acceptodds
Under review as a conference paper at ICLR 2027

CULPRIT: Localizing Mechanism Shifts from Shared Sensors under Varying Conditions

Abstract

Root cause analysis (RCA) for industrial systems must name the faulty component from sensor data. Because sensors typically sit between components, ranking them as root causes mistakes the site of measurement for the site of failure. Our key insight is that a fault is a mechanism shift on a component-level latent variable. Under this view, a system's sensor-to-component association and its operating condition together can localize that shift from sensor data. We introduce CULPRIT, an unsupervised approach to pinpoint the root cause in industrial systems under varying conditions. It matches each fault sample to healthy samples at its operating point and projects the matched residuals onto one latent per component through the association. It then separates root causes from downstream effects along the component edges. We evaluate CULPRIT against 15 baselines on six benchmarks, where it holds the best average Top-1 accuracy of 0.52. We further show that operating-condition adjustment and the sensor-to-component association are each necessary. Matching at the operating point lifts Top-1 accuracy from 0.35 to 0.68 on N-CMAPSS, and one latent per component, instead of a sensor maximum, lifts the average Top-1 from 0.42 to 0.52. Code is available at https://anonymous.4open.science/r/culprit-DC73.

Primary area:
causal reasoning
Keywords:
root cause analysis, fault localization, multivariate time series, latent variable models, mechanism shift, identifiability, causal representation learning, unsupervised learning, industrial systems

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_VLMYOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

Root cause analysis for industrial systems identifies faulty components from sensor data. Since sensors are often positioned between components of interest, deviations in the sensor signal alone might not allow the identification of the failing component. For example, in one system a gas temperature sensor at the exit of a high-pressure turbine also measures inlet temperature of a low-pressure turbine. Under sensor measurement deviation, its value alone can't point to the failing component. A second challenge is the operating condition, since there are systems that move sensors during nominal operation, making it necessary to take into account operating condition for RCA as well. To tackle this the authors introduce CULPRIT, an unsupervised approach for this task, working under varying conditions. It matches fault samples to healthy samples at its operating point and projects the residuals onto one latent per component. It then separates root cause from downstream effects.

Input for the approach is a (multivariate) time series of the sensor readings in a system, containing nominal operation readings and readings during a fault. The duration of the fault is known. Additionally, there are condition channels that influence how nominal operation of the sensing system appears and a system model that describes causality between components and describes which sensors are influenced by multiple components and which are measuring exactly one component (local).

In a first step, for each time step of the faulty sensor reading, the authors find the nearest neighbors in the corresponding clean time series, based on the condition space, e.g. find nominal operation readings at the same Throttle & Altitude.

In step two, they model the effect of the components on the individual sensors using the system model and a linear model that maps the latent states of the components to the standardized sensor reading. To find the latent z for each component, they take all healthy residuals that are connected to the component and run a PCA, to find along with direction the joint sensor readings vary the most. The latent for a component is a standardized projection on it.

At diagnosis time (step 3), the approach calculates an anomaly score for each component, by comparing the healthy latent component distribution with the faulty one. Three quantities are calculated. (1) a displacement score that measures the mean absolute displacement of component c's latent from the latent center over the fault window. (2) Structural separation, a high displacement while a parent shows a high displacement as well is an indication for downstream effects instead of a root cause. They predict the child latent based on the parent latents which cancels out propagation effects and keeps root cause displacement. (3) Velocity score, some faults leave displacement normal but change how fast its sensor move from sample to sample, e.g. a stuck damper. For that they score based on the first-difference variance. The combined score divides separated displacement and velocity by their median over components and takes the maximum of the two scores.

The approach is broadly compared against baselines and across multiple datasets and the components are properly ablated. The results seem to be very strong and the appendix is extensive. Additionally, the authors ship a documented code base with the submission. While I was not able to fully verify the proofs (which is reflected in my confidence score), I enjoyed reading the paper and think the results are great. I especially want to highlight the structure of the paper, the clarity and the consistent notations throughout.

Critical strengths
  • Great structure and clarity (see summary).
  • Interesting alternative to deep learning based anomaly localization methods like Alora with strong empirical results (see summary).
Critical weaknesses

None

Questions for authors
  • Have you tested statistical significance for table 1? I think a rank-based evaluation across datasets would be interesting.
Reviewer_BAEaOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CULPRIT, a method for localizing faulty components from multivariate sensor data under varying operating conditions. Given a healthy reference window, a fault window, and a sensor-to-component mapping, it first matches observations by operating condition, then uses component-wise PCA to recover latent states, and finally combines structural residuals and changes in signal dynamics to rank potential root causes. The method uses closed-form fitting on healthy data rather than neural-network training. Theoretical analysis establishes conditions for successful localization, while experiments across six benchmarks show best or tied-best Top-1 accuracy against 15 baselines.

Critical strengths
  1. The paper addresses an important root-cause localization problem involving shared sensors, varying operating conditions, and downstream fault propagation. Its three-stage pipeline is interpretable and supported by theoretical analysis that explains the roles of condition matching and structural separation. The empirical evaluation is broad, covering six benchmarks and 15 baselines, with best or tied-best Top-1 performance across all benchmarks. Extensive ablation and robustness studies further clarify the contributions and limitations of individual design choices.
Critical weaknesses
  1. The analysis assumes linear sensor mixing, linear component relationships, and a fixed mixing matrix across operating conditions. And in the displacement-fault analysis, the theory requires unchanged structural coefficients. Actual faults may change coupling strengths or introduce nonlinear behavior.
  2. The “training-free” characterization is potentially misleading. CULPRIT estimates component-wise PCA projections, normalization scales, and ridge-regression coefficients from healthy data. Although these steps avoid iterative neural-network training, they still constitute data-dependent model fitting. The authors should clarify the intended meaning of “training-free” and distinguish closed-form estimation from the absence of parameter estimation.
Questions for authors
  1. Does CULPRIT remain effective when faults change component coupling coefficients or introduce nonlinear interactions, rather than shifting latent disturbances under fixed structural relationships?
  2. The details of the training procedures are not clear. How is the dataset split to get results for different stages? How are the parameters for each stage selected? How does localization performance vary with the amount of healthy reference data? The paper specifies minimum sample requirements and fitting caps, but does not establish how much healthy data is needed for reliable matching, latent recovery, and structural separation.
Reviewer_99CQOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CULPRIT, an algorithm that identifies the root cause of anomalies by isolating the faulty component rather than merely flagging an anomalous sensor. A sensor may be shared across multiple components. It first matches samples with similar external conditions to the anomalous sample. It then computes mean and standard deviations from these similar looking healthy samples and uses them to compute condition-adjusted sensor residuals. It then estimates each component's state using principal component analysis (PCA) on its associated sensor residuals. To distinguish root causes from downstream effects, it predicts each component's state from its parents using ridge regression and measures the remaining deviation. It also computes a velocity score to detect changes in sensor dynamics, such as a sensor becoming frozen. CULPRIT combines the displacement and velocity scores to rank components as candidate root causes. The paper provides theoretical guarantees under certain conditions and evaluates the method through benchmark comparisons, ablations, and robustness experiments.

Critical strengths
  1. The empirical analysis seems solid with a variety of datasets and baselines.
  2. The problem statement is relevant because it is indeed significant to not just isolate the faulty sensor but also pin down the component that caused it.
  3. What is interesting in this paper is that they assume that the component values are latent and still find the root causes from within them.
Critical weaknesses

1. Writing I have significant concerns with the writing of the paper. The underlying idea appears relatively simple, but the writing makes it difficult to understand. The dense notation and lack of motivation for several equations make the methodology unnecessarily complicated. I had to read the paper multiple times to understand it. The authors should simplify the writing and provide more intuition.

2. Why does the structural equation fef_e change between normal and anomalous regimes. From what I understand, the mechanism ff represents a physical relationship governed by the laws of physics? Please explain this assumption as to what changes ff during an anomaly using a simple, intuitive example.

3. In Eq. 3, the authors appear to normalise u~t\tilde{u}_t using statistics computed over all healthy samples. Would it be more appropriate to compute the mean and variance using only healthy samples whose external conditions also match those of the anomalous sample? Please explain the rationale behind this choice.

4. What is the main takeaway from Proposition 3.1, and why is the bound useful in practice? Please explain its intuition and, where possible, connect it to the experimental results.

5. Why do the authors assume a linear relationship between ztcz_t^c and its parents in Eq. 5? How would the method perform if this relationship was non-linear?

6. The rationale behind several remaining equations is unclear. Why did the authors choose these particular formulations over alternatives? Why was projection to the first principal direction necessary? What other alternatives did the authors consider. Currently the section reads rather procedural, and lacks motivation and the intuition behind several choices made in the paper.

  1. Include an algorithm atleast in appendix that clearly illustrates the CULPRIT procedure.

Experiments

  • The paper shows that combining sensors into a component score works better than taking the largest sensor score: average Top-1 accuracy improves from 0.42 to 0.52. However, PCA focuses on what varies most during healthy operation, which may not be what changes during a fault. Please test cases where the fault affects a direction that PCA largely ignores, especially when sensors share noise or measure several components. This would show when PCA helps and when it hides the fault.

  • Can normal changes in operating conditions mislead the velocity score? Matching operating conditions improves Top-1 accuracy from 0.35 to 0.68 on N-CMAPSS. However, the velocity score does not use this matching. A healthy sensor may start changing faster simply because the system moves from steady to rapidly changing operation. Please test this situation, both with and without a fault, and report whether the velocity score incorrectly pushes healthy components above the true root cause.

  • I appreciate that the experiments are vast. However, the verbose explanation of the experiments need significant improvements. Please provide the key takeaways from each experiment and try to enhance the explaination for why CULPRIT achieved the best performance against the competing baselines. Do more controlled experiemnts to relate the theory that you have in the paper to practice. If current datasets are not fit fot this, it might be good to come up with synthetic toy cases where the theory can be illustrated.

  1. Robust Root Cause Diagnosis using In-Distribution Interventions (IDI, ICLR 2025) is a recent ICLR 25 paper and should be considered for comparison.

Overall, the writing of the paper needs to be calibrated so that it only feels as complicated as the method deserves it to be. The flow of the paper in its current version is not upto the mark.

Questions for authors

Please see weakness above.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.