🤖 machine learning

SRT: Super-Resolution for Time Series via Disentangled Rectified Flow

The paper introduces SRT, a novel framework that leverages disentangled rectified flow and cross-resolution attention to reconstruct high-resolution time series from low-resolution inputs by decomposing signals into trend and seasonal components, with a scaled-up SRT-large variant demonstrating strong zero-shot capabilities across multiple datasets.

Original authors: Jufang Duan, Shenglong Xiao, Yuren Zhang

Published 2026-06-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jufang Duan, Shenglong Xiao, Yuren Zhang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Blurry Photo" of Time

Imagine you are trying to watch a high-speed race car, but your camera only takes one photo every 10 seconds. You get a photo of the car at the start, then one at the finish line, but you miss the entire race in between.

In the real world, this happens all the time with data. Whether it's heart monitors, factory machines, or weather stations, we often can't record data fast enough because it's too expensive, drains batteries, or clogs up storage. We end up with "Low-Resolution" data—sparse, choppy points that miss the important details.

The Goal: The paper asks: Can we use AI to invent the missing moments between those sparse points, creating a smooth, high-speed video from a low-speed slideshow? This is called Time Series Super-Resolution (TSSR).

Why Existing Methods Fail

You might think, "Just use the same AI that makes blurry photos sharp!" (Image Super-Resolution). But the authors say that's like trying to fix a broken watch by using a paintbrush.

Images are about space (pixels next to each other).
Time Series are about time (events happening in a sequence).
The Difference: In a photo, a blurry edge is just a fuzzy line. In time data, a "blurry" gap might hide a sudden heart attack spike or a machine vibration that causes a crash. Standard AI guesses the missing parts, but often they look too smooth or miss the sharp, critical spikes.

The Solution: SRT (Super-Resolution for Time Series)

The authors built a new system called SRT. Think of SRT as a master chef who doesn't just guess the recipe; they break the dish down into its core ingredients, prepare them separately, and then mix them back together perfectly.

Here is how SRT works, step-by-step:

1. The "Deconstruction" (Disentanglement)

Instead of trying to guess the whole missing picture at once, SRT splits the data into two distinct "flavors":

The Trend (The Slow Walk): This is the general direction the data is going (e.g., the temperature slowly rising over the day).
The Seasonality (The Rhythm): This is the repeating pattern (e.g., the heartbeat's thump-thump or the daily rush hour traffic spike).

Analogy: Imagine a song. The "Trend" is the slow, steady bassline. The "Seasonality" is the catchy, repeating drum beat. SRT separates the bass from the drums so it can fix them individually.

2. The "Magic Bridge" (Implicit Time Function)

The AI needs to know where to put the new data points. It uses a tool called the Implicit Time Function (ITF).

Analogy: Think of the low-resolution data as a few wooden fence posts. The ITF is a magical rope that stretches between those posts, creating a smooth, continuous line that knows exactly where every missing plank should go, even if the fence is crooked. It aligns the "slow" data with the "fast" target timeline.

3. The "Speed Painter" (Rectified Flow)

This is the engine that actually draws the missing details. The authors use a technique called Rectified Flow.

The Old Way (Diffusion): Imagine trying to draw a picture by starting with a cloud of random noise and slowly erasing the noise until the image appears. It takes a long time and can be messy.
The SRT Way (Rectified Flow): Imagine a straight highway. The AI learns the most direct, straight path from "Random Noise" to "Perfect Data." Because the path is straight, it can drive there in just 4 steps (instead of hundreds).
The Result: It generates the missing "Seasonality" (the drum beat) and "Trend" (the bassline) separately, ensuring the sharp spikes and smooth curves are both accurate.

4. The "Conductor" (Cross-Resolution Attention)

Once the ingredients are prepped, SRT needs to mix them. It uses a special mechanism called Cross-Resolution Attention.

Analogy: Imagine a conductor leading an orchestra. The conductor looks at the low-resolution notes (the sparse data) and the high-resolution bridge (the ITF) simultaneously. They tell the musicians exactly how loud or soft to play to ensure the new notes fit perfectly with the old ones, preventing the music from sounding disjointed.

The "Super-Size" Version: SRT-Large

The authors also built SRT-Large.

Analogy: If standard SRT is a talented local chef who knows the neighborhood recipes, SRT-Large is a world-famous chef who has tasted food from every country.
How it works: They trained this model on a massive amount of data from many different industries (retail, power, transport).
The Superpower: Because it has seen so many different "recipes," it can guess the missing details for a new type of data it has never seen before, without needing any training examples for that specific task. This is called Zero-Shot capability.

Why This Matters (According to the Paper)

The paper tested SRT on nine different real-world datasets (like traffic, weather, and electricity usage).

Accuracy: It recreated the missing details better than any previous method, capturing both the smooth trends and the sharp, sudden spikes.
Speed: Because it uses the "straight highway" (Rectified Flow) method, it generates the data incredibly fast—much faster than older AI models that take hundreds of steps.
Flexibility: It works whether the missing data is just "skipped" points (Sampled) or "averaged" points (Aggregated, like turning hourly data into daily averages).

Summary

SRT is a new AI tool that takes choppy, low-speed time data and fills in the gaps to create a smooth, high-speed version. It does this by separating the data into "slow trends" and "fast rhythms," using a smart bridge to align them, and a fast "straight-line" engine to draw the missing pieces. It's faster, more accurate, and more adaptable than previous methods.

Technical Summary: SRT - Super-Resolution for Time Series via Disentangled Rectified Flow

1. Problem Definition

The paper addresses the challenge of Time Series Super-Resolution (TSSR), which aims to reconstruct high-resolution (HR) signals from low-resolution (LR) inputs. While image super-resolution is well-studied, direct transfer of computer vision techniques to time series is hindered by fundamental differences in data priors and dimensionality.

The authors formally distinguish two subtypes of TSSR problems based on the genesis of the LR data:

Sampled Super-Resolution (SSR): Reconstructing missing samples from an undersampled sequence (e.g., $l_i = h_{p_k}$ ).
Aggregated Super-Resolution (ASR): Distributing an aggregated value (e.g., a daily average) into its fine-grained constituents (e.g., hourly values), where the original high-frequency distribution is lost ( $l_i = \frac{1}{\alpha}\sum h_j$ ).

The core difficulty lies in generating credible HR details (such as sharp peaks or transient vibrations) that are absent in the LR input, a task distinct from standard time series imputation which typically assumes missing points within an originally HR sequence.

2. Methodology: The SRT Framework

The proposed SRT (Super-Resolution for Time series) framework utilizes a disentangled approach based on Rectified Flow. The architecture operates through the following pipeline:

A. Decomposition

The input LR series is decomposed into trend ( $\tau$ ) and periodic/seasonal ( $s$ ) components using an Autoformer-style decomposition (Average Pooling with padding). The model aims to generate the residual details lost during downsampling, separating the generation of long-term trends from short-term fluctuations.

B. Implicit Time Function (ITF)

To bridge the granularity gap between LR and HR time axes, the authors introduce the Implicit Time Function (ITF).

Function: ITF acts as a learnable interpolator that aligns LR trend and periodic components to the target HR resolution.
Mechanism: It employs a three-stage process:
1. Temporal Enrichment: Concatenates channels within a dilated window to capture long-term dependencies.
2. Value Prediction: Uses a neural network to predict values on the target time axis based on enriched LR features and relative time offsets.
3. Pattern Smoothness: Aggregates preliminary estimates. For trends, it enforces locality; for periodic components, it incorporates periodicity constraints using distances calculated via Fast Fourier Transform (FFT).
Scaling: For large scale factors, ITF is applied in a cascaded manner (e.g., mapping $L \to 3L \to H$ ) to ensure stability.

C. Rectified Flow Generation

The generation of the residual trend and periodic components is modeled using Rectified Flow, which learns a straight transition path from a prior distribution (Gaussian noise) to the target distribution.

Process: The model solves an Ordinary Differential Equation (ODE) to transform the initial state ( $s_0, \tau_0$ ) to the target state ( $s_1, \tau_1$ ).
Sampling: Inference uses a four-step Euler method, significantly faster than traditional diffusion models requiring hundreds of steps.

D. Velocity Predictor and Cross-Resolution Attention (CRA)

The core generative engine is a decoder-only Transformer that predicts the velocity field ( $v_s, v_\tau$ ) governing the state transitions.

Architecture: Incorporates Rotary Positional Encoding (RoPE) and Pre-Layer Normalization.
Cross-Resolution Attention (CRA): A novel mechanism comprising two sub-layers:
1. First Layer: Cross-attention to the high-resolution conditions aligned by the ITF ( $c_s, c_\tau$ ).
2. Second Layer: Cross-attention to the original LR time series.
  - For SSR, this layer is masked to only attend to available LR time steps, adjusting the generation at known points.
  - For ASR, the LR input is broadcast to all HR steps to learn aggregated segment values.

This design allows the model to condition on both fine-grained structural cues (from ITF) and coarse-grained contextual information (from LR input).

E. SRT-Large (Zero-Shot Capability)

To address scenarios where HR training data is unavailable, the authors propose SRT-large.

Scaling: Increased parameter count (30M) via more attention heads, larger FFN dimensions, and more decoder blocks.
Pre-training: Trained on large-scale, diverse datasets (retail, web search, power, transportation).
Architectural Change: The MLP component of the ITF is removed; the decoder is enhanced to directly infer hidden relations from coarser condition sequences.
Strategy: Implemented as a channel-independent model for univariate series, processing multivariate series dimension-by-dimension.

3. Key Contributions

Problem Formalization: Defined two distinct TSSR subtypes (SSR and ASR) and established a standardized evaluation protocol.
SRT Architecture: Introduced a framework combining time series decomposition, Implicit Time Functions for alignment, and Rectified Flow for efficient, high-fidelity generation.
Novel Mechanisms: Proposed the Cross-Resolution Attention (CRA) mechanism and the Implicit Time Function (ITF) to effectively fuse multi-resolution features.
Zero-Shot Extension: Developed SRT-large, demonstrating strong zero-shot super-resolution capabilities through large-scale pre-training.

4. Experimental Results

Experiments were conducted on nine public datasets (including ETTh1/2, ETThm1/2, Weather, PEMS-SF, and SelfRegulation datasets) across both SSR and ASR tasks.

Performance: SRT consistently outperformed eight baselines (including SRDiff, ResShift, IDM, FlowIE, CSDI, FTS-Diff, Diff-TS, and FlowTS) in both Mean Squared Error (MSE) and Dynamic Time Warping (DTW) distance.
Qualitative Analysis: Visualizations show SRT better reconstructs overall profiles, maintains stability in steady-state regions (avoiding excessive volatility), and recovers sharp peaks in volatile regions more faithfully than baselines.
Efficiency: SRT achieves superior generation speed (ranking 1st among baselines) due to the four-step sampling of Rectified Flow, compared to diffusion models requiring 200+ steps.
Ablation Studies: Confirmed that removing any core component (ITF, CRA, disentanglement, RoPE, Pre-LN) degrades performance, validating the necessity of each module.
SRT-large: Demonstrated state-of-the-art performance in zero-shot settings across diverse datasets and scale factors, though with higher computational costs during inference.

5. Significance and Claims

The paper claims that SRT addresses the fundamental challenges of TSSR by moving beyond simple adaptation of image super-resolution techniques. By explicitly disentangling temporal dynamics (trend vs. periodicity) and utilizing a rectified flow framework guided by cross-resolution attention, the model achieves:

High Fidelity: Accurate reconstruction of both stable and volatile phases.
Efficiency: Fast inference suitable for practical applications.
Generalizability: The ability to handle both sampled and aggregated data types, and the extension to zero-shot scenarios via SRT-large.

The authors acknowledge that the approach relies on structural assumptions (decomposability of series into trend and periodicity) which may not strictly hold for all real-world data behaviors. Future work is suggested to explore adaptive methods for prior selection to further enhance flexibility. The research is presented as a foundational step toward more robust and universally applicable time series super-resolution strategies.