EMG Gesture Recognition Model Adaptation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing hand gesture recognition systems using EMG sensors face challenges in maintaining classification accuracy across different users and sessions due to variations in muscle physiology and sensor placement.
Innovation Solution
A machine learning model with a data adaptation part, encoder, decoder, and classifier is trained using a 3-stage unsupervised domain adaptation method, which updates the data adaptation part based on comparisons between input and decoded sensor data items, without requiring labelled data for the target distribution.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a machine learning model is trained on sensor data from one user or session, then classification accuracy is high for that specific user/session, but classification accuracy deteriorates when applied to different users or sessions
Solution Approach 1:
The model is segmented into two distinct parts: a frozen pre-trained model that handles general feature extraction and classification, and an adaptive domain adaptation part that is trained separately for each new user or session. This segmentation allows the system to maintain high accuracy for the general case while adapting to specific users or sessions without retraining the entire model.
Solution Approach 2:
The model performs preliminary training on a large dataset from multiple users and sessions, freezing the parameters to capture general patterns. This preliminary action establishes a robust foundation that can be efficiently adapted to new users or sessions by training only the domain adaptation part, avoiding the need for complete model retraining.
2Measurement precision
If domain adaptation is performed for each new user or session, then classification accuracy is maintained, but training time and computational resources increase
Solution Approach 1:
The training process is segmented into a one-time preliminary training phase and subsequent per-user/session adaptation phases. The preliminary training establishes the base model, and only the domain adaptation part is trained for each new user or session, significantly reducing the computational burden compared to training the entire model from scratch for each user.
Solution Approach 2:
The model changes its parameters selectively by training only the domain adaptation part for each new user or session, while keeping the pre-trained encoder and classifier frozen. This parameter change approach allows the system to adapt to new users efficiently without the computational cost of retraining the entire model.
Data Source
Figure 1~5
AI summary
Embodiments relate to an apparatus comprising means configured for: - obtaining (T2) a plurality of sensor data items, wherein a sensor data item specifies a set of values determined from at least one sensor, - training (T3), based on the plurality of sensor data items, a machine learning model (10) configured for classifying sensor data items, wherein the machine learning model comprises: - a data adaptation part (11) configured for determining a modified sensor data item (X') based on an input sensor data item (X), - an encoder (12) configured for determining encoded features (F) based on the modified sensor data item (X'), - a decoder (13) configured for determining a decoded sensor data item (Y) based on the encoded features (F), representing an estimation of the input sensor data item (X), and - a classifier (14) configured for determining a class (C) associated with the input sensor data item (X), wherein training (T3) the machine learning model (10) comprises updating parameters of the data adaptation part (11) based on a comparison between an input sensor data item (X) from the plurality of sensor data items and the corresponding decoded sensor data item (Y), while maintaining the encoder, the decoder and the classifier frozen.