Reduce retrieval label noise
Raw classification labels can mix morphology inconsistent instances or highly ambiguous cross class matches. UCR-R removes these cases before retrieval benchmarking.
Dataset Construction
Dataset Construction
TSRBench combines the open UCR-R benchmark and the new industrial CU-RCA dataset. This page explains the UCR-R reconstruction logic, surfaces a traceability example, and summarizes the key facts that define both benchmarks.
Construction principles
Our goal is to make retrieval ground truth more reliable while preserving retrieval difficulty.
Keep semantic and morphology alignment
Retained positives should agree in both semantic meaning and morphology, so retrieval labels match the intended TSR task.
Preserve realistic challenge
The filtering removes misleading labels rather than all variation. The remaining benchmark still contains non-trivial shape diversity and realistic retrieval difficulty.
UCR-R reconstruction pipeline
UCR-R is reconstructed in two stages so that retrieval labels better match morphology and physical meaning.
Stage 1
same class instance filtering
Within one original UCR class, annotators identify the morphology pattern and remove instances that clearly deviate from the class core shape.
- Start from one original class and inspect its morphology pattern rather than isolated local spikes.
- Remove instances that visibly drift away from the class core shape and would act as misleading positives in retrieval.
- Retain the curves that stay consistent with the morphology pattern as the reconstructed retrieval set.
Stage 2
cross class confusable removal
After completing the intraclass filtering, if the morphologies of the two original categories are still highly indistinguishable, then only one of them is selected and retained to construct the retrieval pool.
Why this figure is schematic
The stage two illustration is an explanatory schematic. It clarifies the rule: when two original classes have nearly indistinguishable morphology, they are not treated as clean retrieval labels.
Representative evidence
A real reconstruction case and a compact audit excerpt show how reconstruction decisions are surfaced on the website.
Real reconstruction case
Representative UCR-R reconstruction example
In this real world reconstruction case, the first image below shows five curves from the same category in the initial set. The curve marked in red deviates significantly in the latter half and is therefore removed. The remaining curves with consistent shapes are retained to construct the retrieval pool. The second image shows two other categories of curves, but their main shapes are similar, so they need to be removed entirely.
Traceability example
Audit excerpt
| Dataset | Original class | Decision stage | Decision | Raw class size | Retained in UCR-R |
|---|---|---|---|---|---|
| BME | 1 | Instance level filtering | Keep | 60 | 19 |
| BME | 2 | Fully retained | Keep | 60 | 60 |
| BME | 3 | Class level exclusion | Remove | 60 | 0 |
| Beef | 1 | Class level exclusion | Remove | 12 | 0 |
| Beef | 2 | Fully retained | Keep | 12 | 12 |
| CBF | 1 | Instance level filtering | Keep | 310 | 40 |
Dataset details
TSRBench uses two complementary datasets: one open benchmark for shape consistent retrieval and one new industrial benchmark for incident-centric RCA retrieval.
Open retrieval benchmark
UCR-R
Reconstructed classes
46
Time series
2,600
- Two-stage reconstruction with shape consistent positives.
- Each query is evaluated against a complete candidate pool.
- Designed for multitarget TSR benchmarking with cleaner retrieval labels.
New industrial dataset
CU-RCA
Length per raw long series
11,520
Incident windows
103
- Incident centric retrieval for RCA-oriented analysis.
- Built from two raw long time series segments of the same length（11520）.
- Captures noisy operational telemetry and industrial domain shift.
Why both datasets matter
UCR-R emphasizes open, shape consistent retrieval , while CU-RCA emphasizes noisy industrial telemetry, incident centric tasks, and RCA utility.