Adaptive Principal Component Analysis combined with Fuzzy K-Nearest Neighbors for Activity Recognition Using Multisensor Data Fusion

In the realm of activity recognition, the Fuzzy K-Nearest Neighbors (FKNN) algorithm stands as a pivotal technique for classification tasks. However, the performance of FKNN is heavily reliant on the choice of its hyperparameters, notably the number of neighbors (k) and the fuzziness parameter (m). Optimizing these hyperparameters is a non-trivial challenge, significantly affecting classification accuracy. To address this challenge, an integrated approach that combines Principal Component Analysis (PCA) and an adaptive grid search algorithm for hyperparameter tuning is proposed. Utilizing a dataset captured from a Wireless Sensor Network (WSN) that measures Received Signal Strength (RSS) from sensor nodes placed on a user’s chest and ankles, the methodology initiates with preprocessing to extract time-domain features, such as mean and variance, from the raw RSS data. Strategic ranges for hyperparameters are defined to effectively explore the parameter space. PCA is employed to reduce dimensionality and enhance data structure visibility. The core of the approach lies in adaptive grid search, which iteratively refines the hyperparameter search space based on observed performance metrics, converging on optimal values that yield the highest classification accuracy. The FKNN model, trained with these optimized parameters, is then evaluated on a separate test set. Results demonstrate that this approach not only enhances the classification accuracy of the FKNN algorithm but also achieves a balanced performance in activity recognition using multisensor data fusion.