عنوان المشروع: An Edge-AI Enabled Multi-Sensor Fusion System for Intelligent Physical Rehabilitation Monitoring
الوصف:
An Edge-AI Enabled Multi-Sensor Fusion System for Intelligent Physical Rehabilitation Monitoring
PROJECT DESCRIPTION
INTRODUCTION
The increasing cases of musculoskeletal disorders, aging populations, and sedentary lifestyles have significantly increased the demand for effective physical rehabilitation monitoring systems. Rehabilitation exercises are commonly prescribed to patients recovering from injuries, neurological conditions, or surgeries to restore mobility, strength, and functional independence. Traditionally, physiotherapists supervise these exercises in clinical settings to ensure correct movements and prevent injuries. However, in-person supervision is limited by healthcare resource shortages, high costs, and geographical barriers. As a result, most patients perform rehabilitation exercises at home without supervision, leading to incorrect movements that may delay recovery or cause additional injuries.
Recent advances in wearable sensing technology, particularly Inertial Measurement Units (IMUs) combining accelerometers and gyroscopes, enable fine-grained motion data collection. Machine learning and deep learning can automatically identify exercises and physical activities by analyzing IMU signals. However, most existing solutions rely on cloud computing, where sensor data is transmitted to remote servers for analysis. This introduces latency, network dependency, and privacy concerns. Real-time feedback requires immediate processing, which cloud-based systems cannot guarantee.
Additionally, current research focuses mainly on simple activity recognition (walking, running, sitting) rather than complex rehabilitation exercises with subtle movement differences. Conventional machine learning methods struggle with complex human movements, while deep learning methods often have high computational demands unsuitable for edge devices.
To overcome these challenges, this paper presents an Edge-AI based multi-sensor fusion system for real-time intelligent physical rehabilitation and movement quality detection using wearable IMU sensors. The system is tested on the INSIGHT-LME IMU dataset containing squat, lunge, push-up, sit-to-stand, and arm raise exercises.
LITERATURE REVIEW
Recent developments in wearable sensing and artificial intelligence have advanced human activity recognition in healthcare and rehabilitation. Several studies have proposed transformer-based models, federated learning systems, and hybrid deep neural networks for activity recognition using IMU and EMG sensors. While these approaches have shown promise, most focus on general activity recognition rather than rehabilitation-specific movement quality assessment.
Multi-sensor fusion has been proposed to enhance recognition performance using complementary data sources. Studies have shown that integrating heterogeneous sensor signals improves decision-making compared to single-sensor approaches. However, many frameworks remain cloud-dependent, introducing latency issues that limit real-time feedback capabilities.
Edge computing has emerged as a solution for latency reduction, enabling real-time analytics directly on wearable devices. Some studies have demonstrated edge-enabled activity recognition using Wi-Fi signals or gait analysis. However, these systems either lack multi-sensor fusion, focus on limited movement types, or require specialized hardware.
Quality datasets are essential for effective activity recognition. While several datasets provide multi-sensor inertial measurements, they often do not address real-time inference and movement quality assessment. Recent work has explored lightweight frameworks for rehabilitation motion quality assessment, but these typically use single sensors rather than multi-sensor fusion.
The key gaps in current research include: emphasis on activity categorization rather than movement quality, use of single-sensor or computationally expensive models not optimized for edge deployment, and limited integration of real-time feedback mechanisms. This system addresses these gaps by creating an Edge-AI enabled multi-sensor fusion framework for real-time rehabilitation exercise identification and movement quality assessment.
PROPOSED SYSTEM
The proposed Edge-AI rehabilitation system incorporates wearable sensing, intelligent data processing, and adaptive feedback. The methodology converts raw biomechanical motion data into meaningful rehabilitation information through a systematic pipeline.
Motion data acquisition begins with wearable inertial sensing units, recording tri-axial linear acceleration and angular velocity from the INSIGHT-LME IMU dataset. Inertial sensing is portable, environmentally robust, and privacy-preserving compared to vision-based systems.
A preprocessing phase enhances signal reliability. Low-pass filtering eliminates high-frequency disturbances:
S_f(t) = S(t) * h(t) (1)
where S(t) is the original sensor signal and h(t) is the filtering response function.
Continuous data streams are split into fixed time slices using sliding window segmentation:
W_i = {x_t, x_{t+1}, ..., x_{t+n}} (2)
where n represents the window length.
Feature extraction translates sensor measurements into biomechanical descriptors. Signal normalization ensures consistent scale across sensor channels:
X_norm = (X - μ) / σ (3)
where μ and σ are the mean and standard deviation of the signal distribution.
Processed features are sent to the Edge-AI movement analysis module using a hybrid CNN-BiLSTM network. CNN layers learn spatial correlations among multi-axis sensor channels, while BiLSTM layers learn temporal evolution of motion sequences:
h_t = f(W_x x_t + W_h h_{t-1} + b) (4)
The trained network performs two functions: exercise classification and posture quality evaluation. Movement quality is measured as:
Q = 1 - (|M_pred - M_ref|) / M_ref (5)
where M_pred is predicted movement characteristics and M_ref is reference therapeutic motion.
The trained AI model is deployed on edge devices such as Raspberry Pi or ESP32. Inference latency is estimated as:
L = T_proc + T_trans (6)
where T_proc is local processing time and T_trans is transmission delay. Local inference significantly reduces T_trans, enabling real-time corrective feedback.
The system provides visual, audio, or haptic feedback based on movement assessment. Summarized rehabilitation data is relayed via MQTT-based IoT communication to a cloud analytics environment for long-term progress tracking and therapist visualization.
IMPLEMENTATION AND RESULTS
The proposed framework was evaluated using the INSIGHT-LME IMU dataset, containing motion signals for squat, lunge, push-up, sit-to-stand, and arm raise exercises. Raw signals were preprocessed with normalization, temporal segmentation, and noise filtering. The dataset was split into 80% training and 20% validation.
The model achieved training accuracy of 99.4% and validation accuracy of 96.7% at the final epoch, with minimal overfitting. This represents a 1.7–4.7% improvement over conventional activity recognition models.
The confusion matrix showed correct classifications: 148 squats, 146 lunges, 149 push-ups, 151 sit-to-stands, and 144 arm raises. Misclassifications were minimal, occurring primarily between movements with similar patterns.
The Receiver Operating Characteristic curves demonstrated high discriminative capability with AUC values of approximately 0.972, 0.978, 0.975, 0.981, and 0.976 for the five exercise classes.
Inference latency measurements ranged from 28 to 36 milliseconds, with an average of 32 milliseconds. This represents a 60% reduction compared to cloud-based processing.
Class-wise performance metrics: squat achieved 98% precision, recall, and F1-score; lunge and arm raise achieved approximately 97.3%; push-up achieved 96.7%; sit-to-stand achieved 96.2%.
Table 1: Quantitative Comparison of Exercise Classification Models
Model Accuracy (%) Precision (%) Recall (%) F1 Score (%) AUC
SVM 88.6 87.9 88.1 88.0 0.91
Random Forest 91.3 90.8 91.0 90.9 0.93
CNN 93.7 93.1 93.4 93.2 0.95
LSTM 94.8 94.2 94.5 94.3 0.96
CNN-LSTM Hybrid 95.6 95.0 95.2 95.1 0.97
Proposed Model 96.7 97.0 96.5 96.7 0.98
Edge device inference latency comparison: Smartphone (18 ms), Jetson Nano (28 ms), Raspberry Pi 4 (42 ms), ESP32 (65 ms). All devices achieved real-time performance suitable for wearable healthcare systems.
CONCLUSION
This paper presented an Edge-AI enabled multi-sensor fusion system for real-time intelligent physical rehabilitation monitoring using wearable inertial sensors. The system combines sensor data preprocessing, feature extraction, and deep learning-based activity recognition in an edge computing environment.
The proposed framework achieves 96.7% classification accuracy, 97.0% precision, 96.5% recall, 96.7% F1-score, and 0.98 AUC, outperforming conventional models by 1.1–8.1%. Inference latency ranges from 18–65 milliseconds across different edge devices, enabling real-time feedback.
The system supports both exercise classification and movement quality assessment, which is essential for ensuring patients perform rehabilitation exercises correctly in clinical and home-based settings. The combination of multi-sensor fusion and edge-based processing enhances activity recognition reliability while reducing dependence on cloud computing.
Future work includes integrating additional physiological sensors such as electromyography and pressure sensors, developing personalized patient-specific rehabilitation models, combining computer vision with wearable sensor data for hybrid sensing, and conducting extensive clinical trials to validate practical usefulness in real rehabilitation settings.
تاريخ رفع الملفات: 2026-04-02 15:48:54