LSTM-Powered Intelligent Signal Upsampling in Automotive Telematics for Data Compression and Anomaly Identification

High-frequency automotive sensor signals play a critical role in vehicle diagnostics, predictive maintenance, and safety monitoring. However, continuous logging at fine temporal resolution imposes significant storage, bandwidth, and computational constraints in automotive telematics systems. Consequently, signals are often recorded at sparse sampling rates, leading to the loss of transient dynamics that may be diagnostically relevant. This work addresses boundary-conditioned signal reconstruction under extreme temporal sparsity, where only sparse anchor samples are available at inference time. A Long Short-Term Memory (LSTM)-based conditional reconstruction framework is presented to infer high-frequency signal trajectories from sparsely logged data. The proposed approach enforces temporal continuity across reconstruction windows through state rollover and feedback mechanisms. In addition to signal reconstruction, a reconstruction-consistency–based anomaly detection strategy is introduced, where transient anomalies correspond to segments that cannot be reliably inferred from sparse observations under learned normal dynamics. Reconstruction uncertainty is estimated via stochastic inference and used to indicate anomalous behavior. Experimental results on real-world automotive telemetry data demonstrate that the proposed framework achieves lower reconstruction error than classical interpolation methods, Autoencoders, and standard LSTM baselines, while providing enhanced sensitivity to transient anomalies in sparsely sampled signals. Boundary-conditioned reconstruction with uncertainty-aware analysis provides a practical foundation for storage-efficient signal analysis in automotive telematics.