Robust Human Gait Speed Recognition Under Non-Ideal Conditions for Suspension-Assisted Walking Systems.

Accurate gait speed estimation is vital for mobility assessment, rehabilitation monitoring, and intelligent assistive systems. However, existing methods often struggle under non-ideal conditions such as abrupt speed changes, asymmetric steps, or environmental disturbances. This study proposes a reliable framework for estimating gait speed, which integrates inertial measurement unit (IMU) signals with a hybrid deep learning model based on convolutional neural networks (CNNs) and bidirectional long short-term memory (BiLSTM) networks. Six IMUs were placed on bilateral thighs, calves, and ankles to capture comprehensive lower-limb motion data. After noise filtering, gait cycle segmentation, and Z-score normalisation, features were extracted across time, frequency, and time-frequency domains using Fast Fourier and Continuous Wavelet Transforms (FFT and CWT). A two-stage feature selection strategy combining Random Forest (RF) and Principal Component Analysis (PCA) reduced dimensionality while preserving discriminative power. The CNN-BiLSTM model was trained and tested on data from eight healthy subjects under variable-speed treadmill walking. Results showed superior performance, with an average R2 of 95.2% and a root mean square error (RMSE) of 0.0301 m/s, outperforming convolutional neural network (CNN), long short-term memory (LSTM), GRU, and backpropagation (BP) models. Sensor ablation confirmed the value of multi-site integration, and real-time tests validated online applicability. This work offers a reliable, high-precision solution for gait speed estimation under non-ideal conditions, with potential for wearable gait analysis and intelligent rehabilitation.