An upsampling framework for vehicle low-frequency sampling data based on human-machine-environment fusion

Vehicle reliability fundamentally governs transportation efficiency, energy utilization optimization, and public safety, where signal fidelity determines diagnostic accuracy in health management systems. Conventional vehicle-mounted recorders predominantly employ low-frequency sampling due to cost and power constraints, inevitably causing transient fault feature loss through aliasing and dynamic interference. This paper introduced a human-vehicle-environment fused framework for physics-consistent signal upsampling. The approach systematically combines a power flow-driven stacked autoencoder for hierarchical feature extraction, with tensor product-based fusion of driver behavior, road resistance, and vehicle state parameters as physical constraints. By establishing adversarial learning conditioned on the fusion feature coupled with progressive optimization of time–frequency reconstruction objectives, the framework effectively addresses spectral distortion in conventional methods. Experimental validation on heavy-duty hybrid vehicle data demonstrates statistically significant spectral recovery capabilities, achieving Pearson correlation coefficients of 0.766 and 0.783 with original high-frequency reference signal in fault progression scenarios, supported by Wilcoxon rank-sum test p-values of 0.209 and 0.217. The framework outperforms cubic spline interpolation, standard GAN, and advanced benchmarks (e.g., TCN and UR-CNN) in preserving transient signatures critical for incipient fault diagnosis, with RMSE reductions of up to 65.83% against interpolation and 29.33% against GAN variants. By reconstructing physically consistent high frequency signals from low frequency data, this methodology bridges the gap between constrained onboard recordings and diagnostic requirements.