A multi-modal sensor data fusion processing method in a wireless body area network environment

Abstract

The application discloses a multi-modal sensor data fusion processing method in a wireless body area network environment, obtains surface electromyogram signals and inertial measurement unit data, extracts average absolute amplitude features and attitude deviation features to construct a continuous flow feature sequence after dynamic historical envelope preprocessing and up-sampling time sequence synchronization, monitors wireless body area network link quality and generates a dynamic confidence factor through cross-layer mapping, inputs the feature sequence and the confidence factor into a fuzzy reasoning engine, adaptively adjusts a dual-mode fusion weight based on an asymmetric dual mechanism of a cross-layer soft gate, and finally outputs a motion state quantitative index. The application can effectively suppress data distortion caused by packet loss, improve fusion robustness, and is suitable for resource-constrained edge gateway and microprocessor deployment due to lightweight calculation and low latency.

Description

Technical Field

[0001] This invention relates to a method for multimodal sensor data fusion processing in a wireless body area network environment. Background Technology

[0002] Wireless body area networks (WBANs) achieve continuous acquisition and transmission of multimodal physiological motion signals through sensor nodes deployed locally on the human body. In typical WBAN applications, surface electromyography (EMG) signal sensing nodes and inertial measurement unit (IMU) sensing nodes output digital signals of muscle electrical activity characteristics and limb posture motion characteristics, respectively. These signals are then aggregated into data packets via the WBAN transport layer and sent to the edge gateway or host computer. During transmission, the aforementioned multimodal sensor data often suffers from severe link damage and sudden, continuous packet loss due to human body occlusion and dynamic topology changes.

[0003] In existing technologies, the main approaches to address packet loss in wireless body area networks (WBNs) focus on low-level routing optimization or multi-hop quality of service (QoS) control. While these methods can alleviate packet loss to some extent, they lack a cross-layer coupling mechanism between the underlying physical signaling and the upper-layer data semantics, making it impossible to detect changes in link quality at the application layer and make adaptive adjustments. Regarding application-layer data recovery, existing methods often employ segmented cubic interpolation combined with boundary extrapolation mechanisms to fill in missing data. However, this interpolation process merely serves as a structural placeholder, and under conditions of severe, contiguous packet loss, it is highly susceptible to data distortion and cascading overshoot, leading to the failure of subsequent fusion processing logic.

[0004] In the fusion processing of multimodal sensor data, existing fusion strategies include graph network models, Gaussian filter optimization, and recurrent neural network methods. These strategies highly rely on the absolute continuity of time-series signals. When data packet loss causes time-series breaks, fixed-weight fusion or non-adaptive fusion mechanisms cannot dynamically adjust the participation level of each modal feature according to the real-time link status, resulting in fusion results that deviate significantly from the actual motion state. Furthermore, some existing methods use parameter quantization to construct lightweight recognition networks, but when faced with extreme packet loss, blind data interpolation at the application layer still cannot suppress fusion distortion caused by data loss.

[0005] In summary, existing technologies have not yet solved the problem of high-fidelity fusion processing of multimodal sensor data under the condition of sudden and continuous packet loss in wireless body area networks. There is a lack of a lightweight processing method that can map the underlying link quality status across layers to the application layer fusion decision and adaptively switch the fusion strategy when the link deteriorates. Summary of the Invention

[0006] The present invention provides a method for multimodal sensor data fusion processing in a wireless body area network environment in order to solve the problems existing in the prior art.

[0007] The technical solutions adopted in this invention are as follows:

[0008] A method for multimodal sensor data fusion processing in a wireless body area network environment includes the following steps:

[0009] S1: Acquire surface electromyography signals and inertial measurement unit data;

[0010] S2: Perform dynamic historical envelope preprocessing on the surface electromyography (SEMG) signal and upsampling timing synchronization processing on the inertial measurement unit (IMU) data; use a sliding window to extract the average absolute amplitude feature from the SEMG signal after dynamic historical envelope preprocessing and extract the attitude deviation feature from the IMU data after upsampling timing synchronization processing; construct a continuous flow feature sequence with SEMG signal feature channels and IMU feature channels through the average absolute amplitude feature and attitude deviation feature;

[0011] S3: Monitor link quality at the wireless body area network transport layer and convert the link quality into a dynamic confidence factor through a cross-layer mapping mechanism;

[0012] S4: Input the continuous flow feature sequence and the dynamic confidence factor into the fuzzy inference engine. The fuzzy inference engine adaptively adjusts the fusion weights under the asymmetric dual mechanism based on the dynamic confidence factor through cross-layer soft gating. The asymmetric dual mechanism performs dual-modal deep fusion when the link quality is good, and blocks the surface electromyography signal feature channel and falls back to the inertial measurement unit feature channel as a single-modal backup channel when the link quality deteriorates.

[0013] S5: The fuzzy inference engine calculates and outputs the motion state quantization index based on the adjusted fusion weights.

[0014] Furthermore, the digital signals output by the surface electromyography (SEMG) sensor node and the inertial measurement unit (IMU) sensor node are received through the wireless body area network (BNB) transmission layer; the sampling rate of the SEMG signal is 1000 Hz, and the sampling rate of the IMU data is 100 Hz.

[0015] Furthermore, the dynamic historical envelope preprocessing includes using a high-pass filter with a cutoff frequency of 20Hz to suppress motion artifacts, and using a low-pass filter with a cutoff frequency of 5Hz after full-wave rectification to extract the muscle force envelope.

[0016] Furthermore, the sliding window has a window length of 200ms and a step size of 100ms.

[0017] Furthermore, the upsampling timing synchronization processing uses cubic spline interpolation to upsample the inertial measurement unit data to 1000Hz; the attitude deviation feature is the absolute deviation of the pitch angle relative to the preset gait reference trajectory.

[0018] Furthermore, the cross-layer mapping mechanism includes: calculating the packet reception rate at the media access control layer using a 500ms sliding window, and mapping the packet reception rate to the dynamic confidence factor; the dynamic confidence factor Calculate using the following formula:

[0019] ,

[0020] in, For the span of the listening window, The indicator function takes a value of 1 when a data packet is successfully received and a value of 0 when a packet is lost. For the first The reception status of each sampling point.

[0021] Furthermore, the asymmetric dual mechanism sets three trigger thresholds. When the dynamic confidence factor is greater than 70, the link quality is determined to be good; when the dynamic confidence factor is less than or equal to 70 but greater than or equal to 30, the link quality is determined to be medium; and when the dynamic confidence factor is less than 30, the link quality is determined to be deteriorated. When the link quality is medium, the single-mode output of the inertial measurement unit feature channel is maintained. When the link quality is deteriorated, the surface electromyography signal feature channel is blocked and the output reverts to the inertial measurement unit feature channel as a single-mode backup channel.

[0022] Furthermore, the fuzzy inference engine adopts the Mamdani inference architecture, with the input variables being the average absolute amplitude feature, the attitude deviation feature, and the dynamic confidence factor, and the output variable being the motion state quantization index; the universe of discourse of both the input and output variables is mapped to the interval from 0 to 100.

[0023] Furthermore, the marginal regions of the universe of discourse employ a flat-topped trapezoidal membership function, while the central region of the universe of discourse employs a triangular membership function; the membership function of the dynamic confidence factor is:

[0024] ,

[0025] ,

[0026] ,

[0027] in, , and These are the membership values ​​of the corresponding fuzzy sets. For trapezoidal membership degree operator, This is the membership operator for triangles; the membership functions for the mean absolute magnitude feature and the attitude deviation feature are set similarly.

[0028] Furthermore, the fuzzy inference engine is deployed on an edge gateway or host computer, and the fuzzy inference engine uses the centroid method to calculate the motion state quantization index; the motion state quantization index is mapped to the range of 0 to 100 by global minimum and maximum value normalization.

[0029] The present invention has the following beneficial effects:

[0030] (1) By performing dynamic historical envelope preprocessing on surface electromyography signals and extracting the average absolute amplitude feature using a sliding window, combined with upsampling time-series synchronization processing of inertial measurement unit data and attitude deviation feature extraction, a continuous flow feature sequence is constructed. This continuous flow processing method reduces the dependence on segmentation of specific motion cycles, and the feature extraction process is mainly based on basic arithmetic operations, reducing the computation and storage overhead at the edge.

[0031] (2) By monitoring the link quality at the transmission layer of the wireless body area network and converting the link quality into a dynamic confidence factor through a cross-layer mapping mechanism, the application layer can perceive the changes in the link status before data distortion occurs, providing a basis for the adaptive adjustment of subsequent fusion weights.

[0032] (3) By employing a cross-layer soft gating and asymmetric dual mechanism based on dynamic confidence factors, dual-modal deep fusion is performed when the link quality is good. When the link quality deteriorates, the surface electromyography signal feature channel is blocked and the signal reverts to the inertial measurement unit feature channel as a single-modal backup channel. This mechanism effectively suppresses interpolation artifacts and cascade overshoot caused by missing data. Even under the condition of damaged wireless body area network links, it can still maintain the consistency between the fusion processing results and the actual motion state, and its robustness is higher than that of the fixed-weight fusion method.

[0033] (4) By using a fuzzy inference engine for fusion weight calculation, the processing latency meets the real-time requirements and is suitable for deployment on resource-constrained edge gateways or host computers. In addition, the rule base of the fuzzy inference engine has the technical prerequisite of being converted into a static lookup table, which can avoid complex online matrix calculations and further reduce storage and computing pressure. Attached Figure Description

[0034] Figure 1 This is a logic block diagram of the present invention.

[0035] Figure 2 This is a schematic diagram of the continuous flow feature extraction mechanism.

[0036] Figure 3 This is a time sequence diagram of the cross-layer mapping of the dynamic confidence factor under packet loss channel conditions.

[0037] Figure 4 A three-dimensional decision manifold graph based on a dual-mechanism fuzzy inference engine with cross-layer soft gating.

[0038] Figure 5 This is a performance comparison chart of various algorithms under different packet loss rates. Figure 5 In the middle (a), the RMSE (Root Mean Square Error) comparison curves are shown. Figure 5 (b) shows the Pearson correlation coefficient (PCC) comparison curve.

[0039] Figure 6 A time-domain tracking comparison chart showing the quantitative evaluation results of motion state under a 25% continuous packet loss rate. Detailed Implementation

[0040] The invention will now be further described with reference to the accompanying drawings.

[0041] like Figure 1 As shown, this invention discloses a multimodal sensor data fusion processing method in a wireless body area network (WBN) environment. The main execution component of this method is a data processing terminal deployed on an edge gateway or host computer. It communicates with front-end sensor nodes through the WBN transport layer, forming a complete hardware-software collaborative processing closed loop. The WBN transport layer receives digital signals output from surface electromyography (SEMG) sensors and inertial measurement unit (IMU) sensors. The SEMG signal has a sampling rate of 1000Hz, while the IMU data has a sampling rate of 100Hz. This acquisition of SEMG and IMU data provides the basic data source for subsequent processing. To address the sampling rate mismatch between the two heterogeneous signals, cubic spline interpolation is used to upsample the IMU data to 1000Hz, completing the upsampling timing synchronization process. This ensures complete alignment of the reference timing of the two signals, eliminating timing deviations caused by sampling rate differences.

[0042] like Figure 2 As shown, dynamic historical envelope preprocessing is performed on the acquired surface electromyography (EMG) signals. First, baseline drift is removed, and a fourth-order Butterworth high-pass filter with a cutoff frequency of 20Hz is used to suppress motion artifacts. After full-wave rectification, a fourth-order low-pass filter with a cutoff frequency of 5Hz is used to extract the muscle exertion envelope. The envelope amplitude expression of the surface EMG signal is as follows:

[0043] ,

[0044] in, and These are the low-pass and high-pass filter functions, respectively. This is the DC bias for the signal channel.

[0045] After preprocessing and timing synchronization, a sliding window is used to extract features from the processed signal. The sliding window has a length of 200ms and a step size of 100ms. For the first... A sliding window is used to extract the average absolute amplitude feature of the surface electromyography signal within the window. The calculation formula is as follows:

[0046] ,

[0047] in, The number of sampling points for the sliding window; Index of sampling points within the window.

[0048] The arithmetic mean of the attitude data of the inertial measurement unit within the same sliding window is obtained simultaneously. The absolute deviation between the attitude data and the preset gait reference trajectory is then calculated, and the attitude deviation feature is extracted. The attitude deviation feature is the absolute deviation of the pitch angle relative to the preset gait reference trajectory. The preset gait reference trajectory is derived from the normal gait mean of the subjects in the publicly available motion dataset, and serves as the physical reference for calculating the pitch angle attitude deviation.

[0049] By using the average absolute amplitude feature and attitude deviation feature, a continuous flow feature sequence with surface electromyography signal feature channel and inertial measurement unit feature channel is constructed. All extracted multimodal features are normalized and mapped to the [0,100] interval to provide standardized input for subsequent fusion inference.

[0050] In practical wireless body area networks, shadowing fading and continuous packet loss are often caused by the activities of subjects. Based on the IEEE 802.15.6 standard, a physical layer link impairment model is constructed under the condition of fixed pseudo-random seed, and log-normal shadowing fading is introduced. Its dynamic noise power expression is as follows:

[0051]

[0052] in, The base noise power is derived from the global variance of the original signal; Let be a normally distributed random variable.

[0053] At the application layer receiver, missing data is filled by segmented cubic interpolation and boundary extrapolation. Then, a third-order Savitzky-Golay filter with a frame length of 11 is used for initial smoothing. The interpolation and smoothing at this stage only serve as structural placeholders, aiming to maintain the temporal alignment of the continuous stream feature sequence.

[0054] like Figure 3 As shown, link quality is monitored at the transport layer of the wireless body area network. A cross-layer mapping mechanism is constructed by extracting physical layer link quality indicators. At the media access control layer, packet reception rate is statistically analyzed using a 500ms sliding window, corresponding to 50 data packets and 5 time steps. The cross-layer mapping mechanism converts the packet reception rate obtained from link quality monitoring into a dynamic confidence factor. The dynamic confidence factor at the current time t is... Calculate using the following formula:

[0055] ,

[0056] in, For the span of the listening window; This is an indicator function; it takes a value of 1 if the data packet is successfully received, and a value of 0 if the packet is lost. For the first The reception status of each sampling point.

[0057] When the underlying link deteriorates, the dynamic confidence factor will drop rapidly before the application layer distortion occurs. Based on this cross-layer perception, while the underlying layer passively performs interpolation preprocessing, the top-level fuzzy inference engine drives cross-layer soft gating with the dynamic confidence factor to adaptively block the interference of interpolation artifacts.

[0058] The continuous flow feature sequence and dynamic confidence factor are input into the fuzzy inference engine. The fuzzy inference engine adaptively adjusts the fusion weights under an asymmetric dual mechanism based on the dynamic confidence factor through cross-layer soft gating.

[0059] The asymmetric dual-mechanism system sets three trigger thresholds. When the dynamic confidence factor is greater than 70, the link quality is considered good, and both the surface electromyography (SEMG) signal feature channel and the inertial measurement unit (IMU) feature channel are simultaneously enabled to perform dual-modal deep fusion. When the dynamic confidence factor is less than or equal to 70 but greater than or equal to 30, the link quality is considered moderate, and the IMU feature channel is maintained in single-modal output. When the dynamic confidence factor is less than 30, the link quality is considered degraded, and the SEMG signal feature channel is blocked, reverting to the IMU feature channel as a single-modal backup channel to avoid feature distortion caused by link degradation interfering with the fusion results.

[0060] like Figure 4 As shown, the fuzzy inference engine adopts the Mamdani inference architecture. The input variables are the average absolute amplitude feature, the attitude deviation feature and the dynamic confidence factor, and the output variable is the motion state quantization index. The universe of discourse of both the input and output variables is mapped to the interval from 0 to 100.

[0061] To balance sensitivity to characteristic fluctuations with numerical robustness under extreme conditions, a flat-topped trapezoidal membership function is used in the edge region of the universe of discourse, while a triangular membership function is used in the central region. The membership function expression for the dynamic confidence factor is as follows:

[0062] ,

[0063] ,

[0064] ,

[0065] in, , and These are the membership values ​​of the corresponding fuzzy sets; and These are trapezoidal and triangular membership operators, respectively. The membership functions for the mean absolute amplitude feature and the attitude deviation feature are set similarly. The mean absolute amplitude feature is fuzzified into three fuzzy sets: low, normal, and high. The attitude deviation feature is defined as two fuzzy sets: abnormal and normal. The output motion state quantization index is mapped to three fuzzy sets: poor, medium, and excellent.

[0066] At the collaborative decision-making level, an asymmetric dual-mechanism rule base is constructed, and the minimum operator is used for multivariate pre-submission set computation. The activation strength calculation formula for the i-th rule is as follows:

[0067] ,

[0068] in, , and These represent the membership degrees of the corresponding input variables in a specific fuzzy set. For lower-level backoff rules that do not include surface electromyography (EMG) signal features, this operation is automatically reduced to a two-dimensional intersection operation involving only posture deviation features and dynamic confidence factors. The fuzzy inference engine uses the centroid method to calculate the motion state quantization index. The motion state quantization index is normalized to the range of 0 to 100 through global minimum and maximum normalization before being output, ultimately yielding a quantified result reflecting the motion state.

[0069] To verify the effectiveness and robustness of this method in a wireless body area network packet loss environment, a generalization verification experiment was conducted based on the MovePort public motion dataset. The experiment included 25 subjects to verify the cross-population generalization ability of the method. The core parameters and channel simulation settings of the experiment are shown in Table 1 below.

[0070] Table 1. Core parameter settings for the evaluation method

[0071]

[0072] The experiment simulated a complex wireless transmission environment using a log-normal shadowing fading model with a baseline signal-to-noise ratio of 15 dB and a standard deviation of 4 dB. Packet loss rates of 10%, 15%, and 25% were selected as performance degradation gradients, with 25% burst packet loss approaching the critical state of link interruption. The root mean square error (RMSE) was used to measure the degree of local signal distortion, the mean absolute error (MAE) to reflect the global deviation of the evaluation results, and the Pearson correlation coefficient (PCC) to evaluate the tracking fidelity of motion trends. The ideal evaluation manifold output under completely lossless physical channel conditions was defined as the true baseline, serving as the absolute benchmark for measuring the distortion resistance of each algorithm. To fully verify the performance advantages of the proposed method, the experiment introduced classic RMSE static fusion, the signal quality index (SQI) adaptive fusion mechanism, the gradient boosting tree model (LSBoost), and the lightweight one-dimensional convolutional neural network (1D-CNN) as comparative algorithms. The performance indicators of each method under different packet loss gradients are shown in Table 2 below.

[0073] Table 2 Comparison of evaluation errors and correlations of various algorithms under different packet loss rates.

[0074]

[0075] Under full gradient packet loss conditions, our method exhibits good evaluation accuracy. Taking the scenario with a 25% extreme packet loss rate as an example, its RMSE and MAE reach 11.30 and 4.96 respectively, both of which are better than the comparison algorithms. Compared with the second best performing LSBoost model, our method reduces the evaluation error RMSE by about 10.8% and MAE by 33.3%, demonstrating strong anti-perturbation characteristics.

[0076] like Figure 5 As shown, the evolution trends of core indicators under different packet loss rates are compared. Traditional application layer fusion algorithms have obvious limitations in robustness when facing sudden and continuous packet loss. Conventional interpolation is prone to irreversible cascading overshoot in data missing segments. The RMSE of traditional RMS fusion reaches 18.58 at the 10% degradation node. Even with the introduction of dynamic SQI mechanism, the high-frequency variance mutation caused by interpolation artifacts will mislead the weight allocation logic, and its RMSE rises abnormally to 23.97, making it difficult to effectively maintain the ability to track the real motion trend.

[0077] When channel degradation occurs, data-driven models such as LSBoost and 1D-CNN exhibit a certain degree of systematic uncertainty. As the packet loss rate gradually increases to 25%, the MAE of both models fluctuates significantly, rising to 7.44 and 10.72 respectively. The MAE of 1D-CNN is higher than 10.16 at all degradation gradients, exhibiting saturation oscillations. In scenarios where feature loss and signal distortion are intertwined, it is difficult to achieve stable evaluation output. In contrast, the performance evolution curve of our proposed method is the most gradual. Even under severe packet loss, the PCC remains robustly maintained around 0.80. Paired t-tests on 24 control subjects show that the evaluation error of our proposed method is significantly lower than that of the comparative model (P<0.001), effectively alleviating the evaluation failure problem caused by channel disconnection.

[0078] like Figure 6 As shown, the temporal tracking of the continuous sliding window evaluation trajectory of subjects with abnormal motion states under a 25% packet loss rate is performed. When the time series progresses to the 15 and 37 neighborhoods of the sliding window, the traditional fusion algorithm suffers from significant distorted overshoot due to the distorted interpolation characteristics. In contrast, this method utilizes a low-level link early warning mechanism, where the dynamic confidence factor rapidly decays before cascading overshoot, triggering adaptive compensation rules. By reducing dependence on damaged signals and robust feature tracking, it achieves highly robust motion state evaluation. To quantify the algorithm's time overhead, a single-step inference time test was conducted. The results show that the average time for a single cross-layer evaluation is only 8.42ms, below the real-time threshold, confirming that the asymmetric lightweight rule base can be used for real-time deployment on resource-constrained microcontrollers.

[0079] To verify the contribution of each core module to the method's performance, an ablation experiment was conducted on data with 25% severe packet loss. The performance changes were observed by removing key modules of the method step by step, and the results are shown in Table 3 below.

[0080] Table 3 Comparison of ablation experimental results of algorithm modules

[0081]

[0082] After removing the FIS module, the RMSE increased from 11.30 to 19.55, verifying that this module is the core support for the algorithm to adapt to modal anomalies and achieve nonlinear adaptive control. If the DHE module is removed, the system's resistance to baseline drift is significantly weakened, and the PCC index drops from 0.80 to 0.74, fully demonstrating the advantage of high-frequency continuous envelopes in capturing microscopic motion trends. Furthermore, the absence of the cross-layer gating mechanism, a key hub for adaptively masking interpolation artifacts within the algorithm, leads to a decrease in the system's dynamic fault-tolerant compensation capability, causing the RMSE to rise to 11.97. The microscopic tracking of DHE continuous flow characteristics, the nonlinear compensation of the dual-mechanism manifold structure, and the distortion masking of cross-layer soft gating together construct a complete perception-decision closed loop, fundamentally ensuring the robustness of this method under extremely damaged channels.

[0083] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements without departing from the principle of the present invention, and these improvements should also be considered within the scope of protection of the present invention.

Claims

1. A method for multimodal sensor data fusion processing in a wireless body area network environment, characterized in that: Includes the following steps: S1: Acquire surface electromyography signals and inertial measurement unit data; S2: Perform dynamic historical envelope preprocessing on the surface electromyography signal and upsampling timing synchronization processing on the inertial measurement unit data; A sliding window is used to extract the mean absolute amplitude feature from the surface electromyography (SEM) signal after dynamic history envelope preprocessing, and the attitude deviation feature is extracted from the inertial measurement unit (IMU) data after upsampling and timing synchronization processing. A continuous flow feature sequence with surface electromyography signal feature channel and IMU feature channel is constructed by the mean absolute amplitude feature and attitude deviation feature. S3: Monitor link quality at the wireless body area network transport layer and convert the link quality into a dynamic confidence factor through a cross-layer mapping mechanism; S4: Input the continuous flow feature sequence and the dynamic confidence factor into the fuzzy inference engine. The fuzzy inference engine adaptively adjusts the fusion weights under the asymmetric dual mechanism based on the dynamic confidence factor through cross-layer soft gating. The asymmetric dual mechanism performs dual-modal deep fusion when the link quality is good, and blocks the surface electromyography signal feature channel and falls back to the inertial measurement unit feature channel as a single-modal backup channel when the link quality deteriorates. S5: The fuzzy inference engine calculates and outputs the motion state quantization index based on the adjusted fusion weights.

2. The multimodal sensor data fusion processing method in a wireless body area network environment as described in claim 1, characterized in that: The wireless body area network transmission layer receives the digital signals output by the surface electromyography (SEMG) sensor node and the inertial measurement unit (IMU) sensor node; the sampling rate of the SEMG signal is 1000 Hz, and the sampling rate of the IMU data is 100 Hz.

3. The multimodal sensor data fusion processing method in a wireless body area network environment as described in claim 1, characterized in that: The dynamic historical envelope preprocessing includes using a high-pass filter with a cutoff frequency of 20Hz to suppress motion artifacts, and then using a low-pass filter with a cutoff frequency of 5Hz after full-wave rectification to extract the muscle force envelope.

4. The multimodal sensor data fusion processing method in a wireless body area network environment as described in claim 1, characterized in that: The sliding window has a window length of 200ms and a step size of 100ms.

5. The multimodal sensor data fusion processing method in a wireless body area network environment as described in claim 1, characterized in that: The upsampling timing synchronization processing uses cubic spline interpolation to upsample the inertial measurement unit data to 1000Hz; the attitude deviation feature is the absolute deviation of the pitch angle relative to the preset gait reference trajectory.

6. The multimodal sensor data fusion processing method in a wireless body area network environment as described in claim 1, characterized in that: The cross-layer mapping mechanism includes: calculating the packet reception rate at the media access control layer using a 500ms sliding window, and mapping the packet reception rate to the dynamic confidence factor; the dynamic confidence factor Calculate using the following formula: ， in, For the span of the listening window, The indicator function takes a value of 1 when a data packet is successfully received and a value of 0 when a packet is lost. For the first The reception status of each sampling point.

7. The multimodal sensor data fusion processing method in a wireless body area network environment as described in claim 1, characterized in that: The asymmetric dual mechanism sets three trigger thresholds. When the dynamic confidence factor is greater than 70, the link quality is determined to be good. When the dynamic confidence factor is less than or equal to 70 but greater than or equal to 30, the link quality is determined to be medium. When the dynamic confidence factor is less than 30, the link quality is determined to be deteriorated. When the link quality is medium, the single-mode output of the inertial measurement unit feature channel is maintained. When the link quality is deteriorated, the surface electromyography signal feature channel is blocked and the output reverts to the inertial measurement unit feature channel as a single-mode backup channel.

8. The multimodal sensor data fusion processing method in a wireless body area network environment as described in claim 1, characterized in that: The fuzzy inference engine adopts the Mamdani inference architecture. The input variables are the average absolute amplitude feature, the attitude deviation feature, and the dynamic confidence factor, and the output variable is the motion state quantization index. The universe of discourse of both the input and output variables is mapped to the interval from 0 to 100.

9. The multimodal sensor data fusion processing method in a wireless body area network environment as described in claim 8, characterized in that: The marginal regions of the universe of discourse employ a flat-topped trapezoidal membership function, while the central region of the universe of discourse employs a triangular membership function; the membership function of the dynamic confidence factor is: ， ， ， in, , and These are the membership values ​​of the corresponding fuzzy sets. For trapezoidal membership degree operator, This is the membership operator for triangles; the membership functions for the mean absolute magnitude feature and the attitude deviation feature are set similarly.

10. The multimodal sensor data fusion processing method in a wireless body area network environment as described in claim 1, characterized in that: The fuzzy inference engine is deployed on an edge gateway or host computer. The fuzzy inference engine uses the centroid method to calculate the motion state quantization index. The motion state quantization index is mapped to the range of 0 to 100 by global minimum and maximum value normalization.

Images