Signal or Noise? Understanding Generative Models for Real-World Sensor Time Series

Sep 21, 2026

Upload by Levi

Generative models reshape machine learning’s ability to model complex data distributions, yet real-world observations often take the form of continuous, high-dimensional, noisy sensor time series. Current sensor generation research is fragmented across modalities, datasets, and task formulations, limiting a systematic understanding of when, how, and why models succeed or fail. SensorGen is introduced as a large-scale open study across 14 settings, 4 domains, 7 datasets, and 12 signal modalities, enabling controlled evaluation of multiple model families and uncovering key success drivers.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages38

Reading Time1 h 36 min

AccessFree to download

Introduction

SensorGen Overview

Task Coverage and Standardized Pipeline

Sensor Diversity and Generation Settings

Model Families and Evaluation Findings

arXiv :2607 .04245v 1 [ cs .LG] 5 Jul 2026 2026-7-7 Signal or Noise? Understanding Generative Models for Real-World Sensor Time Series Zitao Shuai1∗ , Zongzhe Xu1∗ , Yuntian Wu2∗ , Sirui Li1 , Tianhong Li3 , Yuzhe Yang1† 1University of California, Los Angeles, 2 Carnegie Mellon University, 3Massachusetts Institute of Technology Generative models have changed how machine learning represents complex data distributions, especially in language and vision, yet many real-world systems are observed instead as continuous, high-dimensional, and noisy sensor time series. Existing generative modeling of sensor data, however, remains fragmented across modalities, datasets, and task formulations, limiting a systematic understanding of when, how, and why generative models succeed or fail in real-world settings. To address this gap, we introduce SensorGen, a large-scale study of sensor-signal generation spanning 14 settings across 4 domains, 7 datasets, and 12 signal modalities. Leveraging SensorGen, we systematically evaluate generative models from five major families and uncover three key findings: (1) flow-matching models provide strong overall performance across most settings; (2) signal properties matter, with demographic covariates improving longitudinal generation and time-frequency modeling improving high-frequency signal generation; and (3) generated signals have practical utility beyond visual realism, with scaling improving generation quality and synthetic data improving downstream performance. Together, SensorGen establishes a broader understanding of design choices, evaluation protocols, and failure modes in real-world sensor data generation. Huggingface: [https://huggingface.co/yang-ai-lab/SensorGen](https://huggingface.co/yang-ai-lab/SensorGen) Code: [https://github.com/yang-ai-lab/SensorGen](https://github.com/yang-ai-lab/SensorGen) Website: [https://yang-ai-lab.github.io/sensor-gen](https://yang-ai-lab.github.io/sensor-gen) 1. Introduction Generative models have reshaped how machine learning imagines data: learning not only to recognize patterns, but to sample, complete, and represent complex distributions [28, 10] . Yet, much of the real world is not expressed as language or images, but as continuous, noisy, high-dimensional sensor time series that record physiology [41], behavior [3], environments [25], and machines [15] . This makes sensor time series a central data regime for modern machine learning, but extending generative modeling advances to sensor signals is non-trivial. Unlike language or natural images, sensor signals are collected across diverse settings and modalities, and vary widely in sampling frequency, time span, sequence length, channel structure, and physical semantics. Existing studies for sensor time series [16, 6, 12], however, remain highly fragmented across modalities, datasets, and task formulations. Current sensor generation methods are often designed for specific applications, with modeling paradigms, architectures, and evaluation protocols tailored to a particular signal type or task. While effective in their target settings, these dedicated designs provide limited evidence about whether the modeling choices generalize across heterogeneous sensor generative problems. As a result, the field lacks a systematic understanding of the factors that shape generation quality across sensor signals. This motivates a central question: When, how, and why do generative models succeed or fail across diverse sensor time series generation settings? Figure 1 | Overview of SensorGen. We present a large-scale study of real-world sensor time-series generation, spanning 14 settings across 4 domains, 7 datasets, and 12 signal modalities. SensorGen establishes to date the broadest coverage of sequence length, frequency, and time span. More details are in Appendix B.1 . To answer this question, we introduce SensorGen, a fully open and large-scale exploration of generative modeling for real-w

FAQs

What problem does SensorGen address in generative modeling for sensor data?

It addresses fragmented research across sensor modalities, datasets, and task formulations by enabling a systematic study of when, how, and why generative models succeed or fail for real-world sensor time series.

What does SensorGen include in terms of coverage and scope?

SensorGen spans 14 generation settings across 4 domains, 7 datasets, and 12 signal modalities, designed to cover diverse sampling frequencies, sequence lengths, and time spans.

What key findings are reported about model family and signal properties?

Flow-matching models provide strong overall performance across most settings, and signal-property-aware design matters: demographic covariates improve longitudinal generation while time-frequency modeling improves high-resolution signal generation.

Sep 21, 2026

Upload by Levi