{"id":"fb354930-c414-4948-ad48-50fe9370971d","arxiv_id":"2506.05958","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A PCA-DBSCAN-SHAP pipeline reduces 15 months of AQUAVALL WWTP sensor data to a small set of explained operation modes, with held-out samples mostly matching regular operation and documented incidents.","lead":"The authors combine PCA, DBSCAN clustering, and SHAP explanations on 15 months of monitoring data from the AQUAVALL wastewater treatment plant to reduce 3,231 sensor variables into a few named operation modes, such as regular operation, heavy rain, and shutdowns. The paper shows how plant managers could view the current operating state as a single point in a low-dimensional, explained space.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SHAP explanations in Tables 4–9 are not well-defined: Eq. (8) requires a predictive model f, but no f is specified for PCA or DBSCAN cluster labels, so the central 'explained operation modes' claim is unsupported.","rationale":"The reader's weakest assumption identifies the same issue: SHAP is applied to outputs for which no predictive function is specified. My independent reading of Sections 2.3, 3.1, 4.1.1 and Tables 4–9 confirms that Eq. (8)'s f is absent. This is load-bearing because the contribution is specifically 'explained operation modes' and the expert validation is anecdotal. I do not see an internal inconsistency in the PCA/DBSCAN pipeline itself; the discovered cluster labels may be useful. The concern is addressable by specifying f or replacing SHAP with a well-defined attribution method, so the verdict remains conditional. The MinPoints=1 issue is secondary and also worth reporting as a sensitivity check.","tokens_in":698,"tokens_out":1079,"duration_ms":40205,"concrete_test":"Request or independently reconstruct the exact f used in Eq. (8), then recompute Table 4 for Ts = 12h as follows: train a surrogate classifier (e.g., random forest or logistic regression) on the 453 standardized original variables to predict the 15 DBSCAN cluster labels, and compute SHAP values for that classifier. Compare the top-10 variables per cluster with Table 4. If the top features are not largely reproduced, the reported SHAP explanations are artifacts of an unspecified f and the explanatory claim fails; if they are reproduced, the concern is resolved by documenting f.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The manuscript's explainability claim rests on Eq. (8), which defines SHAP values only for a specified predictive model f. Sections 2.3 and 3.1 state that KernelSHAP is used to explain PCA and DBSCAN outputs, and Tables 4–9 report per-mode variable importances, but the function f is never defined. For principal components, f could plausibly be a projection or reconstruction loss, but that choice is not stated. For DBSCAN cluster labels, f(S) must be a real-valued model prediction over feature subsets; a clustering assignment is not such a function, and SHAP cannot be applied to it directly unless a surrogate classifier is trained to reproduce cluster membership. No such surrogate is described. Therefore the SHAP attributions are not well-defined, and the central claim that each operation mode is explained by the plant variables is unsupported. This is a correctness risk, not merely a presentation gap: different choices of f would produce different explanations.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an XAI-based unsupervised methodology to identify and explain operation modes of a wastewater treatment plant (WWTP) from SCADA logs. The pipeline consists of data cleaning, resampling at 12-hour and 24-hour intervals, standardization, PCA-based dimensionality reduction to seven principal components, DBSCAN clustering with MinPoints=1 and elbow-selected epsilon values, and KernelSHAP to attribute mode membership to original plant variables. The method is applied to 12 months of AQUAVALL WWTP data for the offline stage and 3 months for an online stage, reporting 15 operation modes for the 12-hour resampling and 13 for the 24-hour resampling, with qualitative expert validation linking several modes to shutdowns, rain episodes, and connection failures.","tokens_in":16188,"tokens_out":4346,"duration_ms":41920,"significance":"If the pipeline is valid, the paper offers a practical tool for plant operators: it compresses thousands of monitored variables into a small set of operation modes and provides per-mode variable attributions, with an online stage that could support real-time monitoring. Strengths of the work include the use of a real industrial dataset, the explicit offline/online split with three months of unseen data, and the effort to validate clusters against plant records and expert knowledge. However, the central explainability step is not formally well-defined, several parameter choices are justified post hoc on the same data used for clustering, and the validation is largely qualitative; as a result, the scientific claims are currently not established at the level expected for a journal publication.","major_comments":[{"comment":"Equation (8) defines Shapley values only for a specified predictive model f, but the manuscript never defines f for the PCA scores or the DBSCAN cluster labels. The text states that KernelSHAP is used to explain PCA and clustering outputs, yet for a clustering assignment no real-valued prediction function f(S) is given; without such a function the SHAP values in Tables 4–9 are not well-defined. Please specify the model or surrogate (e.g., a classifier trained to reproduce cluster membership, or a reconstruction loss for PCA) whose predictions are explained, and re-derive or recompute the attributions accordingly.","section":"§2.3, Eq. (8); §3.1; §4.2.1"},{"comment":"With MinPoints=1, every point satisfies the core-point condition in Eq. (6), so DBSCAN degenerates into connected components of the ε-neighborhood graph and the density-based distinction between core, border, and noise points is lost. The k-distance elbow rule is then applied with k=MinPoints=1, which reduces to nearest-neighbor distances rather than a density estimate. Since the choices ε=1.3 (12 h) and ε=1.1 (24 h) are justified only by this elbow heuristic on the training data, please provide a sensitivity analysis over ε and MinPoints and state explicitly what density property, if any, the resulting clusters are intended to preserve.","section":"§2.2, §4.1.1, §4.2"},{"comment":"The PCA projection and the DBSCAN parameters are fitted on the same 12 months that are then clustered, so the discovered modes are partly in-sample summaries. The online stage provides some independent evidence, but the evaluation is qualitative only, matching three and two new modes to plant records. Please add quantitative validation, such as cluster stability measures, silhouette or Davies–Bouldin indices, and out-of-sample assignment error on the 3-month held-out data, and report how sensitive the number and composition of modes are to the PCA rank and the ε choices.","section":"§4.2, §4.2.1"},{"comment":"The expert validation is not systematic: statements such as \"seem to be\" or \"match with\" are not supported by a defined matching protocol, and only a subset of the discovered modes is explicitly linked to recorded events. Please specify the criteria used to match modes to expert records, report the number of modes checked, and ideally use a blinded or pre-registered comparison to support the claim that the clusters reflect real operation modes.","section":"§4.2.1, Tables 4–9"}],"minor_comments":[{"comment":"There are several typographical errors that should be corrected: \"datset\" near Eq. (5), \"outlayers\" in Section 3.1, \"matching learning\" in Section 4.2, \"variable space space\" in Section 4.2.1, and \"apperance\" in the online 12-hour case.","section":"General"},{"comment":"The notation f(S) is overloaded: it denotes a model prediction in Eq. (8), but later the same symbol is used for explaining PCA and cluster outputs. Please use separate symbols, for example h for the function being explained, to avoid ambiguity.","section":"§2.3, Eq. (8)"},{"comment":"The sentence describing the maximum flow rate of the WWTP is incomplete: \"able to manage a maximum flow rate up to m3 seg\" lacks the numerical value and unit. Please provide the missing quantity.","section":"§4"},{"comment":"The reference list contains formatting and typographical errors, such as \"Phillosphical Transaction\" in [16], inconsistent journal-name styling in [1], and the very long reference [11]; please unify the reference style and correct the typos.","section":"References"},{"comment":"The figures plot only three principal components at a time, without explaining why those particular triples were chosen or whether the remaining components affect the visual interpretation of mode separation. Please add a brief explanation of the component-selection rule for the plots.","section":"§4.2.1, Figures 6, 8, 10, 12"}],"recommendation":"major_revision","confidential_remarks":"This is an engineering case study with a real dataset, and the expert-validation effort is commendable. The central issue is that the SHAP step must be made well-defined before the results can be accepted as explainable; this is likely fixable by specifying a surrogate model and adding quantitative validation, rather than by rewriting the whole paper. The authors may also wish to position the contribution as an engineering demonstration rather than a new methodological result, which would better match the actual content."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper is a serious applied case study with a real 12+3 month dataset from a working WWTP, but its central explainability claim is built on a SHAP application that is never defined. The authors never specify the predictive function f whose Shapley values are computed, so the numbers in Tables 4–9 are not tied to any stated model. That's a load-bearing gap, not a cosmetic one.\n\nWhat's genuinely useful: the preprocessing story is concrete (3,231 raw variables down to 453, missing-value thresholds stated), the two resampling windows (12h and 24h) give different mode structures, and the online stage with three unseen months tests the pipeline on data not used to fit PCA/DBSCAN. The expert-knowledge check—matching modes to recorded shutdowns and storm events—is the right kind of validation for this domain and gives the clusters real-world plausibility.\n\nThe soft spots are real but limited in number. (1) Eq. (8) requires a model f; for PCA and DBSCAN outputs none is specified, so the SHAP attributions are undefined. The authors should either state a surrogate model (e.g., a classifier trained on cluster labels) or drop the SHAP label and use a well-defined attribution for clustering. (2) MinPoints=1 turns DBSCAN into epsilon-neighborhood connected components; that is a legitimate choice given the goal of catching rare events, but the paper should say so explicitly and discuss sensitivity to epsilon. (3) The elbow choices for epsilon and the number of PCs are post hoc; quantitative stability (e.g., silhouette or resampling) would strengthen the mode count.\n\nNone of this kills the case study. The pipeline is a composition of standard tools and the novelty is local—the AQUAVALL modes themselves—but for a systems-engineering venue that is enough if the explanation step is made rigorous. The paper deserves a serious referee; it needs a substantial revision on the SHAP definition before acceptance. I'd bring it to a reading group for the data-cleaning and validation discussion, but I would not cite it until the explainability step is fixed.","headline":"A serious applied case study with real plant data, but the SHAP explanation step is never defined for PCA/DBSCAN outputs, making the central explainability claim unsupported until fixed.","tokens_in":16682,"tokens_out":1482,"would_cite":false,"duration_ms":15422,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A wastewater treatment plant's high-dimensional monitoring history can be compressed into a small set of explained operation modes, giving managers a low-dimensional view of the plant's current state.","keywords":["WWTP","operation modes","eXplainable Artificial Intelligence","SHAP","DBSCAN","PCA","knowledge extraction","dimensionality reduction"],"falsifier":"Re-run the SHAP explanation step with two different explicit surrogate models, for example a classifier that predicts DBSCAN cluster labels from the original variables and a regressor that predicts each principal-component score, and compare the top-ranked variables for each mode; if materially different rankings emerge, the mode explanations are not uniquely determined by the data.","tokens_in":19,"feed_emoji":"💧","tokens_out":5581,"duration_ms":107996,"temperature":0.7,"pith_summary":"This paper proposes a methodology for turning years of high-dimensional wastewater treatment plant (WWTP) monitoring data into a handful of explainable operation modes. Using 15 months of real SCADA data from the AQUAVALL WWTP, it reduces 3,231 monitored variables to 453 usable ones, then to 7 principal components, clusters the data with DBSCAN, and explains each cluster with SHAP values. The result is a low-dimensional space in which each new sample can be located at a glance, with 15 operation modes found at 12-hour resampling and 13 at 24-hour resampling. The detected modes match plant records and expert knowledge—regular operation, maintenance shutdowns, and heavy-rain episodes—so the unsupervised clusters are claimed to reflect real plant states rather than arbitrary groupings.","feed_headline":"15 explained modes summarize a year of WWTP sensor data","feed_subtitle":"Managers can place each new sample in a low-dimensional space and see which plant variables drive the current operation.","key_machinery":"The pipeline is the machinery: expert-guided cleaning and resampling (12h and 24h) reduce the raw SCADA stream, PCA projects the remaining 453 standardized variables onto seven principal components that retain most of the variance, DBSCAN with MinPoints=1 and an epsilon chosen from the k-distance graph discovers dense regions as operation modes, and KernelSHAP attributes each mode and principal component to the original plant variables and facility units. The work of the pipeline is to turn an unlabeled, high-dimensional time series into a labeled, low-dimensional map of plant states expressed in domain terms.","core_discovery":"The central claim is that a large historical monitoring database can be summarized as a small set of explained operation modes, each represented by a point in a reduced variable space, and that this summary is usable in real time. The paper demonstrates this on the AQUAVALL WWTP: after removing irrelevant and low-quality variables, PCA keeps seven components that preserve about 58–60% of the variance; DBSCAN (with MinPoints=1 so that rare events are kept as modes rather than noise) finds 15 modes at 12-hour resampling and 13 at 24-hour resampling; SHAP attributes each mode to the most influential plant variables and units. The authors validate the modes against the plant's logged events and expert knowledge, identifying regular operation, maintenance and power-line shutdowns, and rain-driven episodes among the found modes.","pith_inferences":["Because the paper does not define the prediction function behind its SHAP values, a direct test is to re-run the attributions with an explicit surrogate model, such as a cluster-label classifier or a PC-score regressor, and see whether the mode rankings survive.","The same pipeline should transfer to other wastewater or drinking-water utilities that keep SCADA histories, provided the expert-knowledge validation step is repeated locally.","A practical extension is to turn the reduced-space distance to the regular operation mode into a live anomaly score, alerting managers when a new sample falls far from any known mode."],"forward_implications":["Plant managers can see the current operating state as a single point in a seven-dimensional space and its distance to known modes, without reading thousands of SCADA variables.","New or unknown samples that fall outside the known modes can be flagged immediately and explained in terms of the variables and units involved, as was done for the power-line and connection-failure online samples.","The 12-hour resampling preserves short-term fluctuations and transient states, while the 24-hour resampling yields more stable modes but can hide rapid changes, so the choice of resampling window changes which modes are visible.","When the model is rebuilt with 15 months of data, samples that were previously exceptional can be absorbed into the regular operation mode, showing that incremental updating tracks the plant's evolution."],"supporting_citations":[{"why":"Defines DBSCAN, the density-based clustering algorithm used to discover operation modes.","marker":"[20]"},{"why":"Introduces SHAP and KernelSHAP, the XAI method used to attribute each mode to plant variables.","marker":"[22]"},{"why":"Supplies the PCA formulation the paper uses to reduce 453 variables to seven components.","marker":"[17]"},{"why":"Provides the k-distance graph method used to tune the DBSCAN epsilon parameter.","marker":"[21]"},{"why":"Recent PCA overview supporting the dimensionality-reduction step.","marker":"[25]"}],"fun_headline_variants":["XAI reveals 15 explainable operation modes in WWTP","15 explained modes capture a year of WWTP behavior","Unsupervised XAI finds 15 modes in real WWTP data","WWTP operation distilled into 15 explained modes via XAI","AI explains 15 operation modes from WWTP sensor history"],"cache_read_input_tokens":18944,"weakest_assumption_plain":"The explanation step assumes Shapley-value attributions (SHAP) can meaningfully explain principal components and clusters, but the paper never specifies the predictive model whose output those values attribute, so the variable rankings rest on an unstated surrogate.","fun_headline_variants_meta":{"raw":{"variants":["XAI reveals 15 explainable operation modes in WWTP","15 explained modes capture a year of WWTP behavior","Unsupervised XAI finds 15 modes in real WWTP data","WWTP operation distilled into 15 explained modes via XAI","AI explains 15 operation modes from WWTP sensor history"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00026,"raw_usage":{"total_tokens":1626,"prompt_tokens":1020,"completion_tokens":606,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":636,"completion_tokens_details":{"reasoning_tokens":521}},"tokens_in":636,"tokens_out":606,"duration_ms":5371,"temperature":1.0,"reasoning_tokens":521,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:12:46.218003+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the SHAP explanation step with two different explicit surrogate models, for example a classifier that predicts DBSCAN cluster labels from the original variables and a regressor that predicts each principal-component score, and compare the top-ranked variables for each mode; if materially different rankings emerge, the mode explanations are not uniquely determined by the data.","supporting_citations":[{"cited_title":"Ester, H.-P","cited_arxiv_id":null,"evidence_quote":"Defines DBSCAN, the density-based clustering algorithm used to discover operation modes."},{"cited_title":"Greenacre, P","cited_arxiv_id":null,"evidence_quote":"Recent PCA overview supporting the dimensionality-reduction step."}],"review_version":1}