ts-anomaly-detection

Time SeriesTime-Series-Library

Description

Time Series Anomaly Detection: Custom Model Design

Objective

Design and implement a custom deep learning model for unsupervised time series anomaly detection via reconstruction. Your code goes in the Model class in models/Custom.py. Three reference implementations (DLinear, TimesNet, PatchTST) are provided as read-only.

Evaluation

Trained and evaluated on three anomaly detection datasets:

PSM (25 variables, server machine dataset)
MSL (55 variables, Mars Science Laboratory)
SMAP (25 variables, Soil Moisture Active Passive satellite)

All use seq_len=100, anomaly_ratio=1. Metric: F-score (higher is better).

Code

Custom.py

EditableRead-only

1import torch
2import torch.nn as nn
3
4
5class Model(nn.Module):
6    """
7    Custom model for time series anomaly detection.
8
9    Forward signature: forward(x_enc, x_mark_enc, x_dec, x_mark_dec, mask=None)
10    - x_enc: [batch, seq_len, enc_in] — input time series
11    - x_mark_enc: not used for anomaly detection (None)
12    - x_dec: not used for anomaly detection (None)
13    - x_mark_dec: not used for anomaly detection (None)
14
15    Must return: [batch, seq_len, c_out] — reconstructed sequence

Additional context files (read-only):

Time-Series-Library/models/DLinear.py
Time-Series-Library/models/TimesNet.py
Time-Series-Library/models/PatchTST.py
Time-Series-Library/layers/AutoCorrelation.py
Time-Series-Library/layers/Autoformer_EncDec.py
Time-Series-Library/layers/Conv_Blocks.py
Time-Series-Library/layers/Crossformer_EncDec.py
Time-Series-Library/layers/Embed.py
Time-Series-Library/layers/FourierCorrelation.py
Time-Series-Library/layers/SelfAttention_Family.py
Time-Series-Library/layers/StandardNorm.py
Time-Series-Library/layers/Transformer_EncDec.py

Results

Show per-seed results

Model	Type	f score PSM ↑	precision PSM ↑	recall PSM ↑	f score MSL ↑	precision MSL ↑	recall MSL ↑	f score SMAP ↑	precision SMAP ↑	recall SMAP ↑
dlinear	baseline	0.966	0.987	0.947	0.819	0.897	0.753	0.673	0.899	0.538
patchtst	baseline	0.962	0.989	0.936	0.790	0.887	0.713	0.688	0.904	0.556
timesnet	baseline	0.974	0.985	0.962	0.804	0.888	0.736	0.694	0.899	0.565
timesnet	baseline	0.974	0.985	0.963	0.818	0.895	0.753	0.694	0.901	0.565
timesnet	baseline	0.973	0.985	0.962	0.804	0.883	0.738	0.693	0.898	0.564
timesnet	baseline	0.973	0.984	0.962	0.810	0.892	0.742	0.695	0.898	0.566
anthropic/claude-opus-4.6	vanilla	0.937	0.982	0.897	0.775	0.877	0.694	0.671	0.899	0.535
deepseek-reasoner	vanilla	0.922	0.980	0.871	0.776	0.892	0.687	0.685	0.893	0.556
google/gemini-3.1-pro-preview	vanilla	0.937	0.990	0.889	0.739	0.871	0.642	0.683	0.900	0.550
gpt-5.4-pro	vanilla	0.968	0.985	0.950	0.781	0.880	0.702	0.688	0.904	0.555
openai/gpt-5.4-pro	vanilla	0.969	0.986	0.953	0.801	0.862	0.748	0.687	0.895	0.557
qwen/qwen3.6-plus:free	vanilla	0.933	0.982	0.888	0.317	0.636	0.211	0.685	0.900	0.553
anthropic/claude-opus-4.6	agent	0.964	0.986	0.943	0.791	0.885	0.716	0.673	0.896	0.538
deepseek-reasoner	agent	0.928	0.993	0.872	0.790	0.894	0.707	0.684	0.894	0.554
google/gemini-3.1-pro-preview	agent	0.954	0.986	0.923	0.734	0.863	0.639	0.675	0.902	0.539
gpt-5.4-pro	agent	0.968	0.985	0.950	0.781	0.880	0.702	0.688	0.904	0.555
openai/gpt-5.4-pro	agent	0.969	0.986	0.953	0.801	0.862	0.748	0.687	0.895	0.557
qwen/qwen3.6-plus:free	agent	0.932	0.990	0.881	-	-	-	-	-	-

Agent Conversations

anthropic/claude-opus-4.6

5 steps

deepseek-reasoner

13 steps

google/gemini-3.1-pro-preview