Enhancing the Myoelectric Pattern Recognition Robustness to Electrode Shift by an Autoencoder-Based Feature Calibrator

Electrode shift is considered a pivotal and unavoidable factor affecting the robustness of myoelectric pattern recognition (MPR). Nevertheless, most existing solutions face challenges in simultaneously ensuring plug-and-play compatibility and achieving high classification accuracy. To address this challenge, this study presents a novel method for adaptively calibrating testing data samples using an autoencoder-based feature calibrator. A feature transformation approach, encompassing interpolation, translation, and down-sampling operations, is executed on the original feature map to generate the simulated feature map representing the shifted view. Subsequently, the flattened shifted view features serve as input for the autoencoder network, facilitating the model to learn a more resilient feature representation based on the reconstruction error between the output and the original features. The trained autoencoder is deployed as an independent feature calibrator in the training and testing process of the classifier. The performance of the proposed method was evaluated with data recorded by wearing an 8-channel armband on the forearm of five subjects performing six gestures. The proposed method achieved a high classification accuracy of 87.20±3.53% under the electrode shift condition, outperforming three commonly used comparison methods with statistical significance (p < 0.05). This study provides a practical solution for mitigating electrode shift interference without requiring any additional calibration data, which contributes to enhancing the robustness of MPR systems.