<![CDATA[This paper presents an integrated framework combining Continuous Wavelet Transform (CWT) with multi-stream Long Short-Term Memory and Multi-Layer Perceptron (LSTM+MLP) networks for early wheel slip detection in urban Light Rail Transit (LRT). Conventional Wheel Slip Protection (WSP) systems rely on static thresholds applied to axle speed differences, resulting in detection latency that impacts wheel-rail wear and operational disruptions. Three methodological contributions are proposed: (i) a five-channel CWT feature extraction strategy that leverages the structural hierarchy of LRT trainsets consisting of motor car with cabin (MC), motor car (M), and trailer car (T); (ii) a conventional ground truth labeling workflow using Slip/Slide Status signals as a reference, eliminating expert annotation bias; and (iii) the first distance-based spatial slip hotspot mapping integrated with model prediction for the Jakarta LRT line. The framework is validated on a 50-minute operational dataset consisting of 93,755 time steps with a resolution of 32 ms, collected from the LRT and synchronized via Dynamic Time Warping using 47 Train Control and Management System (TCMS) variables with three different sampling rates. Statistical analysis confirms that the motor-trailer speed difference channel achieves a Cohen's d of 1.615 with a Kolmogorov-Smirnov p < 0.001 test between normal and slip distributions. The integrated LSTM+MLP model achieves an F1 score of 0.9272, an AUC of 0.9958, and a median early detection margin of 1.152 seconds before conventional WSP activation, with 100% of test events detected earlier than the conventional system. Complementary spatial mapping derived from a larger multi-train dataset containing 5,948 events identified 414 high-risk zones along the LRT track, supporting predictive WSP threshold modulation and maintenance scheduling that takes rail wear into account.]]>
