspired by the idea of an artificial bee colony algorithm [53], to reduce the possibility of the HHO algorithm from local optimal search in the iteration process, a limit threshold is introduced into the algorithm, and if an optimal solution has not been obtained within the number of times of the limit threshold, then it will jump out of the local optimal search and enter into the global exploration to carry out parameter optimization. In this paper, we set the limit threshold to 10 through experiments.

Improvement 2: Secondly, to increase the diversity of the population and expand the search range, we consider that both the current and its inverse solutions may yield optimal results. Therefore, an inverse search is performed at the solution’s location, inspired by the opposition-based learning proposed by Tizhoosh [54]. Assuming that the search space is L-dimensional, the inverse solution formula for the $i-th$ individual in the $j-th$ dimensional space is as follows:

(9)${\tilde{X}}_{ij}=L{B}_j+U{B}_j-X_{ij}$

where $L{B}_j,U{B}_j$ are the upper and lower bounds (bounds) on the $j-th$ dimensional. The inverse solution of the current solution for the optimization search iteration increases the search range and enables a more probable optimal solution to be obtained.

2.4. Architecture of the CNN-HHO-SVM Combined Model

The implementation process of the combined CNN-HHO-SVM model is shown in Figure 8. It can be summarized in the following steps:

Step 1: The unprocessed IMU data undergo normalization before being subjected to sliding window processing. The treated data are then used to train the proposed CNN classifier until the model converges.

Step 2: Initiate the HHO algorithm’s parameters, construct the support vector machine classifier, choose the Gaussian kernel function, and then produce the penalty factor $c$ and kernel function parameter $\gamma$ at random to serve as the algorithm’s starting points.

Step 3: Optimize the parameters of the SVM classifier using the HHO algorithm based on the defined strategies. The classification accuracy obtained from cross-validation serves as the fitness value for evaluating the performance of the optimized parameters.

Step 4: Determine whether the maximum number of iterations is reached. If not, jump to Step 3 to continue the search and output the optimal solution location as $c$ and $\gamma$ of the SVM classifier after reaching the iteration.

Step 5: Once the training is finished, CNN is used to extract the feature vectors from the test data. These feature vectors are then categorized using the learned optimum parameter SVM to accurately identify the four motion phases.

2.5. Statistical Analysis

In this study, the Bootstrap method was applied to calculate 95% confidence intervals (CIs) for each model’s performance metrics. Given that the data were derived from the performance metrics of seven subjects, the non-parametric Bootstrap method was selected due to its ability to avoid assumptions about the underlying distribution, making it particularly suitable for small sample sizes. The procedure involved resampling the data with replacement to generate new datasets of equal size, calculating the mean performance metric for each resampled dataset, and repeating this process 1000 times. The 95% CIs were then determined by calculating the 2.5th and 97.5th percentiles of the resulting distribution.

To further evaluate the performance differences between the models, one-way ANOVA was conducted to compare the overall accuracy of the five models. This method assessed whether there were statistically significant differences in the models’ performance, ensuring that the observed variations were not due to random chance.

3. Results

This section focuses on presenting the key data from the experiments, including the model training process and performance evaluation.

3.1. Performance Evaluation with Cross-Validation

Figure 9a shows the CNN learning curve during the training process. The accuracy of the training and test sets eventually converge, with the accuracy of the test set stabilizing at about 94% and the training set at about 96%. This preliminary verifies that the established CNN neural network achieves good classification performance.

Figure 9b shows the variation in the classification error of CNN for the subjects with the training batches. We can see that the training error asymptotically stabilizes after the 600th period. The 192 features from layer FC2 of the trained CNN network are used as new feature vectors representing each input signal and fed to the SVM for learning and testing. To construct the SVM in a hybrid model, we use an RBF kernel. A 5-fold cross-validation method is applied to the training dataset using the HHO optimization algorithm to find the optimal kernel parameters $\gamma$ and penalty factor $c$. These parameters are then used to train the hybrid model.

3.2. Performance of Gait Phase Detection

The performance of the dataset is first analyzed using the standard SVM classifier and the proposed classifier in this paper; the results are shown in Figure 10a. Although the predicted values of the SVM in the HS phase match more closely with the actual values, the other three gait phases suffer from large errors. Figure 10b shows the recognition results using the method of CNN-HHO-SVM, and overall, it can be seen that the phases identified using the proposed method closely correspond to the reference phase.

Table 2 displays the empirical outcomes of the five models over the four stages of HS, FF, HO, and SW. The table includes measurements of accuracy, precision, recall, and F1-score. Based on the information provided, CNN-HHO-SVM outperformed the other four models in all gait phases. CNN performed poorly across the four phases, with a precision score of only 78.57% in the FF phase, while achieving precision scores exceeding 94% in the other three phases. In comparison, CNN-SVM achieved a precision of 81.54% in the FF phase, with all sub-phase scores surpassing 80%. The CNN-LSTM model demonstrated strong performance in the HS and SW phases but struggled in the FF phase, where its precision was 79.46%, slightly lower than CNN-SVM’s 81.54%. Although its recall and F1 scores were competitive in other phases, the FF phase remained challenging. The CNN-GA-SVM model further improved precision in the FF phase to 85.11%, outperforming CNN and CNN-SVM.

Figure 11 presents empirical statistical results for accuracy, macro-precision, macro-recall, and macro-F1 scores across the four models. Compared to CNN, CNN-SVM shows improvements in all metrics. However, the overall performance still falls short of the results achieved by the algorithm proposed in this study. The CNN-GA-SVM model also demonstrates improvements, thanks to the genetic algorithm’s ability to optimize hyperparameters. However, compared to CNN-HHO-SVM, the genetic algorithm (GA) is less efficient in the search process. Under the same number of iterations, the HHO algorithm exhibits a faster convergence rate and higher recognition accuracy.

To assess the classification performance of the models, we first used one-way ANOVA to compare the overall accuracy among the five models. The ANOVA was applied to determine whether there were any statistically significant differences in the mean accuracy scores across the models. The results revealed a statistically significant difference (p < 0.0025).

Following this, Bootstrap resampling was applied to estimate the 95% confidence intervals (CIs) for the mean performance of each model. This non-parametric method involves resampling the observed data with replacement to create multiple new datasets and calculating the mean performance for each resampled dataset. By repeating this process 1000 times, the 95% CIs were determined based on the 2.5th and 97.5th percentiles of the resampled means. The resulting CIs for each model are presented in Table 3.

The overall accuracy achieved by the seven subjects using the CNN, CNN-SVM, CNN-LSTM, CNN-GA-SVM, and CNN-HHO-SVM models was 88.35% ± 3.5%, 91.75% ± 1.8%, 90.2% ± 3.1%, 92.75% ± 1.9%, and 93.67% ± 1.32%, respectively. Figure 12 presents the confusion matrices based on the sample dataset. The CNN-HHO-SVM model outperforms the other models in classification performance. While the FF and HO phases show relatively lower recognition rates, the overall performance is better compared to the other models, with improvements of 0.5%, 8.8%, 0.2%, and 2.4% in the HS, FF, HO, and SW phases, respectively.

4. Discussion

4.1. Analysis of Accuracy and Generalizability

Based on precision, recall, F1-score, and the confusion matrix in Table 2, the results indicate that all models exhibit higher identification accuracy for the HS and SW phases compared to the FF and HO phases. This conclusion is directly supported by the data, as all models consistently show higher performance metrics in these phases. This difference can likely be attributed to the shorter durations of the FF and HO phases within the full gait cycle, which makes these phases more susceptible to misidentification and confusion with adjacent sub-phases.

In terms of model performance, the CNN-LSTM model is the weakest among the four. Despite being designed for temporal data, the lower precision, recall, and F1-scores observed in the LSTM model, particularly for the FF phase, suggest that LSTM may struggle to capture the specific features required for gait recognition. Additionally, LSTM models typically require larger datasets and longer training times to achieve better generalization, which could explain its relatively lower accuracy and macro-F1 scores.

Based on our Statistical Analysis, Table 3 presents the 95% confidence intervals (CIs) for each model’s mean performance. The CNN-HHO-SVM model shows the highest performance, with the narrowest CI range (93.03%, 94.61%), while the CNN model has the lowest and widest CI range (86.69%, 90.63%), indicating lower consistency. Models using optimization methods (CNN-SVM, CNN-LSTM, and CNN-GA-SVM) exhibit similar, more reliable performance with CIs above 90%. These results underscore the advantage of optimization techniques in enhancing both performance and consistency.

The comparison between CNN-HHO-SVM and CNN-GA-SVM under the same number of iterations confirms that the HHO algorithm demonstrates a faster convergence rate and higher recognition accuracy compared to the GA when applied to the same dataset. This is directly supported by the data, as the HHO-based model outperforms the GA-based model in all phases, particularly in the FF and HO phases. This highlights the superior global search capabilities and optimization efficiency of HHO. The HHO’s superior global search capabilities enable it to find optimal solutions in fewer iterations. Unlike GA, which relies on a genetic evolutionary process that can converge slowly, HHO mimics the behavior of Harris’ hawks during hunting, dynamically adjusting its exploration and exploitation strategies. These theoretical advantages are directly supported by the performance metrics, which show faster convergence and higher accuracy in the HHO-based models.

4.2. Limitations and Potential Improvements

This study has some limitations that should be addressed in future work. First, the study involved only healthy individuals. To better assess the model’s robustness and potential clinical applications, future work should include individuals with pathological gait patterns, such as those with stroke or musculoskeletal disorders. This would help evaluate the model’s adaptability to various gait abnormalities and refine its algorithms for broader clinical use, such as in rehabilitation and assistive technologies for individuals with impaired mobility. Second, all experiments were conducted with the hip exoskeleton in zero-impedance mode, meaning no active assistance was provided. Future studies should test the model in dynamic environments where the exoskeleton provides active support, which may affect gait phase recognition performance. Additionally, the motion phase recognition algorithms were tested on a laptop, and transitioning to a microcontroller-based embedded system will be a key challenge. This shift requires optimizing the model for real-time applications, focusing on computational efficiency and low-power hardware suitable for wearable devices. Techniques like model pruning could be explored to improve processing speed and reduce computational demands. Finally, expanding the sample size and testing in real-world environments will be crucial for validating the model’s robustness. This will also help assess the system’s applicability for clinical and industrial uses, particularly in assistive technologies and real-time applications.

5. Conclusions

In this paper, a hybrid CNN-HHO-SVM model for gait phase recognition was proposed and validated using a custom dataset collected from seven subjects during horizontal walking with a hip rehabilitation exoskeleton robot. The experimental results show that the CNN-HHO-SVM model outperforms existing techniques in recognizing the four stages of the gait cycle with higher accuracy and stronger generalization ability. This work significantly advances gait phase recognition for lower-limb exoskeletons, enhancing their adaptability in real-world rehabilitation settings. Future research will expand the dataset to improve model robustness and explore techniques such as model pruning and quantization for real-time processing on microcontrollers.

Author Contributions

Conceptualization, H.W.; Project administration, F.D.; Resources, D.L.; Software, M.W.; Supervision, Y.S.; Validation, Z.P. All authors have read and agreed to the published version of the manuscript.

Institutional Review Board Statement

The study was conducted in accordance with the Declaration of Helsinki, and approved by the Ethics Committee of Jiangmen Maternity and Child Health Care Hospital, Jiangmen, Guangdong, China (protocol code 2024147).

Informed Consent Statement

Informed consent was obtained from all subjects involved in the study.

Data Availability Statement

The datasets generated and analyzed during the current study are available in the GitHub repository at https://github.com/PaceMar1k/HipExoskeleton_PhaseRecognition_Data.git (accessed on 14 December 2024).

Conflicts of Interest

The authors declare no conflicts of interest.

Footnotes

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Figures and Tables

Figure 1. Hip exoskeleton robot structure.

Figure 2. Walking data of seven subjects’ right legs. Mean ± 1 SD across the seven subjects for the measured signals Accel. and Ang. vel. (shaded area). The gait cycle is defined as the period of time from when the heel first makes contact with the ground (0% of the cycle) to when the same heel makes contact with the ground again (100% of the cycle).

Figure 3. Representation of the human gait cycle.

Figure 4. Variations in plantar pressure throughout the full gait cycle.

Figure 5. Plantar pressure data and gait event during seven consecutive walking gait cycles in Subject 1. The orange dashed lines represent the events detected using the threshold-based algorithm introduced earlier. Four events of the gait cycle are marked, labeled from left to right as HS, TS, HO, and TO.

Figure 6. Structure of the proposed model.

Figure 7. CNN architecture.

Figure 8. CNN-HHO-SVM architecture.

Figure 9. CNN learning curves: (a) Trend of loss function of CNN on the customized dataset; (b) Accuracy of CNN on the customized dataset.

Figure 10. Current phase prediction result for volunteer 1: (a) Estimated phases obtained with SVM; (b) Estimated phases obtained with CNN-HHO-SVM.

Figure 11. Evaluation of different models based on accuracy, macro precision, macro recall, and macro F1 score of seven subjects.

Figure 12. Confusion matrix representation for each of the five models, evaluated across four gait phases and seven subjects: (a) CNN model; (b) CNN-SVM; (c) CNN-LSTM; (d) CNN-GA-SVM; (e) CNN-HHO-SVM.

Table 1

CNN detailed data flow.

Layer	Input Size (Height × Weight × Channels)	Output Size
Input	1 × 12 × 6	1 × 12 × 6
Conv1	1 × 12 × 6	1 × 24 × 6
Pool1	1 × 24 × 6	1 × 24 × 3
Conv1	1 × 24 × 3	1 × 48 × 3
Pool1	1 × 48 × 3	1 × 48 × 2
FC (Hidden)	96	192
FC (Output)	192	4

Table 2

Comparison of precision, recall, and F1-scores for different models over four phases.

Model	Phase	Precision (%)	Recall (%)	F1-Score (%)
CNN	Heel Strike	94.10	97.00	95.53
	Foot Flat	78.57	74.58	76.52
	Heel Off	95.03	92.66	93.83
	Swing Phase	98.44	98.37	98.41
CNN-SVM	Heel Strike	94.36	96.04	9519
	Foot Flat	81.54	81.95	81.42
	Heel Off	93.91	94.33	94.12
	Swing Phase	98.61	96.70	97.65
CNN-LSTM	Heel Strike	92.42	95.77	94.06
	Foot Flat	79.46	84.45	80.45
	Heel Off	94.60	93.55	94.07
	Swing Phase	98.55	95.81	97.16
CNN-GA-SVM	Heel Strike	95.77	95.77	95.41
	Foot Flat	82.35	82.95	81.65
	Heel Off	94.37	95.11	94.74
	Swing Phase	98.58	98.06	98.32
CNN-HHO-SVM	Heel Strike	97.41	96.66	96.84
	Foot Flat	85.11	84.71	83.15
	Heel Off	94.98	94.66	94.82
	Swing Phase	98.58	98.22	98.40

Table 3

95% Confidence Intervals for the Mean Performance of Each Model Based on Bootstrap Resampling.

Model	95% Confidence Interval (Lower, Upper)
CNN	(86.69%, 90.63%)
CNN-SVM	(90.15%, 92.14%)
CNN-LSTM	(87.87%, 92.19%)
CNN-GA-SVM	(91.68%, 94.74%)
CNN-HHO-SVM	(93.03%, 94.61%)

References

1. Ma, Y.; Wu, X.; Yi, J.; Wang, C.; Chen, C. A review on human–exoskeleton coordination towards lower limb robotic exoskeleton systems. Int. J. Robot. Autom.; 2019; 34, pp. 431-451. [DOI: https://dx.doi.org/10.2316/J.2019.206-0193]

2. Bernhardt, M.; Frey, M.; Colombo, G.; Riener, R. Hybrid force-position control yields cooperative behaviour of the rehabilitation robot LOKOMAT. Proceedings of the 9th International Conference on Rehabilitation Robotics, 2005. ICORR 2005; Chicago, IL, USA, 28 June–1 July 2005; pp. 536-539. [DOI: https://dx.doi.org/10.1109/ICORR.2005.1501159]

3. Banala, S.K.; Kim, S.H.; Agrawal, S.K.; Scholz, J.P. Robot assisted gait training with active leg exoskeleton (ALEX). Proceedings of the 2008 2nd IEEE RAS & EMBS International Conference on Biomedical Robotics and Biomechatronics; Scottsdale, AZ, USA, 19–22 October 2008; pp. 653-658. [DOI: https://dx.doi.org/10.1109/BIOROB.2008.4762885]

4. Zeilig, G.; Weingarden, H.; Zwecker, M.; Dudkiewicz, I.; Bloch, A.; Esquenazi, A. Safety and tolerance of the ReWalk^TM exoskeleton suit for ambulation by people with complete spinal cord injury: A pilot study. J. Spinal Cord Med.; 2012; 35, pp. 96-101. [DOI: https://dx.doi.org/10.1179/2045772312Y.0000000003]

5. Kawamoto, H.; Kandone, H.; Sakurai, T.; Ariyasu, R.; Ueno, Y.; Eguchi, K.; Sankai, Y. Development of an assist controller with robot suit HAL for hemiplegic patients using motion data on the unaffected side. Proceedings of the 2014 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society; Chicago, IL, USA, 26–30 August 2014; pp. 3077-3080. [DOI: https://dx.doi.org/10.1109/EMBC.2014.6944273]

6. Kazerooni, H.; Racine, J.-L.; Huang, L.; Steger, R. On the Control of the Berkeley Lower Extremity Exoskeleton (BLEEX). Proceedings of the 2005 IEEE International Conference on Robotics and Automation; Barcelona, Spain, 18–22 April 2005; pp. 4353-4360. [DOI: https://dx.doi.org/10.1109/ROBOT.2005.1570790]

7. Xia, Y.; Li, J.; Yang, D.; Wei, W. Gait Phase Classification of Lower Limb Exoskeleton Based on a Compound Network Model. Symmetry; 2023; 15, 163. [DOI: https://dx.doi.org/10.3390/sym15010163]

8. Manchola, M.D.S.; Bernal, M.J.P.; Munera, M.; Cifuentes, C.A. Gait Phase Detection for Lower-Limb Exoskeletons using Foot Motion Data from a Single Inertial Measurement Unit in Hemiparetic Individuals. Sensors; 2019; 19, 2988. [DOI: https://dx.doi.org/10.3390/s19132988] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31284619]

9. Sánchez-Manchola, M.; Arciniegas-Mayag, L.; Múnera, M.; Bourgain, M.; Provot, T.; Cifuentes, C.A. Effects of stance control via hidden Markov model-based gait phase detection on healthy users of an active hip-knee exoskeleton. Front. Bioeng. Biotechnol.; 2023; 11, 1021525. [DOI: https://dx.doi.org/10.3389/fbioe.2023.1021525]

10. Barker, S.; Craik, R.; Freedman, W.; Herrmann, N.; Hillstrom, H. Accuracy, reliability, and validity of a spatiotemporal gait analysis system. Med. Eng. Phys.; 2006; 28, pp. 460-467. [DOI: https://dx.doi.org/10.1016/j.medengphy.2005.07.017] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/16122966]

11. Prakash, C.; Kumar, R.; Mittal, N. Recent developments in human gait research: Parameters, approaches, applications, machine learning techniques, datasets and challenges. Artif. Intell. Rev.; 2018; 49, pp. 1-40. [DOI: https://dx.doi.org/10.1007/s10462-016-9514-6]

12. Du, F.; Chen, J.; Wang, X. Human motion measurement and mechanism analysis during exoskeleton design. Proceedings of the 2016 23rd International Conference on Mechatronics and Machine Vision in Practice (M2VIP); Nanjing, China, 28–30 November 2016; pp. 1-5. [DOI: https://dx.doi.org/10.1109/M2VIP.2016.7827336]

13. Song, J.; Zhu, A.; Tu, Y.; Mao, H.; Zhang, X. Adaptive neural fuzzy reasoning method for recognizing human movement gait phase. Robot. Auton. Syst.; 2022; 153, 104087. [DOI: https://dx.doi.org/10.1016/j.robot.2022.104087]

14. Chinmilli, P.T.; Redkar, S.; Zhang, W.; Sugar, T. A review on wearable inertial tracking based human gait analysis and control strategies of lower-limb exoskeletons. Int. Robot. Autom. J.; 2017; 3, 00080.

15. Strick, J.A.; Farris, R.J.; Sawicki, J.T. A Novel Gait Event Detection Algorithm Using a Thigh-Worn Inertial Measurement Unit and Joint Angle Information. J. Biomech. Eng.; 2024; 146, 044502. [DOI: https://dx.doi.org/10.1115/1.4064435]

16. Sun, M.; Cui, S.; Wang, Z.; Luo, H.; Yang, H.; Ouyang, X.; Xu, K. A laser-engraved wearable gait recognition sensor system for exoskeleton robots. Microsyst. Nanoeng.; 2024; 10, 50. [DOI: https://dx.doi.org/10.1038/s41378-024-00680-x]

17. Reza, S.M.T.; Ahmad, N.; Choudhury, I.A.; Ghazilla, R.A.R. A study on muscle activities through surface EMG for lower limb exoskeleton controller. Proceedings of the 2013 IEEE Conference on Systems, Process & Control (ICSPC); Kuala Lumpur, Malaysia, 13–15 December 2013; pp. 159-163. [DOI: https://dx.doi.org/10.1109/SPC.2013.6735124]

18. Zhang, Z.; Wang, Z.; Lei, H.; Gu, W. Gait phase recognition of lower limb exoskeleton system based on the integrated network model. Biomed. Signal Process. Control; 2022; 76, 103693. [DOI: https://dx.doi.org/10.1016/j.bspc.2022.103693]

19. Chen, W.; Li, J.; Zhu, S.; Zhang, X.; Men, Y.; Wu, H. Gait Recognition for Lower Limb Exoskeletons Based on Interactive Information Fusion. Appl. Bionics Biomech.; 2022; 2022, 9933018. [DOI: https://dx.doi.org/10.1155/2022/9933018]

20. Kolaghassi, R.; Al-Hares, M.K.; Sirlantzis, K. Systematic Review of Intelligent Algorithms in Gait Analysis and Prediction for Lower Limb Robotic Systems. IEEE Access; 2021; 9, pp. 113788-113812. [DOI: https://dx.doi.org/10.1109/ACCESS.2021.3104464]

21. Mutlag, W.K.; Ali, S.K.; Aydam, Z.M.; Taher, B.H. Feature Extraction Methods: A Review. J. Phys. Conf. Ser.; 2020; 1591, 012028. [DOI: https://dx.doi.org/10.1088/1742-6596/1591/1/012028]

22. Lv, P.; Zhang, C.; Yi, F.; Yuan, T.; Li, S.; Zhang, M. A Novel FSVM with PSO for gait phase detection based on elastic pressure sensing insole system. Int. J. Intell. Robot. Appl.; 2024; 8, pp. 596-608. [DOI: https://dx.doi.org/10.1007/s41315-024-00334-1]

23. Farah, J.D.; Baddour, N.; Lemaire, E.D. Gait phase detection from thigh kinematics using machine learning techniques. Proceedings of the 2017 IEEE International Symposium on Medical Measurements and Applications (MeMeA); Rochester, MN, USA, 7–10 May 2017; pp. 263-268. [DOI: https://dx.doi.org/10.1109/MeMeA.2017.7985886]

24. Ziegler, J.; Gattringer, H.; Mueller, A. Classification of Gait Phases Based on Bilateral EMG Data Using Support Vector Machines. Proceedings of the 2018 7th IEEE International Conference on Biomedical Robotics and Biomechatronics (Biorob); Enschede, The Netherlands, 26–29 August 2018; pp. 978-983. [DOI: https://dx.doi.org/10.1109/BIOROB.2018.8487750]

25. Gu, J.; Wang, Z.; Kuen, J.; Ma, L.; Shahroudy, A.; Shuai, B.; Liu, T.; Wang, X.; Wang, G.; Cai, J. et al. Recent advances in convolutional neural networks. Pattern Recognit.; 2018; 77, pp. 354-377. [DOI: https://dx.doi.org/10.1016/j.patcog.2017.10.013]

26. LeCun, Y.; Bengio, Y.; Hinton, G. Deep learning. Nature; 2015; 521, pp. 436-444. [DOI: https://dx.doi.org/10.1038/nature14539]

27. Yang, J.B.; Nguyen, M.N.; San, P.P.; Li, X.L.; Krishnaswamy, S. Deep Convolutional Neural Networks on Multichannel Time Series for Human Activity Recognition. Ijcai; 2015; 15, pp. 3995-4001.

28. Sawada, A.; Miyagawa, T.; Ebihara, A.; Yachida, S.; Hosoi, T. Convolutional Neural Networks for Time-dependent Classification of Variable-length Time Series. Proceedings of the 2022 International Joint Conference on Neural Networks (IJCNN); Padua, Italy, 18–23 July 2022; pp. 1-8. [DOI: https://dx.doi.org/10.1109/IJCNN55064.2022.9892605]

29. Lecun, Y.; Bottou, L.; Bengio, Y.; Haffner, P. Gradient-based learning applied to document recognition. Proc. IEEE; 1998; 86, pp. 2278-2324. [DOI: https://dx.doi.org/10.1109/5.726791]

30. Niu, X.-X.; Suen, C.Y. A novel hybrid CNN–SVM classifier for recognizing handwritten digits. Pattern Recognit.; 2012; 45, pp. 1318-1325. [DOI: https://dx.doi.org/10.1016/j.patcog.2011.09.021]

31. Athavale, V.A.; Gupta, S.C.; Kumar, D.; Savita, S. Human Action Recognition Using CNN-SVM Model. Adv. Sci. Technol.; 2021; 105, pp. 282-290. [DOI: https://dx.doi.org/10.4028/www.scientific.net/AST.105.282]

32. Wu, X.; Zhao, L.; Guo, S.; Zhang, L. Pedestrian inertial navigation based on CNN-SVM gait recognition algorithm. J. Phys. Conf. Ser.; 2021; 1903, 012043. [DOI: https://dx.doi.org/10.1088/1742-6596/1903/1/012043]

33. Agarap, A.F. An Architecture Combining Convolutional Neural Network (CNN) and Support Vector Machine (SVM) for Image Classification. arXiv; 2019; [DOI: https://dx.doi.org/10.48550/arXiv.1712.03541] arXiv: 1712.03541

34. Shi, L.-F.; Liu, Z.-Y.; Zhou, K.-J.; Shi, Y.; Jing, X. Novel Deep Learning Network for Gait Recognition Using Multimodal Inertial Sensors. Sensors; 2023; 23, 849. [DOI: https://dx.doi.org/10.3390/s23020849]

35. Kreuzer, D.; Munz, M. Deep Convolutional and LSTM Networks on Multi-Channel Time Series Data for Gait Phase Recognition. Sensors; 2021; 21, 789. [DOI: https://dx.doi.org/10.3390/s21030789]

36. You, Y.; Lin, J.; Zhang, S.; Huo, W.; Han, J.; Yu, N. An Adaptive Oscillator-Driven Gait Phase Model for Continuous Motion Estimation Across Speeds. IEEE Trans. Instrum. Meas.; 2024; 73, pp. 1-14. [DOI: https://dx.doi.org/10.1109/TIM.2024.3450074]

37. Wang, E.; Chen, X.; Li, Y.; Fu, Z.; Huang, J. Lower limb Motion Intent Recognition Based on Sensor Fusion and Fuzzy Multi-task Learning. IEEE Trans. Fuzzy Syst.; 2024; 32, pp. 2903-2914. [DOI: https://dx.doi.org/10.1109/TFUZZ.2024.3364382]

38. Tang, L.; Shushtari, M.; Arami, A. IMU-Based Real-Time Estimation of Gait Phase Using Multi-Resolution Neural Networks. Sensors; 2024; 24, 2390. [DOI: https://dx.doi.org/10.3390/s24082390]

39. Xu, S.; Dong, H.; Xu, R.; Meng, L.; Ming, D. A Real-Time Gait Phase Detection Method Based on BiLSTM-Attention Model. Proceedings of the 2023 45th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC); Sydney, Australia, 24–27 July 2023.

40. Kim, M.; Lee, D. Development of an IMU-based foot-ground contact detection (FGCD) algorithm. Ergonomics; 2017; 60, pp. 384-403. [DOI: https://dx.doi.org/10.1080/00140139.2016.1174314] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/27068742]

41. Caldas, R.; Mundt, M.; Potthast, W.; de Lima Neto, F.B.; Markert, B. A systematic review of gait analysis methods based on inertial sensors and adaptive algorithms. Gait Posture; 2017; 57, pp. 204-210. [DOI: https://dx.doi.org/10.1016/j.gaitpost.2017.06.019] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/28666178]

42. Liu, K.; Liu, Y.; Ji, S.; Gao, C.; Zhang, S.; Fu, J. A Novel Gait Phase Recognition Method Based on DPF-LSTM-CNN Using Wearable Inertial Sensors. Sensors; 2023; 23, 5905. [DOI: https://dx.doi.org/10.3390/s23135905] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/37447755]

43. Santurkar, S.; Tsipras, D.; Ilyas, A.; Madry, A. How Does Batch Normalization Help Optimization?. Advances in Neural Information Processing Systems; Curran Associates, Inc.: New York, NY, USA, 2018; Available online: https://proceedings.neurips.cc/paper/2018/hash/905056c1ac1dad141560467e0a99e1cf-Abstract.html (accessed on 30 July 2024).

44. Zhou, Q.; Shan, J.; Fang, B.; Zhang, S.; Sun, F.; Ding, W.; Wang, C.; Zhang, Q. Personal-specific gait recognition based on latent orthogonal feature space. Cogn. Comput. Syst.; 2021; 3, pp. 61-69. [DOI: https://dx.doi.org/10.1049/ccs2.12007]

45. Taborri, J.; Palermo, E.; Rossi, S.; Cappa, P. Gait Partitioning Methods: A Systematic Review. Sensors; 2016; 16, 66. [DOI: https://dx.doi.org/10.3390/s16010066] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/26751449]

46. Hearst, M.A.; Dumais, S.T.; Osuna, E.; Platt, J.; Scholkopf, B. Support vector machines. IEEE Intell. Syst. Their Appl.; 1998; 13, pp. 18-28. [DOI: https://dx.doi.org/10.1109/5254.708428]

47. Srivastava, N.; Hinton, G.; Krizhevsky, A.; Sutskever, I.; Salakhutdinov, R. Dropout: A Simple Way to Prevent Neural Networks from Overfitting. J. Mach. Learn. Res.; 2014; 15, pp. 1929-1958.

48. Wang, X.; Gu, Y. Classification of macular abnormalities using a lightweight CNN-SVM framework. Meas. Sci. Technol.; 2022; 33, 065702. [DOI: https://dx.doi.org/10.1088/1361-6501/ac5876]

49. Mirjalili, S.; Lewis, A. The Whale Optimization Algorithm. Adv. Eng. Softw.; 2016; 95, pp. 51-67. [DOI: https://dx.doi.org/10.1016/j.advengsoft.2016.01.008]

50. Heidari, A.A.; Mirjalili, S.; Faris, H.; Aljarah, I.; Mafarja, M.; Chen, H. Harris hawks optimization: Algorithm and applications. Future Gener. Comput. Syst.; 2019; 97, pp. 849-872. [DOI: https://dx.doi.org/10.1016/j.future.2019.02.028]

51. Shehab, M.; Mashal, I.; Momani, Z.; Shambour, M.K.Y.; Al-Badareen, A.; Al-Dabet, S.; Bataina, N.; Alsoud, A.R.; Abualigah, L. Harris Hawks Optimization Algorithm: Variants and Applications. Arch. Comput. Methods Eng.; 2022; 29, pp. 5579-5603. [DOI: https://dx.doi.org/10.1007/s11831-022-09780-1]

52. Issa, M.; Samn, A. Passive vehicle suspension system optimization using Harris Hawk Optimization algorithm. Math. Comput. Simul.; 2022; 191, pp. 328-345. [DOI: https://dx.doi.org/10.1016/j.matcom.2021.08.016]

53. Karaboga, D. Artificial bee colony algorithm. Scholarpedia; 2010; 5, 6915. [DOI: https://dx.doi.org/10.4249/scholarpedia.6915]

54. Tizhoosh, H.R. Opposition-Based Learning: A New Scheme for Machine Intelligence. Proceedings of the International Conference on Computational Intelligence for Modelling, Control and Automation and International Conference on Intelligent Agents, Web Technologies and Internet Commerce (CIMCA-IAWTIC’06); Vienna, Austria, 28–30 November 2005; pp. 695-701. [DOI: https://dx.doi.org/10.1109/CIMCA.2005.1631345]