Estimation of lower limb torque: a novel hybrid method based on continuous wavelet transform and deep learning approach

Shu Xu, Tao Wang, Zenghui Ding, Yu Wang, Tongsheng Wan, Dezhang Xu, Xianjun Yang, Ting Sun, Meng LiPeerJ Comput. Sci 2025 · DOI

TL;DR

This paper introduces a novel hybrid deep learning framework that combines Continuous Wavelet Transform (CWT) with a Multi-Head Self-Attention (MHSA) and Bidirectional Long Short-Term Memory (Bi-LSTM) network to accurately estimate lower limb joint torques from Inertial Measurement Unit (IMU) data. This approach offers a cost-effective, portable, and real-time solution for biomechanical analysis, outperforming existing methods in accuracy and computational efficiency.

What Problem Does This Paper Address?

Estimating lower limb joint torque is crucial for understanding biomechanics, preventing injuries, and guiding rehabilitation. Traditional methods like optical motion capture and force platforms are accurate but expensive, complex, and limited to laboratory settings. While Inertial Measurement Units (IMUs) offer a portable and cost-effective alternative, their data is inherently noisy and non-stationary, making precise feature extraction and torque estimation challenging. Existing deep learning models struggle with this complexity, often facing issues with computational overhead, real-time performance, and robustness in dynamic, real-world environments. This paper aims to bridge this gap by developing a robust, efficient, and accurate IMU-based system for lower limb joint torque estimation.

Methods

The proposed framework integrates several key components:

  • Continuous Wavelet Transform (CWT): This is the initial step, transforming raw IMU signals into time-frequency representations. CWT is chosen for its ability to effectively handle non-stationary signals by providing adaptive time-frequency resolution, smoothing high-frequency noise, and enhancing subtle low-frequency dynamic changes.
  • 1D Convolutional Residual Network (1D Conv ResNet): This module extracts spatial features from the CWT time-frequency images at various scales, delving deeper into the data. Residual connections are used to improve gradient flow and enable the training of deeper networks.
  • Multi-Head Self-Attention (MHSA): Integrated after the Conv ResNet, MHSA dynamically allocates attention weights to refine feature representations. It captures relationships between features across different time points and frequencies, focusing on salient regions and improving the model's understanding of complex spatiotemporal dynamics.
  • Bidirectional Long Short-Term Memory (Bi-LSTM): This module models temporal dependencies. By processing information in both forward and backward directions, Bi-LSTM effectively captures long-range temporal relationships inherent in gait, leading to more accurate and robust torque estimation.

The framework combines these modules in a multi-level fusion strategy to synthesize local and global information, enhancing prediction accuracy and robustness.

Datasets

The study utilized two publicly available datasets:

  • Dataset A: This dataset comprises multimodal biomechanical data from 22 healthy subjects performing six activity scenarios: flat ground walking, ramp ascent and descent, treadmill exercise, and stair ascent and descent. Data was collected using IMUs, a Vicon motion capture system, and a Bertec force plate.
  • Dataset B: This dataset serves as a validation set and includes data from 12 healthy subjects performing dynamic tasks like walking, running, and jumping. IMU sensors were placed on both lower limbs.

Key Insights and Ideas

  • CWT as a powerful pre-processing tool: The paper strongly advocates for CWT's ability to transform noisy, non-stationary IMU data into a richer, more informative time-frequency representation, which is crucial for subsequent deep learning models.
  • Synergy of MHSA and Bi-LSTM: The combination of attention mechanisms (MHSA) for capturing spatiotemporal relationships and recurrent networks (Bi-LSTM) for temporal modeling proves highly effective in modeling complex gait dynamics.
  • Single IMU advantage: The framework demonstrates that accurate lower limb torque estimation can be achieved using only IMU data, significantly reducing the complexity and cost associated with multi-sensor systems.
  • Real-time applicability: The model achieves a single-cycle inference time of 152.6 ms, making it suitable for real-time applications and portable biomechanical monitoring systems.

Key Results

  • High Accuracy: The proposed model achieved excellent performance across various metrics. On Dataset A, it reported an RMSE of 0.16 N·m/kg, an R² of 0.91, and a Pearson correlation coefficient (PCC) of 0.95. In specific scenarios like ramp climbing and treadmill walking, R² values reached up to 0.957 and 0.959, respectively.
  • Robustness: The model demonstrated high prediction accuracy even when subjected to simulated noise and sensor errors (R² > 0.96, PCC > 0.98).
  • Computational Efficiency: The model achieved a computational complexity of O(n · w · d), representing a 37.2% improvement over traditional LSTM models. The end-to-end inference time of 152.6 ms satisfies real-time processing requirements for 100 Hz sampled signals.
  • Superiority over Baselines: The proposed model significantly outperformed various baseline models (ANN, 1D CNN, LSTM, TCN, Time Transformer, TCN-LSTM) in the stair-climbing scenario, showing lower RMSE and higher R² values.

Limitations and Future Work

  • Controlled Environment Data: Data collection and validation were primarily conducted under controlled laboratory conditions. Real-world scenarios introduce greater variability and challenges that may affect generalizability.
  • Sensor Errors and Drift: IMU sensors are susceptible to noise, signal drift, and placement inaccuracies, which can impact accuracy.
  • Inter-individual Variability: The model may not adequately capture variations across different individuals, age groups, or pathological gait patterns.
  • Class Imbalance: The dataset exhibited an intrinsic class imbalance, which was not addressed in this iteration but is acknowledged as a future research direction.

Future work will focus on:

  • Expanding datasets to include diverse real-world scenarios and pathological gait data.
  • Implementing adaptive calibration strategies and advanced sensor error mitigation techniques.
  • Exploring lightweight model architectures and hardware acceleration for on-device deployment.
  • Investigating transfer learning and domain adaptation for improved generalizability.
  • Addressing class imbalance through data augmentation and class-weighted loss functions.
  • Developing more systematic approaches to hyperparameter optimization.

Summary updated

BibTeX
@article{xu2025estimation,
  author = {Shu Xu and Tao Wang and Zenghui Ding and Yu Wang and Tongsheng Wan and Dezhang Xu and Xianjun Yang and Ting Sun and Meng Li},
  title = {{Estimation of lower limb torque: a novel hybrid method based on continuous wavelet transform and deep learning approach}},
  journal = {PeerJ Comput. Sci.},
  year = {2025},
  doi = {10.7717/peerj-cs.2888},
}
Figure