A Synthetic Multivariate Refrigerator Time-Series Dataset for Predictive Maintenance
Islam Benamirouche, Feriel Fass, Djemel Ziou
cs.LG
Quick verdict
Physics-inspired simulator yields 7.07M rows of annotated synthetic fridge time series; strong reproducibility, but validation and baselines are thin.
Summary
The authors release a synthetic dataset of one-minute-resolution multivariate time series from 27 simulated household refrigerators, generated by a simplified physics-inspired simulator covering ambient forcing, door use, thermostat-compressor control, heat exchange, defrost, electrical consumption, and six progressive degradation types (compressor efficiency loss, mechanical wear, condenser fouling, door-seal degradation, fan degradation, defrost-heater failure). Each unit emits 15-20 sensor outputs depending on configuration; the release totals 7,066,161 rows with per-unit failure logs, plus Python generator code, configurations, and documentation (Zenodo, GitHub). Validation examines a 48-hour nominal window (21.0% compressor duty), a single compressor-degradation pre-failure example where two-hour averaged current rises from 0.81 A to 2.3 A, class composition (89.0% nominal, 0.9% failure in the selected series; 158 failures overall, degradations 3-8 days), and sensor heterogeneity (vibration present in 66.7% of units). A failure-prediction evaluation with the authors' companion transferable autologistic model detects 54/59 failures (91.5%) on 10 target units with 101.2 h mean lead time, versus 61.0% without adaptation. Limitations: validation rests on single windows and one degradation example; evaluation uses only the authors' own model with no baselines, confidence intervals, or split sensitivity; physical realism is unvalidated against real measurements; and sensor outputs are frozen during failure intervals, an artifact models may exploit.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Public refrigerator datasets with continuous multivariate monitoring plus explicit degradation-start and failure-interval annotations are scarce, and this release fills that gap (Sec. 1).
- A simplified physics-inspired simulator generates 7,066,161 minutes of data across 27 heterogeneous units with six progressive degradation types and 15-20 sensor outputs per unit (Sec. 2-3, Appendices B-C).
- Pre-failure sensor changes support prediction: the transferable autologistic model [12] detects 54/59 failures (91.5%) on 10 target refrigerators with 101.2 h mean lead time, versus 61.0% without adaptation (Sec. 4.5).
- The Zenodo + GitHub release provides a reproducible basis for developing fault-diagnosis and prognostics methods (Conclusion).
Strengths
Complete generative specification
Appendix B-D give every scheduling rule (Eq. 8-21), phenomenon equation, fixed coefficient, and noise SD (e.g., sigma_I = 0.1 A, sigma_fr = 0.1 degC), plus a quantity-to-column map with rounding rules (Table 5), enabling exact reimplementation of the simulator.
Addresses a real annotation gap
Sec. 1 correctly identifies that public refrigerator data lacks continuous run-to-failure monitoring with explicit degradation-start and failure-interval labels; the closest real alternative (Fonseca et al. [10]) is segment-based. The released per-minute labels (Eq. 4, Table 6) fill this.
Heterogeneity by design
Honest claim calibration
Full release and cost analysis
Weaknesses
Thin, single-example validation
Self-referential predictive evaluation
Simulability bias favors detectors
Frozen outputs during failure intervals
Physical realism untested
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a dataset paper in Data-Descriptor format presenting 7,066,161 rows of synthetic, minute-resolution multivariate time series from 27 independently simulated household refrigerators, plus 27 failure logs, a Python generator, configurations, and documentation (Zenodo DOI 10.5281/zenodo.22076228; github.com/iben0x/fridge-sensor-data). The motivating gap (Sec. 1) is real: public refrigerator data with continuous run-to-failure monitoring and explicit degradation-start/failure annotations is scarce; the closest real dataset cited (Fonseca et al. [10]) contains extracted operating segments, not uninterrupted multi-month histories. Four claims carry the paper: (i) the release fills the annotation gap; (ii) six progressive degradation types (Table 3) with heterogeneous sensor sets (15-20 outputs per unit) support cross-equipment learning; (iii) pre-failure signals are learnable - the companion transferable autologistic model detects 54/59 failures (91.5%) on 10 target units with 101.2 h mean lead time (Sec. 4.5); (iv) the release is a reproducible basis for prognostics research. Claims (i), (ii), and (iv) are well supported by the text; (iii) is supported only in a self-referential sense.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Is the refrigerator dataset real measured data?
No. All 27 refrigerators are simulated with a simplified physics-inspired simulator. The authors explicitly state that validation against real degradation and failure histories remains future work, and that the failure durations reflect simulation settings, not measured real-world timescales.
What is in the dataset and how large is it?
27 CSV time-series files totaling 7,066,161 rows at one-minute resolution with 15-20 sensor columns per refrigerator, plus 27 failure logs (with degradation start, failure start/end, type, severity). The Python generator, configurations, and documentation are on GitHub (iben0x/fridge-sensor-data); the data are on Zenodo (DOI 10.5281/zenodo.22076228).
Which faults are simulated?
Six progressive degradation types: compressor efficiency loss (COMP_DEGRAD), compressor mechanical wear (COMP_WEAR), condenser fouling (COND_DIRT), door-seal degradation (DOOR_SEAL), fan degradation (FAN_DEGRAD), and defrost-heater degradation (DEFROST_FAIL). Each retained failure has a 3-8 day degradation period followed by a 2-21 hour failure interval.
Can the data be used for failure-prediction benchmarks?
Yes. Every minute carries an equipment-state label (fault_indicator), a degradation_indicator, and cumulative fault counters, enabling supervised pre-failure prediction. The paper's own benchmark uses its companion autologistic model (91.5% detection, 101.2 h mean lead time on 10 target units), but no independent baselines have been reported in the paper itself.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-6 are present only as captions in the extracted text; their visual content (trends, distributions) could not be inspected, so figure-based judgments rely solely on the numbers quoted in the body text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.