A novel sample selection method based on kernel entropy component analysis and cumulative distance distribution

Assessing whether incoming samples fall within the applicability domain of an established spectroscopic model is essential for maintaining prediction accuracy during routine instrument operation. Samples identified as novel can be selected for reference analysis and subsequently used for model updating, thereby improving model adaptability to newly emerging sample characteristics. Conventional approaches usually decompose spectra into different projection spaces and evaluate statistics such as Hotelling’s T² and Q -residuals against predefined thresholds. However, these methods often measure novelty by comparing a streaming sample with a global representation of the calibration set, such as its centroid, and therefore do not fully account for the internal distribution of the calibration samples. To address this limitation, we propose an integrated method combining kernel entropy component analysis (KECA) with cumulative distance distribution (CDD). The proposed method first constructs a supervised PLS latent space and then applies KECA to obtain a low-dimensional entropy-component space. Each incoming sample is then projected into this space, and its novelty is quantified from the normalized distance distribution between the sample and all calibration samples. KECA determines the low-dimensional feature space for distance calculation, while CDD aggregates the resulting distances into a quantitative novelty measure. The decision threshold is obtained by rank-based interpolation of the CDD values calculated from the calibration set. Experimental results show that the proposed method can effectively identify novel samples in streaming spectroscopic data.