Raw + robust features
Time space
Raw values, robust standardization, and multiscale local patches expose amplitude shifts, abrupt transitions, and waveform distortion.
SHADE · Methods & evidence
How SHADE works.
Sparse Hybrid Anomaly Detection Engine produces one timestamp-level anomaly score from a time series and its declared known-normal prefix.
Three signal spaces. Five expert views. One score.
SHADE preserves time, frequency, and structural information as separate representations, brings them into a shared temporal context, and fuses five complementary evidence sources at every timestamp.
InputTime series + anomaly-free reference prefixComplete offline sequence · one or more channels
Tri-space representation
Global + local spectra
Frequency space
Compact Fourier summaries and timestamp-aligned local energy preserve persistent rhythms, missing cycles, and short spectral changes.
Whitened dependencies
Structural space
Regularized whitening exposes changes in cross-channel coordination, including faults that remain subtle in each channel alone.
Shared local + long-memory context
Temporal intelligenceConvolution + scalar and matrix memoryLocal behavior and long-range dependencies share one contextual representation.
Five timestamp-level evidence sources
Calibrated multi-scale fusion
Sparse-channel aggregationMean, top fractions, and attention
→Empirical tail calibrationReference-prefix surprise
→Multi-scale temporal supportPoint and extended events
Dense anomaly rankingOne score for every timestamp
02 · Real inference exampleOne MSL series.
One MSL series.
Fifty-five real channels.
The display uses five channels chosen by variance within the label-free normal prefix. Every line and score comes from the saved official-driver artifact.
Series002_MSL_id_1_Sensor_tr_500_1st_900.csv
Channel 5Channel 49Channel 11Channel 39Channel 47
Known-normal prefix boundaryLabeled event
Selected channels
SHADE score
1.00.50.0
0Timestamp1826
What the chart shows: SHADE scored 1,827 timestamps using a 500-point known-normal prefix. The optional gold overlay marks the evaluator’s event at timestamps 900–910. All 11 event timestamps rank among the score’s top 20 values, and the full-resolution view includes every timestamp.
Scoring · Offline, reference-calibratedEvidence · 5 of 55 channels shownRun · July 27, 2026
Strong results across 530 time-series files.
Across single-channel and multichannel tracks, SHADE delivers a mean VUS-PR above 0.65 while producing the same clear, timestamp-aligned score for every series.
TSB-AD performance snapshotJuly 27, 2026
| Track | Series | Mean VUS-PR ↑ | Median VUS-PR ↑ | Minimum | Maximum |
|---|---|---|---|---|---|
| TSB-AD-U | 350 | 0.6511400852 | 0.7472997859 | 0.0008671327 | 1.0000000000 |
| TSB-AD-M | 180 | 0.6505942465 | 0.7451343151 | 0.0120910314 | 0.9999662246 |
530 community-submission evaluation series, with one metric row per series and arithmetic means across rows. Values are reproduced exactly from the paper.
04 · Score aggregation
Strong evidence stays visible.
SHADE carries distinct anomaly hypotheses through scoring, then combines them in ways that preserve localized faults, comparable confidence, and the right temporal support.
Channel aware
Sparse-channel aggregation
Mean, top-5%, top-20%, and attention summaries keep a fault affecting one or a few channels from disappearing inside a full-channel average.
Reference aware
Empirical calibration
Reference-prefix score distributions translate unlike residuals into comparable tail surprise without imposing a parametric distribution.
Duration aware
Multi-scale temporal support
Identity, short, medium, and long supports preserve sharp point anomalies while building sustained evidence for extended events.