acceptodds
Under review as a conference paper at ICLR 2027

Anomaly Detection in Spectral Data Streams: The Degassing Cycle Scenarios

Abstract

Monitoring the elemental composition of materials is of great importance in many scientific and industrial applications. Optical Emission Spectroscopy (OES) is a technique that enables this through the analysis of electromagnetic radiation emitted by a material at different wavelengths, producing spectral data streams. Unlike previous studies on unsupervised anomaly detection in spectral data streams, which mainly focused on air leaks under constant pressure conditions in a vacuum chamber, we collected simulated real-world data within the BV5 Benchmark. This benchmark involves degassing cycles with argon leaks, air leaks, or both. The main objective in BV5 is to distinguish abnormal from normal degassing cycles in real time. However, following previous benchmark settings, we also collected simulated real-world datasets for argon leaks under constant pressure conditions and for temporal anomaly scenarios involving either argon or air leaks. We then provide a progressive evaluation of online anomaly detectors, using only spectral data, across benchmarks. Our results show that OBKNN (TNone), IFASD, OIF, and KitNet consistently outperform competing approaches in spectral anomaly detection. Furthermore, KitNet and OBKNN (TZNorm) exhibit a progressive decline in performance compared to the original benchmark. Lastly, a comparative analysis of methods in degassing cycle scenarios highlights that spectral data can be combined with additional features to enhance anomaly detection performance. To this end, we utilized the collected spectral data and additional features in BV5 to evaluate state-of-the-art online anomaly detection methods. We introduce the Subset Framework (SF), which complements spectral data with an additional feature. We show that incorporating data related to vacuum conditions using SF, such as the electrical current from the OES plasma, improves the detection performance of many online methods. Experiments demonstrate that SF improves the effectiveness of the majority of online detectors, notably SFOBKNN (TNone), SFOBKNN (TZNorm), and SFIFASD, with increases of 0.384, 0.323, and 0.217 in VUS-PR, respectively.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
online anomaly detection, spectral data streams, optical emission spectroscopy, degassing cycles

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_kQhXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates unsupervised online anomaly detection in spectral data streams under varying vacuum conditions. The authors extend existing benchmarks by introducing BV5, which includes new degassing cycle scenarios, different leak types, and additional sensor measurements. Through a comparative evaluation of existing online anomaly detectors, the study highlights their varying robustness under complex operating conditions. Additionally, the authors propose a Subset Framework (SF) that filters unreliable spectral observations using auxiliary sensor information, demonstrating its potential to improve anomaly detection performance.

Critical strengths
  1. Practically relevant benchmark extension. The paper introduces newly collected spectral datasets covering degassing cycles, different leak types, and varying operating conditions. These scenarios extend previous benchmarks and provide a useful resource for evaluating anomaly detection methods under more realistic industrial conditions.
Critical weaknesses
  1. Limited novelty and insights beyond existing benchmarks. The paper primarily extends the spectral anomaly detection benchmarks established in previous work, particularly Varela et al. (AAAI 2026). While the new scenarios are practically relevant, the authors do not sufficiently demonstrate what fundamentally new challenges or insights they introduce. The cumulative comparisons mainly reveal changes in algorithm rankings, without systematically explaining the underlying failure mechanisms or identifying new research directions. Consequently, the incremental contribution may be insufficient for a standalone benchmark study.

  2. Insufficiently controlled evaluation of the proposed Subset Framework. The experiments comparing SF with the original detectors use different hyperparameter configurations, including substantially different window sizes and ensemble sizes. Therefore, the reported performance improvements cannot be reliably attributed to the proposed filtering strategy alone. Controlled comparisons under matched hyperparameters are necessary to establish the actual contribution of SF.

  3. Unclear effects of data filtering on detection performance. SF excludes a substantial proportion of spectral observations and directly assigns zero anomaly scores to excluded instances. This raises concerns about whether the improvements reflect better anomaly detection or simply a restriction of the effective detection space. The paper lacks comparisons with simple rule-based or random filtering strategies, as well as analyses of anomaly coverage and missed detections. These omissions weaken the evidence supporting SF's effectiveness.

  4. Incomplete benchmark coverage and evaluation. Although the paper evaluates several traditional online anomaly detectors, the coverage of recent deep learning approaches remains limited, with KitNet being the only evaluated deep learning method. Relevant approaches such as ARCUS are discussed but not included. Furthermore, despite emphasizing real-time anomaly detection, the evaluation primarily focuses on VUS-PR without systematically assessing computational efficiency, memory consumption, or detection latency. This limits the comprehensiveness and practical value of the benchmark.

  5. Limited dataset diversity and reproducibility. The newly introduced datasets are collected from a single controlled experimental setup, with fixed leak intensity and relatively few abnormal events in several scenarios. Consequently, their representativeness across different industrial systems and operating conditions remains unclear.

Questions for authors
  1. Reliability of anomaly annotations. According to Appendix B, anomaly labels are primarily determined by the timing of leak valve operations. However, the actual spectral response may not align precisely with valve opening and closing events. How do the authors account for potential response delays, and how is the accuracy of the ground-truth annotations validated?

  2. Hyperparameter tuning protocol. The hyperparameters are selected using AUC-PR on only two labeled datasets, including one spectral dataset from BV4. Could the authors clarify how this supervised tuning procedure aligns with the intended unsupervised setting? How sensitive are the conclusions to the choice of tuning datasets, particularly given the different operating conditions in BV5?

Reviewer_TCt4OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces BV5, a benchmark of optical-emission spectral streams collected from controlled vacuum experiments, and proposes the Subset Framework (SF) for online anomaly detection. SF uses an expert-defined plasma-current interval to select spectra for detector processing and assigns zero anomaly scores outside that interval. The paper compares existing detectors across three benchmarks and reports substantial VUS-PR improvements for OBKNN and IFASD on BV5 degassing

Critical strengths

• BV5 incorporates degassing cycles, different leak compounds, and humidity conditions, providing a useful extension to existing spectral benchmarks focused on stable vacuum conditions. • The experiments cover distance-based, tree-based, and neural detectors across three benchmarks, with the appendices documenting data collection procedures and hyperparameter settings.

Critical weaknesses

• The SF comparison changes OBKNN’s window fraction from 0.20 to 0.02 and ensemble size from 30 to 240, making it difficult to attribute the reported gains to SF alone. • Section 5.3 compares SF with detectors using only auxiliary features, but omits a simple current-range indicator baseline and detectors using spectra and auxiliary features jointly, leaving SF’s advantage over these alternatives unclear. • The paper does not evaluate sensitivity to the expert-defined 7.75–8.25 mA interval or its transferability across acquisition setups, leaving the robustness of the filtering rule insufficiently established. • SF assigns zero scores to excluded observations, including some labeled anomalies in Table 2, but event-level recall, false-alarm rates, and detection delays are not reported to assess the practical consequences.

Questions for authors

• How do the gains change when SF and the original detectors use identical hyperparameters and are evaluated on the same timestamps after the initial training phase? • How does SF compare with a simple current-range indicator baseline and detectors using spectra and auxiliary features jointly under the same tuning budget? • Was the current interval determined independently of evaluation labels, and how sensitive are the results to changes in its bounds? • What event-level recall, false-alarm rates, and detection delays are obtained when all observations assigned zero scores by SF remain in the evaluation?

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

This paper studies unsupervised online anomaly detection on spectral data streams obtained from Optical Emission Spectroscopy (OES) in vacuum systems. The authors point out that prior benchmarks primarily focused on static, constant-pressure chambers with simple air leaks. To reflect more realistic operating conditions, the paper introduces the BV5 benchmark, which incorporates dynamic degassing cycles (transitions from atmospheric pressure to vacuum), alternative leak compounds (Argon), and 54 auxiliary physical sensor features. The paper presents a progressive evaluation of several streaming anomaly detectors (e.g., OBKNN, IFASD, OIF, KitNet) across benchmark tiers. To deal with sensor stabilization noise during degassing cycles, the authors also describe a "Subset Framework" (SF) that filters spectral instances using a domain-specified range on the plasma regulation current.

Critical strengths
  1. Extending spectral stream anomaly detection to include degassing cycles and different chemical leaks (Argon alongside air) addresses a clear gap in existing literature. Static pressure setups overlook the sensor noise and drift that typically occur during chamber pump-down.
  2. Evaluating algorithms sequentially across three tiers (Original benchmark →\rightarrow BV4 →\rightarrow BV5) offers a helpful view of how methods hold up as the operating environments become progressively more non-stationary.
  3. The analysis correctly identifies that physical OES sensors produce noisy, uncalibrated spectra right after chamber connection/disconnection, and demonstrates how this specific operational phase corrupts online detectors that lack robust filtering.
Critical weaknesses
  1. The paper frames SF as an algorithmic contribution, but Section 3.3 (Eq. 1) shows it is a straightforward interval filter: L0={S(t,⋅):x(t)∈[lc,uc]}L_0 = \{S(t, \cdot) : x(t) \in [l_c, u_c]\}, with [lc,uc][l_c, u_c] hardcoded to [7.75,8.25] mA[7.75, 8.25]\text{ mA}. Filtering out data points based on a static expert rule and assigning a default score of 0 is standard operational preprocessing, so presenting it as an independent learning framework feels somewhat inflated.
  2. According to Table 2, applying the current filter to the five non-degassing scenarios (SAR1–SAR3, TNH, TAR) filters out 100% of the anomalous instances, reducing the effective anomalous duration to 0 seconds. In real-world deployments, an anomaly detection system usually cannot know a priori whether a cycle is normal, degassing, or constant unless an external mode-switch exists. Because SF breaks completely on stable operating conditions, its practical utility as an autonomous stream preprocessor is questionable.
  3. In Table 6, p_window_size was tuned to 0.20 for key methods (OBKNN, IFASD, OIF) on the full streams. However, for SF in Table 7, p_window_size was set to a fixed default of 0.02. In addition, Table 2 shows that SF drastically shrinks the number of evaluated time steps (e.g., down to 271–289 steps for HNH–LAR). It is unclear whether the performance gain in Figure 2 comes from the physical filtering itself or from the combination of a much shorter stream and a different sliding window ratio.
  4. In Table 8, the standard deviations across the seven degassing scenarios are very large (e.g., 0.410±0.300.410 \pm 0.30 for IFASD, 0.576±0.190.576 \pm 0.19 for OBKNN, 0.682±0.400.682 \pm 0.40 in Table 9). With standard deviations often rivaling the mean values, the relative rankings among methods appear somewhat unstable across different runs and scenarios.
Questions for authors
  1. In Section 3.3, the text states that instances outside [lc,uc][l_c, u_c] receive a default anomaly score of 0. When calculating VUS-PR in Figure 2 and Table 8, are these zero-scored points included in the full 4200-length sequence, or is the metric evaluated strictly over the filtered subset L0L_0? If evaluated over the entire sequence, assigning zero to long non-anomalous segments could inflate precision simply by eliminating false alarms during unmonitored periods.
  2. The current range [7.75,8.25] mA[7.75, 8.25]\text{ mA} was chosen with domain experts. How sensitive are the detection results to this specific window? What happens if the plasma current shifts slightly over time due to electrode wear or chamber degradation?
  3. Given that SF leaves zero anomaly instances in constant-pressure scenarios (Table 2), how should a practitioner deploy this pipeline in an automated environment where the system transitions between degassing and steady-state operation? Is an upstream regime-classification model required to turn SF on and off?
  4. Why was p_window_size changed from 0.20 (in Table 6) to 0.02 (in Table 7) when evaluating the SF models? Did the authors test whether running the baseline models (without SF) with p_window_size = 0.02 changes the comparison in Figure 2?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.