The latest data from the China Disabled Persons’ Federation (CDPF) shows that by the end of 2023, 12,463 rehabilitation organizations had been established across the country. The growth in this number not only reflects China’s investment in and emphasis on rehabilitation for people with disabilities, but also demonstrates the continuous expansion and popularization of rehabilitation healthcare resources across the country. With the establishment of these rehabilitation institutions, the application of rehabilitation aids is also spreading rapidly, while rehabilitation robots and human-computer interaction technology are gradually becoming key areas of academic research, promoting the continuous development of rehabilitation medicine and rehabilitation technology.

Currently, researchers at home and abroad have made significant progress by using different algorithms to realize the prediction of joint angles through surface electromyographic signals (sEMG). Utilizing the properties of surface EMG signals for fast and accurate real-time discrimination is one of the important foundations of the lower limb movement rehabilitation control system. By analyzing the surface EMG signals, researchers can extract key information reflecting the activity state of the muscles, thus predicting the movement angle of the joints.

Riccardo Fratti et al. proposed a multiscale convolutional neural network (MSCNN) that was pretrained with various strategies to improve inter-subject generalization [1]. Zhou et al. proposed a fusion recognition model that integrates a surface electromyography (sEMG)-based convolutional neural network, a transformer encoder, and a long and short-term memory network to perform four different gait movements [2]. Zhang et al. proposed Bidirectional Long Short-Term(BILSTM), a machine learning based human movement intention recognition algorithm to recognize actions such as falling, walking and turning [3]. Khorram et al. proposed a novel deep neural network (DNN) method based on smoothed, filled outliers, detrended and normalized (SFDN) dataset for classification of SEMG signals in hand movement recognition [4]. Su et al. proposed using a time delay artificial neural network to recognize human movement intention to predict human joint force of ankle eversion and inversion based on sEMG and angular velocity signals [5]. Liu et al. proposed a CNN-LSTMT (CLT) model based on surface electromyographic signals (sEMG) to solve the problems of limited classification and low recognition accuracy in existing muscle fatigue state recognition methods [6]. Charikleia Angelidou et al proposes a subject-specific pattern recognition (PR) and classification strategy using kinematic data and surface electromyographic (EMG) signals to recognize user intent to transition from a rigid to a compliant surface. Using a K-Nearest Neighbors (K-NN) methodology in combination with an Artificial Neural Network (ANN) [7]. Yuewan of Shanghai Jiao-Tong University predicted surface electromyography signals, but the LSTM neural network utilized was difficult to process long sequences of data and the prediction and identification of single signals was slightly poorer [8]; Yudong et al. Prediction by a single joint angle signal, the overall has some data processing difficulties, and the prediction effect has some error [9]; Relatively good motor intention prediction by both surface EMG signals and muscle strength by Jinming [10]; In order to solve the problem of inaccurate gait recognition due to the reliance on a single signal source, this study plans to collect electromyographic signals from the surface of the three major muscles of the lower limb, and to measure the angle of the knee and ankle joints simultaneously. Based on these data, a multi-source information flow limiting sliding time window algorithm fusing surface EMG signals with lower limb joint angles is proposed. The fused signals are processed by current limiting, filtering and feature ex-traction, and these data are subsequently used as inputs to the BPGN for accurate recognition of lower limb joint movement intentions. The relevant thought processes are detailed in Fig. 1.

FIGURE 1.

Flow chart of lower limb motion intention recognition based on multi-source information fusion.

Show All

The traditional sliding time window method does not delineate the start and end points of a fixed time window; the sliding time window ends at the time node reached by each request, and the starting point is the width of the time window one time window forward from the end point. Sliding time window, can prevent the traffic surge in any time window, but in some time windows are prone to a large number of duplicate data, so this paper proposes a sliding time window algorithm to limit the flow, as shown in Fig. 2 below.

FIGURE 2.

Current limiting sliding time window.

Show All

The sub-window in the figure is a quarter of the sliding time window (the width of the sliding window is an integer multiple of the width of the sub-window). Fig. 2 shows that the request arrives with statistics of the data traffic of the sub-window at that point in time, and after obtaining the data of the sub-window, the summation is performed, and it passes if the threshold is not exceeded; on the contrary, the flow is limited.

The formula is expressed as:

View Source\begin{equation*} \textrm {Q}=\int \limits _{t_{i}}^{t_{i} +\Delta t} {\textrm {x}\left ({{\textrm {t}}}\right )^{2}} \textrm {dt} \tag {1}\end{equation*}

If $\Delta t$ is too large, the range of the time window becomes consequently larger, leading to a lack of precision in extracting the start and end points of the action for that time phase and if the range is too small it leads to too little data being collected and the data being easily interfered with. Therefor the width of the sliding time window can be determined based on the sampling frequency of the signal.

When multiple source signals are used as input signals at the same time, the width of the sliding time window can be determined according to the sampling frequency and the number of samples to ensure that the synchronized data of multiple source signals can be acquired at the same time within the same time window.

The formula is expressed as:

View Source\begin{equation*} \Delta \textrm {t}_{\textrm {i}} =\frac {S_{i}}{f_{i} } \tag {2}\end{equation*}

In this paper, five healthy volunteers without limb injury and sports-related diseases were selected for the experiment, and the relevant personal data of the volunteers are shown in Table 1, with reference to the WHO’s BMI classification standard ($18.5\le$ BMI$\le 23$ is the normal range) the volunteers were all normal stature. In the table, A1 and A2 are two 25-year-old women, and A3, A4 and A5 are three 25-year-old men. All five volunteers signed an informed consent form before the experiment and understood the experimental process, which would not cause any physical injury to the volunteers, and all volunteers were aware of all the contents of the experiment and the purpose of the experimental study.

TABLE 1 Volunteer Data

During the experiment, a treadmill was utilized to achieve a uniform walking motion for the volunteers, and the treadmill was set at a speed of 4 km/h. When the walking maneuver occurred, the surface electromyographic signals of the three muscles of the leg, including the tibialis anterior, gastrocnemius, and rectus femoris, changed significantly. Therefore, surface EMG signal acquisition device after-market comparison, choose Jiangsu Wuxi Porunim Technology Co., Ltd products - EMG_06. EMG_06 is a six-channel muscle-electricity sensor, which contains the front end of the acquisition of analog circuitry and the back end of the digital filtering circuitry. The six-lead EMG module is equipped with STM32 development kit to collect the output signal from the middle end. The six-lead EMG sensor composition and electrode pads are shown in Fig. 3 below. The equipment is shown in Fig. 3, where ① is the upper computer acquisition and display interface, ② is the hardware circuit acquisition system, and ③ is the non-invasive acquisition signal dressing.

FIGURE 3.

Multi-source signal acquisition system construction.

Show All

Meanwhile, two nine-axis IMUs with a frequency of 200 Hz were used as angle sensors, the sensors are shown in Fig. 3, ④, to collect the angle information of the knee and ankle joints, respectively. Transmission of the collected data to the computer via Bluetooth reduces the signal interference of the line to the collection process. Before the experiment, 75% medical alcohol is used to clean the skin and thus reduce the impedance, and during the experiment, attention is paid to the simple environment of the acquisition location to avoid the interference of other equipment and lines, and the multi-source signal acquisition system is shown in Fig. 3 below. The test acquisition results are shown in Fig. 4, 5, 6, 7, and 8.

FIGURE 4.

Primary electromyographic signal of tibialis anterior muscle.

Show All

FIGURE 5.

Primary electromyographic signals of the gastrocnemius muscle.

Show All

FIGURE 6.

Rectus femoris primitive electromyographic signal.

Show All

FIGURE 7.

Measured knee angle.

Show All

FIGURE 8.

Measured ankle angle.

Show All

The sEMG has an effective frequency range of 20-200 Hz and is subject to low frequency of 0-20 Hz as well as industrial frequency interference of 50 Hz during the acquisition process. The text selects a band-pass filter to filter out the low-frequency interference in the experiment, and an Infinite Impulse Response (IIR) digital trap is used to filter out the work-frequency interference during the experiment.

Although the initial filtering has removed some of the noise, the amplitude of the signal still exhibits large random fluctuations. To further improve the stability and availability of the signal using full-wave rectification, full-wave rectification can effectively smooth the signal waveform and reduce the amplitude of the random fluctuations. This provides a more reliable input base for subsequent data analysis and model training.

View Source\begin{equation*} \textrm {sEMG}_{\textrm {rac}} \left ({{ n }}\right )=\left |{{\textrm {sEMG}\left ({{ n }}\right )}}\right | \tag {3}\end{equation*}

$$\begin{equation*} \left |{{H\left ({{ \omega }}\right )}}\right |^{2}=\frac {1}{1+\left ({{\omega /\omega _{c}}}\right )^{2n}} \tag {4}\end{equation*}$$ View Source\begin{equation*} \left |{{H\left ({{ \omega }}\right )}}\right |^{2}=\frac {1}{1+\left ({{\omega /\omega _{c}}}\right )^{2n}} \tag {4}\end{equation*}

To solve the problem of large order of magnitude difference between the various dimensions of the signal, all the data is normalized to a number between [0, 1]. The mapminmax function is used in MATLAB for normalization.

View Source\begin{equation*} y\left ({{ i }}\right )=y_{\min } +\left ({{y_{\max } -y_{\min }}}\right )\ast \frac {x\left ({{ i }}\right )-x_{\min }}{x_{\max } -x_{\min } } \tag {5}\end{equation*}

In the above equation, $x(i)$ denotes the original signal sequence; $x_{max}$ and $x_{min}$ are the maximum and minimum values of $x(i)$ , respectively; $y(i)$ is the normalized signal sequence; $y_{max}$ and $y_{min}$ are the maximum and minimum values of $y(i)$ , respectively, and to ensure that the original signals are normalized to between [0, 1], $y_{max}=1$ , $y_{min}=0$ .

Based on the above processing, the processed surface EMG signal was obtained as shown in Fig. 9, 10, and 11.

FIGURE 9.

Preprocessed tibialis anterior muscle electromyographic signals.

Show All

FIGURE 10.

Preprocessed electromyographic signals of the gastrocnemius muscle.

Show All

FIGURE 11.

Rectus femoris electromyographic signals after preconditioning.

Show All

The BPGN algorithm combines the capabilities of Back Propagation Neural-networks (BPNN) and Generalized regression Neural-networks (GRNN). BP are known for their strong learning ability and optimization performance, and can continuously adjust the weights through the error back propagation algorithm to approximate complex nonlinear functions, but the training process may be slow to converge and easy to fall into the local optimum. On the other hand, the generalized regression neural network is a radial basis function-based network with fast convergence and good nonlinear mapping ability, and has high robustness and fault tolerance. After fusion, the BPGN algorithm makes full use of the advantages of the two networks to achieve more efficient nonlinear mapping and prediction ability. It also has higher robustness.

The window width affects the accuracy of the experiment, and in this paper the surface EMG signals were acquired at a frequency of 1000 Hz and 3000 sample points were selected. The joint angle signal was sampled at 200 Hz, and 600 sample points were selected to determine a window width of 3 seconds (approximately two gait cycles) for both signals. Data synchronization of the surface EMG signal and the joint angle signal can be ensured simultaneously.

In this paper, we use a three-layer Sigmoid-type neural network.

View Source\begin{equation*} f\left ({{Net_{i}}}\right )=\frac {1}{1+e^{-Net_{i} /\textrm {T}}} \tag {6}\end{equation*}

$$\begin{equation*} x_{j} =\sum {i\omega _{ij} x_{i}} \tag {7}\end{equation*}$$ View Source\begin{equation*} x_{j} =\sum {i\omega _{ij} x_{i}} \tag {7}\end{equation*}

$$\begin{equation*} y_{j} =\sum {j\omega _{jk} x_{_{j}}} \tag {8}\end{equation*}$$ View Source\begin{equation*} y_{j} =\sum {j\omega _{jk} x_{_{j}}} \tag {8}\end{equation*}

FIGURE 12.

Three-layer neural network structure model.

Show All

In neural networks, the number of hidden layer neurons has a significant impact on the prediction accuracy of the model. Therefore, determining a reasonable number of hidden layer neurons is the key to the successful establishment of a BP neural network prediction model. The choice of the number of nodes will directly affect the fitting effect, resulting in lower prediction accuracy. Currently, there is no uniform mathematical formula for how to select the number of neurons in the hidden layer of BP neural networks. This study refers to the following empirical formula in this issue:

View Source\begin{equation*} H=\sqrt {n+m} +a \tag {9}\end{equation*}

Learning rate decreases as the difference of the exponent grows with the number of training rounds when using neural networks for deep learning

View Source\begin{equation*} \alpha =0.95^{epoch\_num}\ast \alpha _{0} \tag {10}\end{equation*}

The learning rate plays a key role in the neural network training process. Learning rate is too small will lead to the number of iterations to increase the parameter convergence slowly, learning rate is too large will lead to the local optimum, or even cannot converge. To determine the suitable learning rate, 0.1, 0.2, 0.3, 0.4, 0.5 and 1 were selected for iterative simulation. The results of the recognition rates corresponding to the six groups of learning rates are 93.27%, 92.81%, 91.14%, 90.83%, 90.11% and 83.69% by experimental comparison. From the above results, when the learning rate is 0.1, the recognition rate results perform better, so a learning rate of 0.1 is chosen for neural network learning training. Through the above analysis, after constant adjustment of the weights, the learning rate of 0.1 allows the model to obtain proper robustness, which never makes the prediction accuracy increase.

Surface EMG signals are commonly used to assess the accuracy of regression model and neural network prediction to evaluate the accuracy of regression model and neural network prediction, the mean square error and the coefficient of determination are commonly used as the prediction accuracy evaluation indexes, and they can mutually verify the quality of the regression model when they exist at the same time.

Mean Square Error (MSE): the closer the result is to 0, the more accurate it represents its prediction.

View Source\begin{equation*} MSE=\frac {1}{n}\sum \limits _{i=1}^{n} {\left ({{y_{i} -\hat {y}_{i}}}\right )} ^{2} \tag {11}\end{equation*}

Coefficient of determination ($R^{2}$ ): a range of values (0-1), the closer the value is to 1, the more capable the regression model is represented.

View Source\begin{equation*} R^{2}=1-\frac {\sum \nolimits _{i=1}^{n} {\left ({{y_{i} -\hat {y}_{i}}}\right )^{2}}}{\sum \nolimits _{i=1}^{n} {\left ({{y_{i} -\bar {y}_{i}}}\right )^{2}} } \tag {12}\end{equation*}

In the above equation, $y_{i}$ is the true observation, $\bar {y_{i}}$ is the true observation mean, $\hat {y_{i}}$ is the predicted value, and n is the sample size.

In MATLAB, the number of neurons in the hidden layer is set to 11, and the learning rate of the training process is 0.1. 1000 samples are selected for each set of prediction data, in which two thirds of the data are used to train the neural network, and the other one third is used for prediction, and the prediction of knee and ankle angles is carried out with the volunteer A1 as an example, and three simulations are carried out for the data provided by each volunteer, namely, multivariate current-limiting-based sliding time window BP neural network generalized algorithm for single signal recognition, multi-source signal recognition based on multivariate restricted flow sliding time window BP neural network generalized algorithm and multi-source information recognition based on BP neural network.

As an example, a single signal is trained with a multi-source current limiting sliding time window BPNN algorithm, and the error convergence during training is shown in Fig. 13.

FIGURE 13.

MSE for single signal prediction.

Show All

The data shown in Table 2 below are the evaluation metrics of the five volunteers’ intent to recognize a single lower extremity joint angle signal via the lower extremity joint angle.

TABLE 2 Results of Evaluation Indexes for Identification of Joint Angle Signals via a Single Lower Limb Joint

As an example, the error convergence in training is shown in Fig. 15 for a multi-source signal for multi-source current limiting sliding time window BP neural network generalized algorithm training. Fig. 16 shows a comparison of the results of the BP neural network for motor intent recognition after fusing surface EMG signals and lower limb joint angle signals from multiple sources, and Fig. 17 shows a comparison of the results of the generalized algorithm of the BP neural network for motor intent recognition after fusing surface EMG signals and lower limb joint angle signals from multiple sources with a current-limited sliding time window.

FIGURE 14.

Comparison of recognition results of single signal of volunteer A1.

Show All

FIGURE 15.

MSE for multi-source signal prediction.

Show All

FIGURE 16.

BP neural network motion intention recognition after multi-source fusion of surface electromyography and lower limb joint angle signals for Volunteer A1.

Show All

FIGURE 17.

Multi-source current-limited sliding time window BP neural network generalization algorithm for motion intention after fusing surface electromyography and lower limb joint angle signals from multiple sources for Volunteer A1.

Show All

Table 3 shows the results of evaluation metrics based on BP neural network for lower limb movement intention recognition in five volunteers after fusing multiple signal sources.

TABLE 3 Evaluation Index Results of BP Neural Network Recognition by Multi-Source Signal Fusion

Table 4 shows the results of the evaluation indexes for the recognition of lower limb motor intention in five volunteers after fusion of signals from multiple sources.

TABLE 4 Evaluation Index Results of MLS+BPGN Neural Network Recognition by Multi-Source Signal Fusion

Fig. 18 shows the comparison of the RBF neural network motion intention recognition results after fusion of multi-source current-limited sliding time window surface EMG signals and lower limb joint angle signals

FIGURE 18.

Volunteer A Recognition of motor intention based on RBF neural network after multisource fusion of surface electromyography signals and lower limb joint angle signals.

Show All

Table 5 shows the results of the evaluation indexes for the recognition of lower limb motor intention in five volunteers after fusion of signals from multiple sources.

TABLE 5 Evaluation Index Results of MLS+ RBF Neural Network Recognition by Multi-Source Signal Fusion

The comparison of Table 3 and Table 4 shows that the recognition effect of multi-source current limiting sliding time window BP neural network generalized algorithm is better than the recognition effect of BP neural network; the comparison of Table 2 and Table 4shows that the accuracy of the recognition effect of a single signal is lower than that of the recognition effect of multi-source signal fusion; From the comparison of Tables 4 and 5, it can be seen that the recognition accuracy of the generalized algorithm of multi-source current limiting sliding time window BP neural network is better than that of RBF neural network.

Comparison results showed that the average error of knee joint angle recognition was reduced by 10.8° and the average error of ankle joint angle recognition was reduced by 7.2° for multi-source signal fusion compared to single signal. The multivariate flow-limited sliding time window BP neural network generalization algorithm reduced the average knee joint error by 13.6° and the average ankle joint error by 8.5° compared to BP neural network intent recognition. The multivariate flow-limited sliding time-window BP neural network generalization algorithm reduced the knee mean error by 11.2° and the ankle mean error by 7.4° compared to the RBF neural network intent recognition. From the experimental results, BPGN has faster recognition response speed and higher recognition accuracy than BP and RBF; the recognition results of multi-source signal fusion are also more accurate, and in summary, BPGN is suitable to be applied in the research of lower limb movement intention recognition.

Focusing on the challenges of human posture complexity and lower limb motion in-tent recognition, this study significantly improves the accuracy of motion intent recognition by innovatively introducing a multi-signal fusion approach. While traditional single-source approaches have limitations in capturing complex motion intentions, this study provides richer information inputs by integrating multi-source data, which enhances the predictive capability of the model.

By proposing a multi-source current limiting sliding time window algorithm, this study effectively solves the synchronization and flow control problems in multi-signal fusion. The algorithm optimizes data processing and transmission through the time window mechanism to ensure the consistency of temporal data and overcome the problem of temporal disorder in traditional data processing. The experimental results show that the multi-signal fusion method significantly improves the accuracy of motion intent recognition and verifies its effectiveness in the analysis of complex systems.

In terms of algorithms, this study used a multi-source current-limited sliding time window generalization algorithm combined with a BP neural network generalization algorithm for the recognition of gait motion patterns. The recognition accuracy is further improved by utilizing its nonlinear mapping capability. Combining multi-signal fusion and advanced neural network algorithms, this study successfully constructs an efficient motion intent recognition system, which significantly improves the recognition rate.

Overall, this study not only provides new ideas and methods for multi-signal processing theoretically, but also demonstrates its potential application in the field of motor intent recognition in practice. The multi-signal fusion method improves the accuracy and robustness of the model by integrating multi-source data, which provides a powerful technical support for sports science and rehabilitation medicine.

Although this study has made significant progress in multi-signal fusion and motion intent recognition, there are still many areas that deserve further exploration an optimization. First, future work will be devoted to optimizing the speed and accuracy of the algorithm. Although the multi-source current limiting sliding time window algorithm per-forms well in this study, the processing speed of the algorithm still needs to be further im-proved in real-time applications. By introducing more efficient algorithm optimization techniques and hardware acceleration methods, computational latency can be further reduced and real-time responsiveness can be improved.

Second, future research will explore the applicability of multi-signal fusion methods in more motion modes and application scenarios. For example, different types of motion (e.g., upper body motion, whole body motion) may require different signal fusion strategies and algorithm optimizations. In addition, the research will delve into the application of multi-signal fusion in rehabilitation medicine to provide more precise and personalized rehabilitation programs for patients with motor dysfunction through integration with biomechanics and neuroscience.

Finally, future work will focus on the scalability and robustness of multi-signal fusion techniques in real-world applications. The performance and potential of the method in real-world environments will be validated through collaboration with industry and healthcare organizations. Further developments also include the development of portable and wearable devices that will enable the multi-signal fusion technique to be widely used in everyday life and drive innovation in sports science and rehabilitation medicine.

In summary, the multi-signal fusion method proposed in this study provides a new technical path for motor intention recognition, which has important theoretical and practical significance. Future research and work will be oriented towards state analysis of rehabilitation training muscles based on surface EMG signals, obtaining lower limb muscle movement feedback information during patient characterization rehabilitation training through multi-source sensor fusion, and realizing intelligent quantitative assessment of rehabilitation outcomes by combining with traditional assessment scales. Fully mobilize the enthusiasm of patients to participate in rehabilitation training. Bringing more innovations and breakthroughs to the field of sports science and rehabilitation medicine.