Bridging the Resolution Gap: Conditional Generative Model for High- Fidelity Time-Series in Automotive Applications

Abstract

The automotive sector increasingly relies on high-resolution vehicle data for advanced analytics, yet readily available customer vehicle data is often sparse and lowresolution due to hardware, cost, and privacy constraints. Existing time-series generative models frequently struggle with long sequences (exceeding 1000 timesteps) and lack robust mechanisms for controlled generation using continuous conditional inputs. This thesis aims to bridge this data resolution gap by developing and evaluating advanced machine learning models capable of generating high-fidelity, longsequence time-series data from low-resolution conditional inputs. We evaluated current state-of-the-art time-series generative models for their efficacy in handling long sequences, identifying Variational Autoencoders (VAEs), particularly TimeVAE, as the most promising base due to their fast training time and superior predictive performance. A novel conditional generative model was then developed by integrating TimeVAE with a Transformer encoder, adapting the PatchTST architecture to encode the conditional information. Key architectural enhancements, including Rotary Positional Embeddings, Layer Normalization, and specifically, a residual integration approach for incorporating conditional information, were explored. Further improvements like "free bits" and "conditional prior" were implemented to mitigate posterior collapse and enhance overall model performance. Our findings indicate that state-of-the-art models generally underperform on long sequences , and naive conditional integration is ineffective due to complex gradient flows. However, the proposed residual integration improved the model’s ability to leverage conditional information. The combined application of "free bits" and "conditional prior" alleviated posterior collapse, leading to a more robust model. Crucially, the purely data-driven model was able to generate physically plausible long-sequence time-series data, with three key physical metrics showing less than 16% deviation from real signals, without explicit physics-based training. This demonstrates a viable solution for controlled high-fidelity time-series generation in automotive applications