Anomaly Detection on WADI
49.51F1 ScoreTranAD
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| TranADTraining=Complete training data2022.01 | 49.51 | 35.29 | 82.96 | 89.68 | — | |
| OmniAnomalyTraining=Complete training data2022.01 | 42.6 | 31.58 | 65.41 | 81.98 | — | |
| GDNTraining=Complete training data2022.01 | 42.6 | 29.12 | 79.31 | 87.77 | — | |
| MTAD-GATTraining=Complete training data2022.01 | 41.69 | 28.18 | 80.12 | 88.21 | — | |
| CAE-MTraining=Complete training data2022.01 | 41.17 | 27.82 | 79.18 | 87.28 | — | |
| MSCREDTraining=Complete training data2022.01 | 37.41 | 25.13 | 73.19 | 84.12 | — | |
| MAD-GANTraining=Complete training data2022.01 | 35.88 | 22.33 | 91.24 | 80.26 | — | |
| USADTraining=Complete training data2022.01 | 30.56 | 18.73 | 82.96 | 87.23 | — | |
| DAGMMTraining=Complete training data2022.01 | 14.12 | 7.6 | 99.81 | 85.63 | — | |
| MERLINTraining=Complete training data2022.01 | 11.74 | 6.36 | 76.69 | 59.12 | — | |
| LSTM-NDTTraining=Complete training data2022.01 | 2.71 | 1.38 | 78.23 | 67.21 | — | |
| Ti-iLSTMFeat. (Number of features)=10, Win. (Window size)=10, HW (Hardware)=PLC-HW, Detection Focus=Logic-layer inconsistencies, PLC-D (Solution deployed on PLC)=Yes2026.05 | 0.96 | — | — | — | 0.95 | |
| Digital Twin + MLFeat. (Number of features)=NR, Win. (Window size)=NR, HW (Hardware)=GPU, Detection Focus=Anomalies in system behaviour, PLC-D (Solution deployed on PLC)=No2026.05 | 0.94 | — | — | — | — | |
| Graph NN (edge-cond.)Feat. (Number of features)=51/127, Win. (Window size)=5, HW (Hardware)=GPU, Detection Focus=Sensor/actuator relations, PLC-D (Solution deployed on PLC)=No2026.05 | 0.85 | — | — | — | — | |
| Masked GNNFeat. (Number of features)=51/123, Win. (Window size)=5, HW (Hardware)=GPU, Detection Focus=Correlation by masking, PLC-D (Solution deployed on PLC)=No2026.05 | 0.83 | — | — | — | — | |
| ASTRO2026.05 | 0.7885 | 0.9024 | 0.7002 | — | — | |
| VAE-LSTMFeat. (Number of features)=51/123, Win. (Window size)=NR, HW (Hardware)=GPU, Detection Focus=Latent encoding + temporal, PLC-D (Solution deployed on PLC)=No2026.05 | 0.75 | — | — | — | — | |
| DGNN2026.05 | 0.6497 | 0.8017 | 0.5467 | — | — | |
| IT-DT2026.03 | 0.615 | 0.785 | 0.505 | — | — | |
| TranAD2026.03 | 0.602 | 0.751 | 0.501 | — | — | |
| GDN2026.03 | 0.57 | 0.71 | 0.475 | — | — | |
| GDN2026.05 | 0.5692 | 0.975 | 0.4019 | — | — | |
| USAD2026.03 | 0.542 | 0.682 | 0.45 | — | — | |
| TranAD2026.05 | 0.4951 | 0.3529 | 0.8296 | — | — | |
| GRN2026.05 | 0.4828 | 0.3584 | 0.7398 | — | — | |
| PCA2026.03 | 0.395 | 0.551 | 0.306 | — | — | |
| MAD GAN2026.05 | 0.373 | 0.4144 | 0.3392 | — | — | |
| DAGMM2026.05 | 0.3609 | 0.5444 | 0.2699 | — | — | |
| AE2026.05 | 0.3435 | 0.3435 | 0.3435 | — | — | |
| LSTM VAE2026.05 | 0.2482 | 0.8779 | 0.1445 | — | — | |
| USAD2026.05 | 0.2328 | 0.9947 | 0.1318 | — | — | |
| OmniAnomaly2026.05 | 0.2296 | 0.9947 | 0.1298 | — | — | |
| MERLINTraining data percentage=20%2022.01 | 0.1174 | — | — | 0.5912 | — | |
| OmniAnomalyTraining data percentage=20%2022.01 | 0.1017 | — | — | 0.7913 | — | |
| PCA2026.05 | 0.0986 | 0.3953 | 0.0563 | — | — | |
| MAD-GANTraining data percentage=20%2022.01 | 0.0937 | — | — | 0.5383 | — | |
| CAE-MTraining data percentage=20%2022.01 | 0.0781 | — | — | 0.6109 | — | |
| USADTraining data percentage=20%2022.01 | 0.0733 | — | — | 0.7011 | — | |
| TranADTraining data percentage=20%2022.01 | 0.0649 | — | — | 0.7688 | — | |
| DAGMMTraining data percentage=20%2022.01 | 0.063 | — | — | 0.6497 | — | |
| MTAD-GATTraining data percentage=20%2022.01 | 0.052 | — | — | 0.6267 | — | |
| MSCREDTraining data percentage=20%2022.01 | 0.0413 | — | — | 0.6029 | — | |
| GDNTraining data percentage=20%2022.01 | 0.0412 | — | — | 0.6121 | — | |
| LSTM-NDTTraining data percentage=20%2022.01 | 0 | — | — | 0.6637 | — | |
| GWO + AutoencoderFeat. (Number of features)=NR, Win. (Window size)=NR, HW (Hardware)=Server, Detection Focus=Feature sel.+reconstruction, PLC-D (Solution deployed on PLC)=No2026.05 | — | — | — | — | 0.99 |