- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper proposes SLRCA, a lightweight method for root cause localization in multivariate time series. A ridge-fitted VAR(1) captures the linear lagged dependencies with variable-specific coefficients, and a single small TCN shared across all variables predicts the remaining VAR residual from each variable's own residual history and three cross-variable summary statistics. Variables are then ranked by their normalized prediction errors over the fault window, using a persistent-shift score combined with a pulse score calibrated on normal data. A proposition bounds the risk gap between shared and variable-aware predictors, and a toy example shows how VAR residualization can make sharing lossless. Experiments on four synthetic systems and SWaT, MSDS and WADI, with seven deep baselines and several ablations, report AC@1 of 0.667 on SWaT and 0.652 on MSDS using roughly 9k to 25k parameters in total.
Critical strengths
-
The method is simple and very compact (around 9k to 25k parameters in total), which is relevant for deployment on shared or limited hardware.
-
On SWaT and MSDS, the architecture-controlled comparisons in Table 5 support the design. VAR + shared TCN beats VAR alone, raw + shared, VAR + independent TCNs and the PatchTST variant under the same scoring.
-
Testing the fixed models on newly generated fault trajectories, without refitting or recalibration, is a useful check that the results are not tied to the original event set.
-
Proposition 1 appears correct under its assumptions, and the toy example in A.3 gives clear intuition for why residualizing before sharing can remove conflicting prediction targets.
Critical weaknesses
-
The combined architecture mainly helps on SWaT and MSDS, only 43 events in total. VAR alone matches or nearly matches it on three synthetic datasets, Raw+Shared is better on Lorenz96, and VAR alone has higher MRR on WADI (Table 5). Hence, I am not convinced the improvement generalizes beyond these two datasets.
-
Only 20 SWaT attacks are evaluated, without clear selection criteria. The dataset providers list 41 attacks, of which 36 are considered valid and are used in other RCA works on SWaT. AERCA uses the same count of 20, but are these the same attacks and labels? Please list the selected attacks, explain the exclusions, and report results on all 36 valid attacks.
-
The MSDS results are unusual: eight configurations (five baselines in Table 2, three variants in Table 5) score exactly zero AC@1 with only ten candidates, while SLRCA gets 15/23 correct with zero variance across seeds. Please provide the label distribution, a constant-ranking baseline, and per-event top-ranked variables.
-
How were the dataset-specific settings in Table 3 selected? Normal validation loss does not explain the scoring-window choice. The MSDS drop from 0.652 to 0.507 ± 0.153 under the alternative settings in Table 6 makes this important. Were fault events used for tuning?
-
Scoring appears to account for a substantial part of the gains. VAR alone with score S beats every published baseline in MRR on SWaT and WADI. What happens when compatible baselines receive the same calibration and scoring?
-
Baseline training needs justification: SARAD gets up to three epochs, TranAD five, and PatchTST-RCA 400 to 800 updates. Were these models trained to convergence? AERCA is missing on WADI without explanation, and lightweight baselines such as n-sigma, RCD, and CIRCA, as well as recent RCA methods evaluated on SWaT such as ShaTS[1] and CondAttr[2], are missing.
-
Using faulty history may allow persistent faults to become predictable, weakening the residual at the source. Does evaluation start at the true fault onset? The Lorenz96 failure makes diagnosis-delay and window-length sensitivity important.
-
Important design choices lack ablations. Does spectral scaling to 0.9 hurt slowly varying industrial signals? Do the shared summary channels spread fault effects into other variables' predictions? Constant-channel handling and actuator root causes also need clarification.
-
Proposition 1 does not establish that fitted VAR reduces the sharing gap or that sharing improves root-cause rankings, so it offers limited support for the proposed architecture. The novelty is also unclear given prior work on neural residual prediction and per-series statistical models with shared neural networks [3], neither of which is discussed.
References:
[1] Franco, M., Gómez, Á. L., & Maimó, L. F. (2025). ShaTS: A Shapley-based Explainability Method for Time Series Artificial Intelligence Models applied to Anomaly Detection in Industrial Internet of Things. ArXiv. https://arxiv.org/abs/2506.01450
[2] Mishra, Shashank, et al. "Conditional Attribution for Root Cause Analysis in Time-Series Anomaly Detection." Joint European Conference on Machine Learning and Knowledge Discovery in Databases. Cham: Springer Nature Switzerland, 2026.
[3] Smyl, Slawek. "A hybrid method of exponential smoothing and recurrent neural networks for time series forecasting." International journal of forecasting 36.1 (2020): 75-85.
Questions for authors
-
Which 20 SWaT attacks are used, and why were the others excluded? How does SLRCA perform on all eligible attacks, with any exclusions explained?
-
What is the root-cause label distribution on MSDS? How does a constant-ranking baseline perform, and which variable does each method rank first in each event?
-
How were the settings in Table 3 selected, in particular the scoring-window length, clipping and scale estimator? Were fault events used at any stage?
-
If compatible baselines such as GDN or TranAD use the same normalization and score S, how much of the margin remains?
-
Were SARAD, TranAD and PatchTST-RCA trained to convergence? Why is AERCA missing on WADI?
-
Does the scoring window start at the true fault onset? How do the results change if diagnosis starts a few steps later?
-
The full model offers little benefit over VAR on three synthetic datasets and WADI, while Raw+Shared is stronger on Lorenz96. Is there a measurable property of the data that tells in advance whether the shared TCN will help?
-
Is there any evidence that the VAR residuals on the real datasets come closer to the common-optimum condition of Proposition 1 than raw observations? Shared versus per-variable validation loss would be indirect evidence, but even that is not reported.
-
Nonlinear cross-variable effects can only enter the shared TCN through permutation-invariant summaries, which lose which variable produced them. How would the method capture nonlinear interactions that VAR cannot, and is there evidence that it does so on the real datasets?
-
SWaT and WADI are closed-loop systems where a source and its controller response can deviate in the same step. What assumptions does the method need to separate them, and what is the intended scope of the RCA claim?
Additional suggestions
-
Ablations without the three summary channels and without spectral scaling on the real datasets.
-
Per-variable calibration of the persistent score on normal windows, since residual autocorrelation can differ across variables.
-
Runtime and memory comparison with the baselines, not only for SLRCA.