acceptodds
Under review as a conference paper at ICLR 2027

Hierarchical Granger-Causal Root Cause Analysis of Anomalies in Industrial Multivariate Time Series

Abstract

Root cause analysis (RCA) of industrial multivariate time series aims to identify the variables directly responsible for anomalous events. This task is challenging because disturbances can propagate across interacting subsystems, causing many sensors to respond to a small number of root variables. Existing methods mainly model dependencies at the variable level, which may mix subsystem-level dynamics with local variable interactions. We propose HiRCA, a hierarchical Granger-causal predictive framework that models these two levels separately. HiRCA uses a soft engineering prior on subsystem membership to learn latent subsystem states and their temporal dependencies. At the variable level, it separates the effects of self-history, shared subsystem context, and cross-variable interactions on prediction. For diagnosis, HiRCA first aggregates calibrated innovation scores to identify candidate subsystems and then prioritizes their variables in the final ranking. On two synthetic benchmarks, HiRCA shows strong hierarchical graph-recovery performance and achieves Avg@10 scores above 0.90 for root cause localization. Across four real-world datasets, HiRCA ranks among the top three in nearly all evaluations, including 16 top-two variable ranking results and 15 best time-variable ranking results. These results demonstrate the effectiveness of hierarchical predictive modeling for structured, coarse-to-fine root cause localization in industrial multivariate time series.

Primary area:
learning on time series and dynamical systems
Keywords:
root cause analysis, multivariate time series, Granger causality, industrial anomaly diagnosis, hierarchical modeling

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_nFFDOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

HiRCA uses an engineering subsystem partition as a soft prior to learn latent subsystem states and their temporal graph. Each observed variable is predicted using self-history, contemporaneous target-excluded subsystem context, and gated cross-variable histories. Calibrated residual scores support subsystem screening followed by candidate-first variable ranking. Experiments evaluate graph recovery and root localization on two hierarchical synthetic systems and four industrial datasets, reporting means and standard deviations over five seeds. The reported localization results are competitive, although the additional benefit of subsystem screening varies across datasets.

Critical strengths
  1. The distinction between shared subsystem variation and additional variable-level predictive effects is well motivated. The target-excluded construction provides a concrete way to operationalize this distinction, and Appendix A appropriately limits its theoretical result to population predictive support rather than claiming general causal identifiability.
  2. The evaluation covers both variable and joint time-variable localization, with five-seed variability. Reported results are promising: HiRCA leads variable Avg@10 on TEP (0.472) and joint Avg*@500 on three of four real datasets (Table 4). The real-data ablation and limitations section candidly report cases where screening hurts.
Critical weaknesses
  1. Two construction details weaken the claimed correspondence between the learned mechanism and its diagnostic/graph outputs. Equation (32) weights each cross-variable path by co-membership and the learned gate probability, but the graph score in Eq. (33) omits both factors. A nearly disabled path can therefore retain a large reported graph score; the graph-recovery results need justification against the effective predictor. Separately, Appendix D, lines 837–841, replaces the target-excluded context with the full pooled representation for effectively singleton slots. This reintroduces the current target into its own predictor and permits a self-reconstruction shortcut that can suppress the residual used for diagnosis. The frequency and empirical impact of this fallback are not reported.
  2. The experiments do not isolate the contribution of the learned hierarchy. The only component ablation retains the entire hierarchical predictor and calibration, removing only screening (Section 4.4). Its synthetic Avg@10 gains are 0.004 and 0.001; real joint Avg*@500 is unchanged or slightly worse on every dataset (Tables 5–7). A matched flat/fixed-membership predictor with the same calibration and comparable access to contemporaneous observations is needed. The synthetic generator also uses the correct partition and jointly perturbs selected variables and their parent subsystems (Section 4.1), leaving robustness to inaccurate priors or variable-only faults untested. Real-data partitions, event construction/splits, and baseline tuning are insufficiently specified to assess reproducibility and comparison fairness.
Questions for authors
  1. Does the implementation of Eq. (33) include the co-membership and gate factors omitted from the equation? How do graph-recovery results change when edges are scored by their effective gated contribution?
  2. How often is the singleton fallback used on each dataset, and do the main RCA results hold when its context is replaced by a strictly target-independent alternative?
  3. Can you provide a matched flat/fixed-membership control and the exact real-data membership maps, train/validation/calibration/test construction, and baseline tuning protocol? These would clarify whether the gains come from learned hierarchy rather than prior information, calibration, or evaluation choices.
2 comments

Thank you for identifying the mismatch between the equation and implementation. We corrected the effective graph score and context fallback and updated the experiments.

W1/Q1. Effective graph scores. The original equation and path-score implementation omitted soft co-membership and gate factors. We now multiply each path norm by co-membership and gate probability before hard support and target-wise normalization, aligning scores with effective predictive paths. Using identical model weights, candidate support and normalization, we apply the same tail-gap rule to both datasets, searching upper tails containing 2–50% of eligible edges separately for subsystem and within-subsystem graphs. Within-subsystem variable-graph results are:

DatasetScoreF1HD↓AUPRCAUROC
LinearPath0.6960.1800.6929320.887167
LinearEffective0.6780.2000.6907270.885000
NonlinearPath0.6170.2500.6206850.833167
NonlinearEffective0.6350.2250.6200140.832333

Results remain close after weighting. The main text and appendix consistently use the new effective score, with the equation and graph tables updated accordingly.

W1/Q2. Strict target exclusion. The former pooled fallback could include the current target. We replaced it with non-target pooling and zero context when peer mass is insufficient. Variable and joint time-variable localization results under this mechanism are shown below as five-seed means:

DatasetVariable Avg@10Joint Avg*@500
Linear0.9060.986
Nonlinear0.9000.983
TEP0.4710.124
Coolant0.8390.586
Hydraulics0.6800.303
Probe0.8630.628

Strong localization performance remains with strict target exclusion. These results are used throughout the main text and appendix.

W2/Q3. Hierarchical modeling. We added No-hierarchy, which removes subsystem representations, assignments, latent graphs and context, and No-context, which removes only context. Retrained models use the same residual calibration fitted on normal-operation data. Full and No-context retain two-stage inference; No-hierarchy uses direct ranking. Five-seed mean variable Avg@10 is:

DatasetFull (Two-stage)No-hierarchy (Direct)No-context (Two-stage)
Linear0.9060.8140.823
Nonlinear0.9000.7490.799
TEP0.4710.4080.475
Coolant0.8390.7990.836
Hydraulics0.6800.7450.679
Probe0.8630.8050.835

Full exceeds No-hierarchy on five datasets, supporting hierarchical prediction.

W2/Q3. Membership priors. Fixed freezes correct assignments while training other modules. Random preserves group sizes, shuffles the prior used for initialization and assignment regularization, and keeps assignments learnable. Mean Avg@10 with two-stage inference is:

DatasetFullFixedRandom
Linear0.9060.9020.800
Nonlinear0.9000.8940.767
TEP0.4710.4770.366
Coolant0.8390.8380.820
Hydraulics0.6800.6790.694
Probe0.8630.8710.763

Full exceeds Random on five datasets, while Fixed is generally close to Full, characterizing the contributions of engineering structure and assignment learning.

Joint synthetic perturbations model direct variable faults together with concurrent shared changes; root labels identify directly perturbed observed variables. The appendix now specifies the construction of Linear and Nonlinear, including generating mechanisms, fault injection, root labels, data splits and window settings.

The real-data appendix lists complete memberships, splits, event counts and training/calibration settings. TEP follows the six-group framework in MOSAIC [1], with XMEAS_37–41 additionally assigned to Stripper; causRCA groups follow the second device level of node paths. K/K_max is 6/8, 6/8, 5/7 and 3/5, with 15, 25, 22 and 34 fault events. Normalization and residual calibration are fitted on normal training data; validation controls early stopping. Baselines use their respective variable-level inputs and fitting procedures, while HiRCA additionally uses engineering assignments.

[1] Fan et al. MOSAIC, 2026. TEP preprocessing.

Dear Reviewer,

We have uploaded our rebuttal and revised manuscript, and we hope that our responses have helped clarify the concerns you raised. We would sincerely appreciate it if you could take a moment to review our responses at your convenience. If you have any remaining questions or concerns, we would be more than happy to provide further clarification.

We truly appreciate your time and consideration, and thank you once again for your valuable feedback.

Best regards, The Authors

Reviewer_zW8jOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

In industrial multivariate time series, a fault at one sensor propagates and hides which variable was hit first. HiRCA uses engineering subsystem assignments as a soft prior to learn latent subsystem states and their temporal dependencies. To predict a variable, it combines that variable's own history, a subsystem context with the target excluded, and the histories of other variables. Training uses normal-operation data only. At diagnosis, calibrated prediction residuals are aggregated to screen candidate subsystems, and the variables that survive screening are pushed up the root cause ranking.

Critical strengths
  • Feeding engineering subsystem information into both the predictive model and the diagnosis step is a sensible use of structure that industrial systems already have.
  • Latent subsystem modeling plus a target-excluded context gives a principled way to separate the variation a subsystem shares from what one variable adds on top of it.
  • The evaluation covers synthetic and real datasets and looks at graph recovery, variable localization, and joint temporal localization. Repeated runs with reported variability make the numbers easier to trust.
Critical weaknesses
  • The ablations do not isolate the components. Comparing subsystem screening against direct ranking on one already-trained model only tests the inference procedure, and even there the benefit is inconsistent across datasets. I would want ablations on the subsystem representation, the engineering prior, the target-excluded context, and residual calibration, plus a capacity control, before crediting any of the gains.
  • The real-data setup lacks the detail needed to reproduce or interpret it: engineering membership maps, data splits, the number of fault events, and training and calibration settings all need documenting. Some ranking metrics saturate across methods on the manufacturing datasets, so they cannot separate localization quality at all.
Questions for authors

Please refer to weakness 1-2.

2 comments

Thank you for the suggestions. We added ablations of the hierarchical representation, context and engineering prior, and documented the experimental setup.

W1. Hierarchical representation and context. No-hierarchy removes subsystem representations, assignments, latent graphs and context, retaining self-history and cross-variable histories. No-context removes only context. Each model is retrained with the same data, training protocol and residual calibration fitted on normal-operation data. Full and No-context retain two-stage inference; No-hierarchy uses direct ranking. Five-seed mean variable Avg@10 is:

DatasetFull (Two-stage)No-hierarchy (Direct)No-context (Two-stage)
Linear0.9060.8140.823
Nonlinear0.9000.7490.799
TEP0.4710.4080.475
Coolant0.8390.7990.836
Hydraulics0.6800.7450.679
Probe0.8630.8050.835

Full exceeds No-hierarchy on five datasets, supporting the complete hierarchical predictor. Context contributes strongly on synthetic data, with dataset-dependent gains on real data.

W1. Engineering priors. Fixed freezes correct assignments while training the other modules. Random preserves group sizes, shuffles the initialization and assignment-regularization prior, and keeps assignments learnable. Mean Avg@10 with two-stage inference is:

DatasetFullFixedRandom
Linear0.9060.9020.800
Nonlinear0.9000.8940.767
TEP0.4710.4770.366
Coolant0.8390.8380.820
Hydraulics0.6800.6790.694
Probe0.8630.8710.763

Full exceeds Random on five datasets, and Fixed is generally close to Full. These results characterize the contribution of engineering structure and the dataset-dependent behavior of assignment learning. Both ablation tables are added to Section 4.

The appendix now specifies the construction of Linear and Nonlinear, including generating mechanisms, fault injection, root labels, data splits and window settings.

W2. Real-data setup. TEP follows the six-group framework in MOSAIC [1] for 47 variables; we additionally assign XMEAS_37–41 to Stripper. causRCA groups follow the second device level of node paths, with top-level nodes in Global. K/K_max is 6/8, 6/8, 5/7 and 3/5 for TEP, Coolant, Hydraulics and Probe, with 15, 25, 22 and 34 labeled fault events. Complete memberships and split-specific window counts are added to the appendix.

TEP uses d00.dat for training, d00_te.dat for validation and d01_te.dat–d15_te.dat for testing. causRCA normal files are split by file with training fraction 0.8, and fault runs are held out for testing. Normalization and residual calibration are fitted on normal training data; validation controls early stopping. AdamW uses base/graph learning rates 0.001/0.005, up to 2000 epochs per stage and patience 100. The appendix also specifies each baseline's inputs and fitting procedure.

All three manufacturing datasets have nine variables, so AC@10 is always one. We now state this property and assess localization jointly through lower cutoffs, Avg@10 and time-variable metrics.

[1] Fan et al. MOSAIC, 2026. TEP preprocessing.

Dear Reviewer,

Thank you very much for your time and for your thoughtful comments on our manuscript. We have uploaded our rebuttal and revised manuscript in response to the concerns you raised, and we sincerely hope that our responses have adequately addressed your concerns. Please do let us know if there are any remaining concerns or if further clarification would be helpful.

Thank you again for your valuable feedback.

Best regards, The Authors

Reviewer_rX8LOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes HiRCA, a hierarchical Granger-causal model for root cause analysis in industrial multivariate time series. Using an engineering subsystem map as a soft prior, it learns latent subsystem states and a cross-subsystem temporal graph, predicts each variable from self-history, a target-excluded subsystem context, and sparse cross-variable edges, and localizes root causes by first screening subsystems with calibrated innovation scores and then ranking their variables. Experiments cover graph recovery and root-cause ranking on two synthetic systems and four real datasets.

Critical strengths

S1. The idea of exploiting engineering subsystem structure for RCA is well motivated and practically relevant: industrial plants do come with P&ID-level groupings, and separating shared subsystem variation from conditional variable-level effects is a sensible way to reduce confounded edges in flat Granger graphs. The three-branch decomposition with a target-excluded context is a clean formulation.

S2. The evaluation is reasonably broad, and the time-variable localization results on Coolant, Hydraulics and Probe are clearly better than the compared baselines.

Critical weaknesses

W1. The central claim, that the hierarchy improves RCA over a flat model, is not isolated by the experiments. The only ablation (Table 5 - 7) compares two-stage versus direct ranking using the same hierarchical predictor. There is no ablation that removes the subsystem-context branch, the latent subsystem graph, or the engineering prior to obtain a flat version of the same predictor and calibration pipeline. On the synthetic data the two-stage gain is negligible, and on real data it is mixed and hurts on Hydraulics. Since the synthetic generator is constructed exactly according to the model's hierarchical assumptions, the strong synthetic graph-recovery numbers do not establish that the hierarchy helps in practice. The real-data variable-ranking results, where HiRCA is best on only TEP and is beaten by CausalRCA on Coolant and Hydraulics and by epsilon-Diagnosis on Probe, do not settle it either. The abstract's conclusion, e.g., "top three in nearly all evaluations", may not reflect the mixed results.

W2. Missing details on the engineering prior for the real datasets, which the method depends on. The paper does not state how the subsystem membership was defined for TEP, Coolant, Hydraulics and Probe, how many subsystems were used, or how sensitive results are to this choice, even though Appendix G identifies prior mismatch as the main limitation. Together with the deterministic zero-variance results on Hydraulics, the fairness of the baseline comparison, which baselines received the same prior, if any, is hard to judge.

Questions for authors

Q1. It is better to add an ablation of the same training and calibration pipeline with (a) no subsystem-context branch and no latent subsystem graph (flat predictor), and (b) a random or uninformative membership prior, on both synthetic and real datasets, so that the contribution of the hierarchy can be assessed.

Q2. How were engineering subsystem memberships obtained for TEP and the three causRCA datasets, and what are K and K_max in each case? Were any baselines given access to this information?

Q3. Why is the standard deviation exactly zero across five seeds for HiRCA on Hydraulics in Table 4 and for several baselines on the real datasets? Does this reflect a single test event, or a deterministic pipeline?

Q4. On the synthetic data, AERCA has higher AC@1 than HiRCA. Could the authors comment on why the top-1 result is worse while AC@10 is much better, and on how the subsystem perturbations interact with the root labels?

2 comments

Thank you for the suggestions. We added hierarchy and engineering-prior ablations and updated the manuscript.

W1/Q1. Hierarchical modeling. No-hierarchy removes subsystem representations, assignments, latent graphs and context, retaining self-history and cross-variable histories. No-context removes only the context branch. Each model is retrained with the same data, training protocol and residual calibration fitted on normal-operation data. Full and No-context retain two-stage inference; No-hierarchy uses direct ranking. Five-seed mean variable Avg@10 is:

DatasetFull (Two-stage)No-hierarchy (Direct)No-context (Two-stage)
Linear0.9060.8140.823
Nonlinear0.9000.7490.799
TEP0.4710.4080.475
Coolant0.8390.7990.836
Hydraulics0.6800.7450.679
Probe0.8630.8050.835

Full exceeds No-hierarchy on five datasets, supporting the contribution of hierarchical modeling to localization. Context contributes strongly on the synthetic systems, with dataset-dependent effects on real data. Section 4 includes this table, and the abstract's performance summary is updated.

W1/W2/Q1. Engineering priors. Fixed freezes correct engineering assignments while training the remaining modules. Random preserves group sizes, shuffles the prior used for initialization and assignment regularization, and keeps assignments learnable. With two-stage inference and the same normal-operation residual calibration, mean Avg@10 is:

DatasetFullFixedRandom
Linear0.9060.9020.800
Nonlinear0.9000.8940.767
TEP0.4710.4770.366
Coolant0.8390.8380.820
Hydraulics0.6800.6790.694
Probe0.8630.8710.763

Full exceeds Random on five datasets, while Fixed is generally close to Full. Engineering groups provide useful structure, and the benefit of adapting assignments varies across datasets. This table is added to Section 4.

The appendix now specifies the construction of Linear and Nonlinear, including generating mechanisms, fault injection, root labels, data splits and window settings.

W2/Q2. Membership sources. TEP follows the six-group framework in MOSAIC [1] for 47 variables; we additionally assign XMEAS_37–41 to Stripper. causRCA groups follow the second device level of node metadata paths, with top-level nodes in Global. K/K_max is 6/8, 6/8, 5/7 and 3/5 for TEP, Coolant, Hydraulics and Probe. The appendix lists complete memberships, splits and training settings. The four RCA baselines use their respective variable-level inputs; HiRCA additionally uses engineering assignments. The new ablations examine hierarchy and membership priors separately.

Q3. Standard deviations. Coolant, Hydraulics and Probe contain 25, 22 and 34 labeled fault events, fixed across five training seeds. Ranking metrics stay identical when the relevant root ranks stay identical. With nine variables, AC@10 is always one. Hydraulics Avg@10 is 0.679509±0.000049 and joint Avg*@500 is 0.302834±0.000089; both SDs round to 0.000 at three decimals. Main-text tables consistently use three decimals, and the appendix documents event counts.

Q4. AC@1 and AC@10. Each synthetic event has 4–16 directly perturbed roots. AC@1 measures the first-ranked hit, whereas AC@10 reflects multi-root coverage. On Nonlinear, AERCA/HiRCA achieve AC@1 of 0.922/0.864 and AC@10 of 0.598/0.941, illustrating their different ranking strengths. Parent-subsystem shifts model concurrent shared process changes, and root labels identify directly perturbed observed variables. The appendix clarifies this construction.

[1] Fan et al. MOSAIC, 2026. TEP preprocessing.

Dear Reviewer,

I hope this message finds you well. I’m writing to follow up on my previous note, as the discussion period will close in less than a day. Could you kindly let us know whether our responses have addressed your concerns, or if there are any remaining points you would like us to consider? We would be happy to respond promptly. If our response has been satisfactory, a brief confirmation would be greatly appreciated.

Thank you again for your time and for reviewing our paper.

Best regards, The Authors

Comments on the paper

1 comment

Dear Area Chair and Reviewers,

We provide this concise summary of our rebuttal and revisions for your efficient review. Reviewers widely recognized the practical relevance of our engineering-subsystem-aware RCA framework, the principled target-excluded context design, the comprehensive multi-dataset evaluation with multi-seed results, and the candid discussion of limitations. To address all raised concerns thoroughly, we have supplemented key ablation experiments, corrected implementation details, and completed full experimental documentation in the revised manuscript.

Specifically, we 11 added a matched flat baseline and engineering prior ablations under a unified training pipeline, verifying consistent gains of hierarchical modeling on most datasets; 22 supplemented complete subsystem memberships, data split protocols and full experimental configurations for all datasets to ensure reproducibility; 33 corrected the effective graph scoring formula and singleton context fallback mechanism, with stable overall performance after fixes; 44 clarified the interpretation of near-zero standard deviations, AC@1/AC@10 semantics and metric saturation on small-variable datasets. Full point-by-point responses are available in our official rebuttal comments. Thank you for your time and valuable feedback.

Best regards, Authors of Paper 31280

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.