- 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.