WaveletKAN-LSTM Multi Source Fusion for Fire Risk Ranking in Battery Containers
Authors
Files
Abstract
Battery container monitoring requires reliable inspection priorities when heterogeneous sensor streams indicate different degrees of concern. This study presents a multi-source ranking framework that combines long short-term memory with a wavelet-based Kolmogorov-Arnold readout and evaluates its contribution to inspection prioritization. The evaluation dataset contains 1,200 container records arranged into 100 inspection groups, with eight synchronized channels and 24 observations per record. Complete groups are assigned to training, validation, and testing to preserve separation between ranking lists. Recurrent weights and wavelet centres remain fixed, while regularized readout coefficients are fitted using training labels. Across five recurrent initializations, the hybrid achieves a normalized discounted cumulative gain at three of 0.9737, top-three recall of 0.8600, and mean absolute relevance error of 0.1343. Its gain over the recurrent linear readout is 0.0088, with a conditional group bootstrap interval from 0.0037 to 0.0149. Snapshot Ridge achieves a higher ranking score of 0.9763. Under standardized noise with a standard deviation of one unit, the representative hybrid score is 0.9294; replacing forty percent of input entries with training means reduces it to 0.8933. These results identify a nonlinear temporal readout benefit and establish the static estimator as a reference for evaluation across different operating conditions. The output supports relative inspection prioritization within a comparison group; interpreting it as an event probability requires separate calibration.
