@article{4682,
  author = {Ricardo Rodríguez Jorge},
  title = {A Novel Feature Engineering Framework for Capacitive Sensor Time Series Classification: Distinguishing Water and Oil Immersion Through Sequential Pattern Analysis},
  journal = {Progress in Signals and Telecommunication Engineering},
  year = {2026},
  volume = {15},
  number = {1},
  doi = {https://doi.org/10.6025/pste/2026/15/1/14-25},
  url = {https://www.dline.info/pste/fulltext/v15n1/pstev15n1_2.pdf},
  abstract = {This paper presents a novel feature engineering framework for time series classification of capacitive sensor
data to distinguish between water and oil immersion conditions. Capacitive sensors are widely used in
industrial monitoring and human machine interaction applications, yet accurately classifying short sequential
signals remains challenging due to temporal dependencies and subtle pattern variations. The dataset consists
of ten consecutive intensity measurements from sensor electrodes, labeled according to immersion medium.
While raw sequential values contain discriminative information, their high correlation and limited length
make it difficult to extract meaningful temporal dynamics.
To address this, we design a comprehensive feature set encompassing trend based, complexity-based, stability,
and position specific descriptors. Key features include linear slope, first last difference, entropy, variance,
stability, and change point counts, which collectively encode global trends, variability, and structural
irregularities within each sequence. Experimental evaluation using Logistic Regression, Random Forest,
SVM, and k-NN demonstrates that engineered features substantially enhance classification performance.
Notably, combining raw and engineered features yields near perfect performance, achieving 99.93%
accuracy and a ROC-AUC of 0.99999 with Logistic Regression, confirming that the proposed descriptors
provide complementary discriminative power.
Feature importance and coefficient analyses further reveal that temporal trend strength, entropy, and signal
stability are critical in differentiating water from oil immersion, reflecting underlying dielectric property
differences. The proposed framework balances predictive accuracy and interpretability, offering a scalable
and transparent solution for short sequence binary sensor classification problems.},
}