acceptodds
Under review as a conference paper at ICLR 2027

FactoryNet: A Large-Scale Dataset toward Industrial Time-Series Foundation Models

Abstract

Industrial machines are actuated systems, yet most open time-series corpora record sensor outcomes alone; the few that do log commanded signals expose them in vendor-specific forms that differ from machine to machine, so commanded intent and measured response cannot be compared across embodiments. We introduce FactoryNet, a pretraining corpus for industrial time-series: 113M datapoints across 57k end-to-end task executions (15k real, 42k synthetic) on seven embodiments and four tasks, with 27 annotated anomaly types alongside healthy baselines and counterfactual pairs. Every signal in every source is mapped into a single control-theoretic schema, Setpoint, Effort, Feedback, Context (S-E-F-C), which expresses any actuated system in a common representational frame and makes commanded-versus-realized dynamics readable as explicit prediction residuals. Using the corpus, we find that a task-phase model trained on simulation, with no real phase labels, recovers roughly two thirds of the real-data ceiling on both robots, and that transfer holds within a robot family but rarely across manufacturers. Fault signatures behave differently: an uncalibrated simulator gives no detectable fault signal on either robot, and aligning the feature distributions does not repair it. What separation there is comes from Effort: motor current distinguishes faults where measured tool motion is close to chance, and calibrating the simulator to reproduce that one channel recovers a payload fault on the real robot. We read this as simulation capturing the stable structure of a task while fault behaviour stays diverse and fault-type specific. We release FactoryNet as a growing, multi-embodiment substrate for industrial foundation models.

Primary area:
datasets and benchmarks
Keywords:
industrial time-series, cross-embodiment transfer, foundation models, anomaly detection, robotics datasets, forward dynamics, sim-to-real

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_vUt4OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors present factorynet, a multi-embodiment industrial time-series corpus spanning 56,591 end-to-end task executions (simulated & real), 113M logged timesteps, 7 embodiments, 4 tasks, 27 annotated anomaly types with healthy baselines and counterfactual pairs. An embodiment is a distinct machine platform (e.g. a different robot arm). The benefit of this collection is, that it not only contains the resulting physical state signals of machines, but rather the full control loop, including targets and actuation effort, properly labeled and structured. This enables one to structurally separate cause and effect and to work with the commanded-versus-realized dynamics measured during recording.

In the presented collection, every signal is mapped to a machine-agnostic control-theoretic schema: setpoint (commanded intent, e.g., target joint positions), effort (actuation, e.g., motor current or torque), feedback (measured outcome, e.g., actual position), context (environment, e.g., payload mass). Additionally, they present two studies, (1) they investigate if the schema supports anomaly detection, and (2), they investigate transfer of a monitoring model between datasets (sim to real and robot to robot).

(1) instead of reconstruction-based anomaly detection, they fit an MLP to map setpoints to efforts and compute the per-episode mae between the estimate and real effort as their anomaly score. Their approach lands mid-pack. Switching to a mapping from setpoint to feedback ends up being the strongest approach in their evaluation. They argue that their schema can serve as a reduced search space to select input and target channels for anomaly detection.

(2) They test two tasks: phase identification (which stage of the pick-and-place cycle is running) and binary fault detection. The presented results are interesting and highlight the benefits of having real and simulation data in the collection.

Overall the paper is in a good shape and is a timely and important contribution for deep learning research in the manufacturing sector. The work took notable standardization efforts and presents a senseful schema for future dataset releases that contain closed-loop episodes of manufacturing machines. The huggingface page is well documented. Having that many fault episodes in the collection is indeed interesting for anomaly detection and robustness research, but the ratio seems a bit skewed towards faulty data (72/28). To improve the presentation I recommend to add a brief list of contributions to the end of the introduction and to add a visualization of one episode of the dataset to the main paper. The anomaly detection experiments could be improved with more recent baselines, more details and metrics. Since the main contribution is the presented collection and standardization effort, I recommend acceptance.

Critical strengths
  • The curated dataset and structuring is exactly the standardization effort the field needs for the work on foundation models and more general assessment of algorithmic improvements across use cases and embodiments.
Critical weaknesses

None

Questions for authors
  • What exactly was the bootstrapping setup that you mentioned in your paper?
  • Could you provide additional details about the anomaly detection evaluation setup? (e.g. scoring, score aggregation, additional metrics)
  • You hypothesize that existing TSFMs can't disentangle static payload biases from dynamic behavior, have you performed experiments on that?
Reviewer_e1z1OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2(+1)
Soundness
1
Read review
Paper summary

This paper introduces FactoryNet, an industrial time-series dataset organized around a unified control-theoretic schema: Setpoint, Effort, Feedback, and Context (S-E-F-C). The schema places heterogeneous actuated systems within a common representational framework and enables deviations between commanded and realized behavior to be examined through prediction residuals. However, the presentation needs clarification. In particular, the paper should distinguish healthy-only anomaly detection from supervised fault detection. Although the former is evaluated in 4.1, the transfer experiments in 4.3 primarily address binary fault detection using labeled fault examples. The baseline selection also requires clearer justification: while the single-machine evaluation includes established anomaly-detection methods, the transfer experiments rely mainly on general-purpose time-series classifiers. Important experimental details, including the simulator calibration procedure and the real data used for calibration, should be presented more clearly in the main text. Furthermore, the main structured baseline underperforms several established methods and performs similarly to random channel selection. Although an alternative role pairing achieves strong results, this advantage is demonstrated on only one benchmark and within one architecture family. Broader validation is needed to establish whether the proposed schema offers consistent benefits.

Critical strengths
  1. The problem is really interesting to the domain. And an industrial-phase focus dataset is novel and important for further evaluating the baselines.
Critical weaknesses
  1. What is the relation between fault type and anomaly detection? I think the problem the paper focuses on is not clear and is consistently changing. Also, the methods used to evaluate the dataset are mostly time-series models. Does the paper work on time-series fault detection or time-series anomaly detection? If anomaly detection is the focus, why only use the AUROC as the main metric? Why not use anomaly detection accuracy as a more direct metric?
Questions for authors
  1. What is the relationship between fault detection and anomaly detection? Is the paper focusing on fault detection or anomaly detection?
7 comments

We thank the reviewer for a careful review. The central point is fair: the paper evaluated several tasks without naming them, which made the problem look as if it shifts. The revision names them at the start of Section 4.

What is the relationship between fault detection and anomaly detection? Is the paper focusing on fault detection or anomaly detection?

  • [W1, Q1] Tasks. FactoryNet is not built around a single task. It provides data and a schema from which several tasks and benchmarks can be defined: the sources come with healthy episodes, labelled faults, counterfactual pairs, simulated twins and task-phase labels, all mapped to the same four roles. The paper evaluates three of these tasks but did not name them. The revision states them up front and uses the names in the section titles, Figure 1 and the abstract:
    • Task A, healthy-only anomaly detection (Section 4.1–4.2): models see only healthy episodes in training and score unseen episodes, with no fault labels.
    • Task B, supervised fault detection under domain shift (Section 4.3): a classifier is trained on labelled faults in one domain (simulation or another robot) and tested on another.
    • Task C, phase recognition (Section 4.3): a classifier labels which stage of the task is running, trained in one domain and tested on another; it asks whether task structure transfers.
  • [W1, Q1] Relationship. In the standard terms of fault diagnosis (Isermann, 2006; Gertler, 1998) and anomaly detection (Chandola et al., 2009):
    • A fault is a physical cause, such as a wrong payload or a stuck valve. An anomaly is what the data show: behaviour that differs from healthy operation.
    • Fault detection asks whether a fault is present. It can be learned without fault examples, by treating faults as anomalies (Task A; the usual industrial setting, where faults are rare, and the residual approach of model-based fault diagnosis, Chow & Willsky, 1984), or from labelled faults (Task B).
    • Both are fault detection; they differ in the training data they assume. FactoryNet supports both because it records healthy and labelled fault episodes on the same machines.
  • [W1] Fault type. Which anomaly a fault produces depends on its type, and this is the question S-E-F-C is meant to answer; the paper did not say so clearly. A residual reacts only to a fault that changes the signal it predicts, and the fault type decides where in the control loop that change appears. In fault-diagnosis terms, residuals between roles form a set of structured residuals: each fault type moves a known subset of them. The revision adds a per-fault table across datasets (now in the Appendix):
    1. Position faults appear in Setpoint→Feedback: voraus-AD invalid gripping position, AUROC 97.5; Graabæk et al. object outside the box, 0.975.
    2. External forces appear in measured force or torque: voraus-AD cable caught on the robot, Setpoint→measured torque 99.9 in all four model families; Graabæk et al. TCP wrench, Setpoint→TCP force 0.957, against 0.582 for Setpoint→motor current.
    3. Motor-electrical faults appear in Effort: voraus-AD miscommutation, 97.6 to 100 for every Effort target, against 59 to 73 for measured torque.
    4. Faults a feedback loop compensates appear in Effort, not Feedback. On the Tennessee Eastman process, a step in condenser cooling-water temperature (IDV5) is regulated out of the measured temperatures, but the valve command stays shifted: Feedback→Effort detects it at 100, PCA on all variables at 66 to 72. On the PRONTO plant (Stief et al., 2019), the air-flow controller compensates an air leak, and Effort targets reach 83 to 93.

A single score over all signals hides this structure; residuals between roles expose it.

The remaining points follow in the next comments.

If anomaly detection is the focus, why only AUROC? Why not anomaly-detection accuracy?

  • [W1] AUROC is the protocol of the voraus-AD benchmark (Brockmann et al., 2024), so it keeps our rows comparable with the published baselines. We agree that it hides the operating point. The revision also reports AUPRC, and F1, balanced accuracy and accuracy at a threshold fixed before testing at the 99th percentile of healthy validation scores, so no test label sets it (see Appendix). On voraus-AD:
Inputs → targetAUROCAUPRCF1Bal. acc.Acc.
Feedback + Context → Effort93.296.088.284.585.1
All signals → Effort80.188.450.766.056.9
  • [W1] We report balanced accuracy next to accuracy because 64% of the voraus-AD test episodes are faulty, so flagging everything already scores 64% accuracy. The all-signal model loses most at a fixed threshold: its healthy scores are widely spread, so the threshold sits high.

The transfer experiments rely mainly on general-purpose time-series classifiers rather than anomaly-detection methods.

  • [Summary] Task B is a supervised question (do simulated fault examples help?), so a classifier is the right tool. We chose MiniRocket because it is fixed and not pretrained, so transfer cannot come from a pretraining corpus.
  • [Summary] To close the gap the reviewer points out, the revision adds a healthy-only detector under the same shift: a Setpoint→current residual MLP trained only on simulated healthy KR10 episodes. It separates simulated faults at AUROC 0.99 but real faults at 0.43 [0.36, 0.49] (0.52 with a fixed current scaling), while the same model trained on real healthy episodes of the same session reaches 0.63 to 0.75. Anomaly detection therefore shows the same simulation-to-real gap as the classifiers.

The simulator calibration procedure and the real data used for calibration should be in the main text.

  • [Summary] Section 4.3 now describes them. The simulator was calibrated to one fault, a payload change on the UR3:
    • The command lag was fitted on all 109 real healthy episodes of that program.
    • The disturbance-observer response and the fault onset at the grasp were fitted on its 99 real fault episodes.
    • Only the timing-noise model was fitted on the half not being scored.
  • [Summary] Because the other components saw both evaluation halves, we call the result a case study, as the paper did. On the held-out half B, motor-current AUROC rises from 0.62 uncalibrated to 0.92 to 0.99 calibrated (real-data ceiling 0.99). Tracking-error AUROC over the carry phases rises from 0.40 to 0.58 to 0.66, and to 0.68 to 0.71 with the timing-noise model (ceiling 0.89 when trained on real half A). A confirmation on a held-out session is listed as a limitation.

The main structured baseline underperforms established methods and is similar to random channel selection; the better pairing is shown on one benchmark and one architecture.

  • [Summary] The reviewer is right on all three points. The revision now reports one fixed rule, Feedback + Context → Effort, across model families and systems.
voraus-AD, mean AUROC: inputs → EffortMLPCNNGRURidge
Fixed rule: Feedback + Context92.585.885.480.4
All signals79.480.372.062.4
  • [Summary] The rule beats all signals in every family, by 5 to 18 points. A pretrained forecaster does not remove the need for roles: Chronos-Bolt, which sees each torque channel without the commanded motion, is at chance (52.1), against 86.0 for the same channel predicted from the Setpoint.
Same healthy-only protocol, mean AUROCResidual, Effort targetRandom inputs → EffortPCA
Tennessee Eastman process (21 faults)90.888.9 ± 1.386.4
PRONTO plant, real (3 faults)79.270.5 ± 10.257.4
Air-handling unit, real (5 faults)85.173 to 8582.4
AURSAD robot (3 faults)64.248.351.0
  • [Summary] On all three process systems, an Effort-target residual is the strongest healthy-only detector we ran. Random inputs come close when they also predict Effort: the gain comes from choosing Effort as the target, which needs a schema that separates Effort from Feedback; the exact inputs matter less. The appendix also reports the limits: residuals help little on AURSAD (Isolation Forest 55.3) and the strongest published detectors stay ahead (MVT-Flow 93.6 on voraus-AD; LOF 0.95 against 0.93 on Graabæk et al., 2023).
  • [Summary] What the revision claims. Not a higher single score, but three things:
    1. Knowing which residual sees which fault, on any machine (first comment), across robots, process plants and building HVAC.
    2. A default that holds in every model family: on voraus-AD, feeding every signal into one model is 5 to 18 points worse than the role rule.
    3. One structure for every machine, so a task defined once (anomaly detection, fault classification, phase recognition, transfer) runs on every source. This is what makes the transfer study in Section 4.3 possible.

Presentation

  • [Summary] Section 4 opens with the definitions above and the three tasks. Each experiment states its task, its training data and its metric in one line. The calibration procedure is in the main text, the additional metrics and model families are in appendices, and the per-fault table connects fault types to residuals. Section titles, Figure 1 and the abstract use the task names.
Reviewer_2EvgOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a new dataset for anomaly detection mainly for several types of industrial robots and one milling machine. Therefore, they generated own dataset and curated existing datasets into their new comprehensive dataset. They augmented the datasets with controlled anomalies and enriched it with simulations in Isaac Gym. Additionally, they introduce a new taxonomy "Setpoint-Effort-Feedback-Context" (S-E-F-C) that categorizes the recorded data into commanded intent, actuation energy, measured physical outcome, and environmental variables and aims to provide a common interface for training foundation models on multiple types fo robots. The evaluation narrowly shows that the proposed S-E-F-C slicing of the dataset provides advantages in anomaly detection compared to models that use all signals. A transferability study provides a limited insight into the transferability from sim to real and between robots.

Critical strengths

Topic

The paper touches an interesting and imporant topic:
There is a real gap for large and detailed industrial datasets and benchmarks. 


## Evaluation
The paper evaluates two important questions on the dataset: 
Whether the proposed taxonomy offers advantages, and whether the dataset allows evaluating sim to real and robot to robot evaluations.
Critical weaknesses

Novelty

The novelty of the paper is little: 
The dataset is partly compiled from existing datasets and the introduced S-E-F-C taxonomy inspired by IEC 81346.
Though it shows advantages in a single evaluation setup, a full proof remains. 


## State of the Art

There are multiple references for similar datasets missing in the state of the art. 
Most obviously (if you want to remain in a robot domain) Grabaek et al. 2023 (10.1109/access.2023.3289068). 
If you want to cover other industrial problems, Moddemann et al. 2024 (https://doi.org/10.1109/ETFA65518.2025.11205680) , Goldrick et al. 2019 (10.1016/j.compchemeng.2019.05.037).


## Overclaim

The paper overclaims its contribution.
The majority of the datasets are robot datasets.
There is only one CNC milling dataset included. 
Hence, the title and claim that it is a "Large Scale Dataset for Industrial Time-Series Foundation Models" is an overstatement.
I would consider narrowing the contribution down to industrial one-arm robots. 
There is additionally no general overview over and baseline results over the complete dataset and tasks that allow using it as a benchmark. 

# Writing & Presentation
The writing of the paper and the presentation of the results appears a unstructured and anecdotal.
A clearer presentation of the evaluation results would help to contextualize the results and evluate the dataset better.
Over the paper it remains unclear what the main contribution and its novelty is, the dataset, the taxonomy, both, and what it has to do with foundation models.
Questions for authors

How well does the dataset allow to evaluate a transfer between all the included robots? Does the S-E-F-C taxonomy slicing also improve performance for other models than the MLP regressor?

7 comments

We thank the reviewer for the careful review and for agreeing that large, detailed industrial datasets are a real gap and we are committed to solve this. The reply is split in: Weaknesses W1–W4 first (Part 1), then the two Questions Q1–Q2 (Part 2).

Part 1: Weaknesses

  • W1. Novelty
    • "The dataset is partly compiled from existing datasets and the introduced S-E-F-C taxonomy inspired by IEC 81346. Though it shows advantages in a single evaluation setup, a full proof remains."
    • Compiled from existing data. Of the 56,591 executions, 4,185 come from public sources. Of the 14,847 real executions, 10,662 (72%) were recorded by us. Re-expressing public sets in the shared schema is what makes cross-machine comparison possible.
    • S-E-F-C and IEC 81346. We agree and now say so: S-E-F-C is presented as a practical convention inspired by IEC 81346, not as a new theory. Its value is the empirical claim, now tested on robot arms, process plants and building HVAC (W3 and Q2).
    • Single evaluation setup. The ablation now covers seven systems besides voraus-AD and, on voraus-AD, four model families. We do not claim a full proof: see W3 and Q2 for the results.
  • W2. State of the art (missing references)
    • Grabaek et al. 2023; Moddemann et al. 2025; Goldrick et al. 2019.
    • All three are now cited and in Table 1. We ran our method on the two with public data, under each benchmark's own protocol:
Benchmark and metricRole residual (ours)Baselines
Graabæk et al., per-sample over 8 anomaly types, ROC-AUC / PR-AUCFeedback + Context → Effort: 0.857 / 0.681PCA 0.647 / 0.352; LOF (best published) 0.953 / 0.826
Graabæk et al., with LOF's program-node clock added as Context0.928 / 0.788LOF without the clock 0.905 / 0.750
IndPenSim, 10 faulty vs held-out healthy batches, batch AUROC (MLP / ridge / CNN)99.6 / 94.7 / 100PCA 97.3; Isolation Forest 81.3
  • Graabæk et al. Role choice moves ROC-AUC from 0.58 (Context only) to 0.86. LOF's lead comes mostly from a program-node clock among its 25 features, which our schema files as Context; with the same clock, the residual is within 0.03 of LOF.
  • Goldrick et al. (IndPenSim). Recipe feeds are Setpoint and the pH and temperature PID outputs are Effort. The original paper reports only qualitative PCA charts, so these are the first numbers on it. Every pairing that predicts the PID outputs from Feedback is within 2 points of the top; with 25 test batches the margin over PCA is small, and we claim no more.
  • Moddemann et al. (HAI-CPPS) models no actuator response, so only Setpoint → Feedback pairings are possible, and there are no published detection numbers to compare against.
  • Graabæk et al. also log target and actual current, as voraus-AD and AURSAD do, so we removed any wording suggesting that FactoryNet alone records Setpoint and Effort. What FactoryNet adds is the harmonisation of many machines, real and simulated, under one control-loop schema.
  • W3. Overclaim
    • The majority of the datasets are robot datasets; only one CNC milling dataset; consider narrowing to industrial one-arm robots; no general overview over, and baseline results on, the complete dataset and tasks.
    • Scope beyond robot arms. We agree that the corpus as submitted is weighted toward arms: they make up 56,573 of 56,591 episodes, and the revision states this in a coverage table by machine class. We keep the industrial framing because the claim is not about arms. S-E-F-C is defined for any actuated machine, and the finding is that the residual between commanded and realised behaviour carries the fault signal. We tested this on five non-robot systems with labelled faults, under the healthy-only protocol of Section 4.2:
SystemBest role residualRole-agnostic baseline
Tennessee Eastman process, mean AUROC over 21 faults90.8 (Feedback → Effort)PCA 86.4
PRONTO multiphase-flow pilot plant (real), mean AUROC over 3 faults79.2PCA 57.4
Air-handling unit (LBNL, real single-zone), mean AUROC over 5 faults85.1PCA 82.4
Water pump (SKAB, its own protocol), F10.70 (role choice range 0.06–0.70)best published detector 0.78
Penicillin fermentation (IndPenSim), batch AUROC: MLP / ridge / CNN99.6 / 94.7 / 100PCA 97.3
  • In every case the most informative residual relates Effort to Feedback, in one direction or the other, the same relation that carries the fault signal on the robots. The clearest case is a cooling-water fault on the Tennessee Eastman process (IDV5): the condenser loop regulates it out of the measured temperatures, but the valve command stays shifted, so Feedback → Effort detects it at 100 AUROC against 66 to 72 for PCA. The adapters for these sources are released with the code. Wherever a sentence rests only on arm evidence, the revision now writes "industrial robots".
  • No overview or baselines over the complete dataset. Added. A new table gives healthy-only baselines for each public fault source, with splits and scripts, reported as AUROC, AUPRC and balanced accuracy at a fixed threshold. Mean AUROC (best role residual / PCA / Isolation Forest): voraus-AD 92.5 / 69.3 / 74.3; AURSAD 64.2 / 51.0 / 55.3; IndPenSim 99.6 / 97.3 / 81.3. Best role residual / PCA: Tennessee Eastman 90.8 / 86.4; PRONTO 79.2 / 57.4; air-handling unit 85.1 / 82.4. No method wins everywhere, and AURSAD remains open, which is what a possible benchmark needs.
  • W4. Writing and presentation
    • Unstructured and anecdotal; unclear what the main contribution and novelty is (dataset, taxonomy, both) and what it has to do with foundation models.
    • The introduction now ends with three numbered contributions:
      1. a corpus of 56,591 executions, 10,662 of them new real laboratory recordings with injected faults and 41,744 simulated;
      2. a control-loop schema with released adapters that put heterogeneous machines into one dataloader;
      3. findings that the schema makes measurable: which roles carry fault signal, and what transfers from simulation and between machines.
    • Section 4 now opens with a table that states each question, its setting and its result. Every experiment is labelled Task A, B or C.
    • The link to foundation models is that a pretraining corpus for industrial machines needs commanded and realised signals in a shared frame across vendors, and contributions (1) and (2) provide this.

Questions

  • Q1. How well does the dataset allow evaluating transfer between all the included robots?
    • The revision adds a full phase-recognition matrix over every pair that shares the pick-and-place task (macro-F1):
    • Within a robot: real UR3 0.93, real KR10 0.66.
    • Between the real robots: 0.15 (UR3 → KR10) and 0.23 (KR10 → UR3), or 0.32 and 0.40 on the five stages both robots share (chance 0.20).
    • Leave-one-arm-out in simulation: held-out UR arms 0.88 to 0.93, KR10 0.56.
    • Pooled simulated arms → real robots: 0.32 to 0.41.
    • The transfer is strong inside the UR family and weak across manufacturers. The dataset therefore supports transfer evaluation, and shows that cross-manufacturer transfer is still unsolved.
    • Why cross-manufacturer transfer is weak. We tested whether it is a scale problem. Per-robot standardisation or datasheet scaling changes macro-F1 by at most 0.09, and not consistently. Stage timing explains most of the gap: giving every labelled stage the same duration (an oracle that uses labels) lifts UR3 → KR10 and KR10 → UR3 from 0.35 and 0.40 to 0.79 and 0.82. The two task programs spend different shares of the cycle in each stage (the return stage is 18.8% of the UR3 cycle and 7.7% of the KR10 cycle), and a window covers one eighth of a cycle.
  • Q2. Does the S-E-F-C slicing also improve performance for models other than the MLP regressor?
    • Yes for the main effect. We reran the voraus-AD ablation with ridge regression, a causal temporal CNN and a GRU (mean per-category AUROC, standardised residuals):
MLPCNNGRURidge
Best role pairing92.587.585.480.4
All 124 signals → Effort79.480.372.062.4
  • Predicting Effort from a role-chosen input beats predicting it from every signal in all four families. The ranking of pairings agrees with the MLP.

Comments on the paper

2 comments

We thank the reviewers for their careful reading. Three points were shared, so we addressed them together:

  • What the paper claims The contribution is the collection and its standardisation: every signal is mapped to Setpoint, Effort, Feedback and Context, so that heterogeneous machines share one control-loop schema. We then test how the schema helps on three tasks, which are the worked examples of its use. The submitted text did not name the task of each experiment, so the revision now states them up front. Task A is healthy-only anomaly detection, where the schema decides which signals are inputs and which is the target; it tests whether the schema supports anomaly detection (§4.1–4.2). Task B is supervised fault detection under domain shift, and Task C is phase recognition; both test whether a monitoring model transfers from simulation to real and from robot to robot (§4.3). Results are reported with AUPRC, F1 and balanced accuracy at a fixed threshold, in addition to AUROC.
  • Is the S-E-F-C benefit general? The role ablation now runs beyond voraus-AD and beyond robots: on two more robot benchmarks (AURSAD; Graabæk et al.) and on five systems that are not robots, a water-pump testbed (SKAB), a penicillin fermentation process (IndPenSim), the Tennessee Eastman chemical process, the PRONTO multiphase-flow pilot plant and building air-handling units (LBNL). On voraus-AD it also runs with four model families: MLP, ridge, temporal CNN and GRU. Two results hold throughout. A role-chosen Effort target beats the all-signal default on voraus-AD in every family. And a residual that relates Effort to Feedback is the best or tied-best pairing on Graabæk et al., SKAB, IndPenSim, the Tennessee Eastman process, PRONTO and the real air-handling unit, so the same relation carries the fault signal in a robot arm, a pump, a bioprocess, a chemical plant and a building HVAC system.
  • Can the corpus serve as a benchmark? Yes, and it is built to support several. FactoryNet gives the data and one common schema; a benchmark is then a choice of task (A, B or C), of sources, of split and of metric, all expressed through the same dataloader and the same four roles. The revision shows two ways to use it. (1) Task A on every public fault source: a new table gives healthy-only baselines per source (the best role residual, PCA, Isolation Forest and a zero-shot foundation model where it applies), with splits and scripts released. (2) Tasks B and C across machines: the phase-recognition matrix over every robot pair, with leave-one-arm-out and simulation-to-real splits.
AuthorsOpenReview ↗

Discussion summary: e1z1 raised the score, revised paper uploaded

Dear Area Chair,

As the discussion closes, here is where the paper stands. Reviewer e1z1 raised the score after reading the revised PDF, and wrote that it reads more clearly and is easier to understand. Reviewer vUt4 recommended acceptance from the start. Reviewer 2Evg could not find our changes at first; the revised PDF is now online, and on their concern about novelty, 93% of the 56,591 executions are new.

The data is multimodal, as e1z1 suggested we emphasise: every robot episode carries kinematic, electrical, mechanical, thermal and controller signals, part of the laboratory episodes also carry scene and wrist-camera video, and both real robots have simulated twins for the sim-to-real study.

The reviews led to three main changes:

  • The paper opens with the dataset and the schema, lists three numbered contributions, and labels every experiment Task A, B or C.
  • The schema's benefit now holds in four model families on voraus-AD (5 to 18 AUROC points over feeding every signal) and on seven further systems, five of them not robots.
  • Healthy-only baselines for every public fault source, with splits and scripts, a phase-recognition matrix over all robot pairs, and AUPRC, F1 and balanced accuracy.

We have since rewritten the paper further and look forward to the reviewers' replies. We recorded the 10,662 real executions ourselves because this kind of data was missing, and we are keen to keep growing it with more machines so that it becomes a shared foundation for industrial machine learning. Thank you for your time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.