Knock detection and diagnosis in engines based on the WPKNN approach

Knock detection is crucial for ensuring engine reliability and performance. Conventional approaches, such as the maximum amplitude of pressure oscillations (MAPO), face limitations including poor noise robustness, low sensitivity to incipient knock, and reliance on empirically defined thresholds. This study proposes a knock detection and diagnosis framework WPKNN based on weighted PCA and k-nearest neighbors. By constructing a time-sensitive weighted principal component score sequence, the method effectively captures weak incipient knock features, significantly enhancing detection sensitivity and accuracy. Furthermore, a novel statistic KNNpcW is introduced that integrates local neighborhood variations and global principal component information, enabling precise identification of knock. It overcomes the limitations of the MAPO method, which relies solely on a single-peak indicator. In addition, kernel density estimation (KDE) is employed to adaptively determine dual thresholds for slight and severe knock, thereby improving the knock risk management. To support knock tracing, an advanced contribution-based diagnostic mechanism is established. It combines enhanced complete decomposition contribution (ECDC) and enhanced partial decomposition contribution (EPDC) methods, which accurately identify the key parameters most responsible for knock evolution. Experiments on a turbocharged gasoline engine under various knock conditions confirm that the WPKNN framework improves detection reliability, and provides robust support for intelligent combustion control and active knock suppression strategies.