{"id":"4b8ac13a-2144-4800-8769-fd86a0990a19","arxiv_id":"2411.13268","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A calibrated nine-sensor IoT array with sequential Monte Carlo Bayesian inference localizes a water vapor source to within 10-20 percent of the sensor spacing.","lead":"Nine IoT humidity sensors are combined with a Bayesian particle filter to locate the source of a water vapor plume in a room. In synthetic and real tests the estimated source position ends up within 10 to 20 percent of the sensor spacing.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real validation is effectively 1D: both tested sources lie on the sensor array's symmetry axis, so the circular Gaussian plume model is never challenged in the perpendicular direction.","rationale":"The paper makes a modest, plausible engineering claim, and the SMC implementation is standard. The numerical tests in Sec. IV.A are self-consistency checks: data are generated from Eqs. (2)–(4), so they validate the sampler, not the environmental model. The real experiment is the only evidence for the central claim. Its two source positions both sit on the line y=128, a symmetry axis of the cross array. Because the likelihood depends only on pairwise distances, the posterior's y-marginal is symmetric and its mean is exactly the true y=128; the reported error is effectively the x-error. This does not prove the algorithm wrong, but it means the circular-Gaussian model has never been challenged by an asymmetric configuration. A buoyant steam plume in a real room is likely to be advected or asymmetric, and such a bias would first appear in the y-direction for off-axis sources. The reader's weakest assumption named the same model risk; this critique sharpens it by explaining why the chosen geometry hides it. The requested conditions (data release, baseline comparison, sensitivity analysis) remain appropriate, and adding an off-axis validation is necessary before the accuracy claim can be generalized. Therefore the conditional verdict stands unchanged.","tokens_in":12986,"tokens_out":9351,"duration_ms":114019,"concrete_test":"Repeat the real measurement protocol with the water-vapor source placed at an off-axis position that breaks the array symmetry, e.g., (98,80) cm or (60,100) cm in the same 256×256 cm arena, keeping the same nine-sensor cross, calibration, 500-Ω threshold, 20-s moving average, and 10-repeat averaging. If the mean center-of-mass error in the y-coordinate, or the total error, exceeds the sensor spacing ds=60 cm, or if the SMC belief develops a systematic y-bias, then the collinear geometry is load-bearing and the generalized accuracy claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. IV.B reports real localization errors of 10–20% of the sensor spacing ds for two source locations, (98,128) and (38,128) cm. The nine-sensor cross is symmetric about both axes, and both sources lie on the horizontal symmetry line y=128. For any likelihood that depends only on distances to the sensors (Eqs. 3–5), and with a symmetric prior, the posterior in y is exactly symmetric when the true source sits on that line. Hence the SMC center-of-mass y-coordinate is unbiased by construction, and the reported error is essentially a one-dimensional along-row interpolation error. The Gaussian plume model of Eq. (2) is circular and smooth; a real buoyant plume distorted by convection, which the authors explicitly neglect in Sec. III.A, would introduce a bias in y that these two collinear tests cannot detect. Thus the headline claim 'error consistently lower than sensor spacing' is supported only for a favorable 1D geometry, not for general 2D localization.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a gas source localization method that combines nine low-cost IoT sensors with a sequential Monte Carlo (SMC) Bayesian inference scheme. The environment is modeled by a stationary Gaussian plume (Eq. 2) and the sensor readings are converted into Poissonian hit counts (Eqs. 3-4). The method is tested first on synthetic data generated from the same model and then on real water-vapor experiments with two source locations. The authors report localization errors of 10-20% of the sensor spacing in both settings and claim that the error is consistently lower than the distance between adjacent sensors.","tokens_in":13090,"tokens_out":3944,"duration_ms":45835,"significance":"If the general two-dimensional localization claim is supported, the paper would provide a useful demonstration that a simple Bayesian Monte Carlo approach, combined with inexpensive distributed IoT hardware, can localize an indoor gas source with sub-sensor-spacing accuracy. The strengths of the manuscript are the concrete end-to-end system description, the repeated real experiments (10 repetitions per location), and the algorithmic detail in the appendix that makes the SMC implementation reproducible. However, the current experimental validation is too narrow to establish the headline claim: the two real source locations lie on the symmetry axis of the sensor array, so the reported accuracy is essentially a one-dimensional along-row result.","major_comments":[{"comment":"The real-world validation tests only two source positions, (98,128) cm and (38,128) cm, both lying on the horizontal symmetry line y=128 of the nine-sensor cross. Because the likelihood (Eqs. 3-5) depends only on distances to the sensors, and the prior is uniform, the posterior is exactly symmetric about y=128 for a true source on that line. Consequently, the SMC center-of-mass y-coordinate is unbiased by construction, and the reported 10-20% of ds error is essentially an along-row interpolation error. To support the claim of general 2D source localization, off-axis source positions must be tested, or at minimum the separate x and y errors should be reported; as it stands, the experiment does not challenge the model in the perpendicular direction.","section":"Sec. IV.B, Fig. 8"},{"comment":"The synthetic experiment generates sensor data from the same Gaussian plume and Poisson hit model that is later used for inference (Eqs. 2-4). This is a valid self-consistency check of the SMC implementation, but it cannot validate the environmental model. The only model-adequacy evidence is the real-data experiment, which is limited to two collinear source locations and only 10 repetitions each. The paper should either add more real source configurations, including off-axis ones, or explicitly characterize how the reported accuracy degrades as the model mismatch increases.","section":"Sec. IV.A"},{"comment":"No baseline or comparison method is reported for the real-data experiment. The claim that the Bayesian SMC algorithm localizes the source to 10-20% of the sensor spacing would be much stronger if compared to simple alternatives such as the nearest-sensor estimate, the weighted centroid of the hit counts, or a direct least-squares fit of the plume model. Without such a baseline, it is unclear how much of the reported accuracy is due to the Bayesian machinery and how much is already contained in the sensor geometry and the hit-rate pattern.","section":"Sec. IV.B"},{"comment":"The preprocessing pipeline relies on hand-tuned choices: the moving-average window of 20 s and the hit threshold of 500 Ω. These choices directly determine the discrete hit counts h that enter the likelihood, so the reported localization performance may be sensitive to them. The paper should include a sensitivity analysis over these parameters, or at least justify the chosen values with a principled criterion, before the 10-20% accuracy can be considered a robust property of the method.","section":"Sec. IV.B and Fig. 7"}],"minor_comments":[{"comment":"The exponent in Eq. (2) appears to be missing the square on the distance: it should be exp(-||r - r_s||^2 / lambda^2), consistent with Eq. (4). The current notation exp(-||r - r_s|| / lambda^2) is dimensionally inconsistent with the rest of the paper.","section":"Eq. (2)"},{"comment":"The acknowledgments paragraph is duplicated verbatim before the Abbreviations section; one copy should be removed.","section":"Acknowledgments"},{"comment":"The abbreviation 'IAQ Indoor Air Quality' is listed twice in the Abbreviations section.","section":"Abbreviations"},{"comment":"The shaded region in the bottom panels is described only as a confidence interval; please state explicitly whether it is one standard deviation, a 95% interval, or another quantile, and over how many experiments it is computed.","section":"Fig. 8"},{"comment":"The statement that the algorithm 'could converge to the correct source location typically within such a shorter time frame' would benefit from a quantitative criterion defining convergence, rather than a visual inspection of the figures.","section":"Sec. IV.B"},{"comment":"The priors for lambda and Q~ are described only as 'uniform' in Sec. IV.A, while Table A1 lists ranges; please state explicitly whether the uniform priors are in the parameters themselves or in some transformed variables, since this affects the SMC implementation.","section":"Sec. IV.A and Table A1"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a plausible and potentially useful integration of IoT hardware with Bayesian SMC for gas source localization, and the implementation details are commendable. The central issue is not the algorithm itself but the mismatch between the headline claim of general 2D localization and the validation, which is effectively one-dimensional because both real source locations lie on the symmetry axis of the sensor array. This is fixable within the manuscript's scope by adding off-axis experiments or by substantially tempering the claim, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a clean, well-written demonstration of a known method (Bayesian SMC) on a new IoT sensor platform (Sensiplus/Sensibus) with real water vapor data. The algorithmic pieces are standard and the authors say so. The genuinely new content is the end-to-end integration: nine low-cost sensors on a cross, careful calibration, and an honest admission that the dispersion model is a Gaussian plume that neglects convection.\n\nThe main soft spot is the geometry of the validation. Both real source locations, (98,128) and (38,128), lie on the horizontal symmetry axis of the sensor cross (centered at 128,128). For any likelihood that depends only on distances to sensors, the posterior in y is exactly symmetric, so the SMC mean y-coordinate is unbiased by construction. The reported 10-20% of sensor spacing is therefore essentially a one-dimensional along-row interpolation error. The stress-test note is correct: the Gaussian plume model is never challenged in the perpendicular direction, and a real buoyant plume distorted by convection could bias y in a way these two collinear tests cannot detect. The paper's headline claim is unqualified, but the evidence only supports 1D localization.\n\nOther gaps: no comparison to simpler baselines (e.g., hit-count centroid or least-squares plume fit), only 10 repeats per location, and hand-picked preprocessing choices (500 ohm threshold, 20 s moving average) with no sensitivity analysis. The synthetic test is a self-consistency check, not a prediction. Code and data are not released, which limits reproducibility.\n\nThat said, the paper is not circular: the source position is not used to set lambda or Qtilde; those are estimated online by the SMC. The calibration procedure is careful, and the real-data plots are convincing within the narrow geometry tested. This is a useful reference for anyone building a static-array gas localization system, and it is a fair example of what standard Bayesian methods achieve in a realistic indoor setting.\n\nI would send this to peer review, but with a request for major revision: test at least one off-axis source location, add a baseline comparison, analyze sensitivity to the preprocessing thresholds, and ideally release the data. As is, the conclusion should be qualified to say 'errors below sensor spacing for sources on the sensor array's symmetry axis.'","headline":"A solid engineering integration of SMC with IoT gas sensors, but the real-data validation only tests sources on the sensor array's symmetry axis, so the headline claim is broader than the evidence.","tokens_in":13744,"tokens_out":4837,"would_cite":false,"duration_ms":51228,"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":"Nine IoT sensors locate a gas source within 20% of their spacing","keywords":["gas source localization","Bayesian inference","sequential Monte Carlo","Gaussian plume model","IoT sensor network","water vapor sensing","Poisson detection model","indoor air quality"],"falsifier":"Run the identical nine-sensor array and SMC algorithm in a room with a steady cross-draft, for example a household fan producing a 0.5 m/s wind across the plume, and measure the average localization error against the known source position; if the error systematically exceeds the 60 cm sensor spacing or is biased downwind, the Gaussian plume assumption fails. A second check would be to compare the measured time-averaged concentration profile along the sensor line with the exponential $e^{-d^2/\\lambda^2}$ shape of Eq. (2); a mismatch predicts exactly where the method will break.","tokens_in":12709,"feed_emoji":"💨","tokens_out":12764,"duration_ms":106612,"temperature":0.7,"pith_summary":"This paper tries to establish that a network of nine cheap, low-power IoT gas sensors, running a Bayesian inference algorithm on their discrete 'hit' counts, can localize a gas source in an indoor room to better than the spacing between sensors. The authors test this with water vapor released from a kettle on the floor, using a nine-sensor array hanging 110 cm below the ceiling, and report that the average localization error falls to 10-20% of the 60 cm sensor spacing within about 15 minutes. This matters because gas leak detection in homes and workplaces could then be done with inexpensive distributed hardware instead of costly analytical instruments, and the same probabilistic machinery might extend to turbulent environments where gas signals are intermittent. The paper's central burden is to show that a deliberately simplified Gaussian plume model, one that ignores convection and buoyancy, still contains enough information for the Bayesian update to find the source.","feed_headline":"Nine IoT sensors locate a gas source within 20% of their spacing","feed_subtitle":"A Bayesian Monte Carlo algorithm reads water-vapor hits from nine cheap sensors and pinpoints the source in minutes.","key_machinery":"The central object is the Gaussian plume model of Eq. (2), $c(\\mathbf{r}-\\mathbf{r}_s) = \\frac{Q}{\\pi u \\lambda^2} e^{-\\|\\mathbf{r}-\\mathbf{r}_s\\|^2/\\lambda^2}$, paired with the Poisson detection model of Eqs. (3)-(4), where each sensor records hits $h_i$ with mean hit rate $\\mu(d_i) = \\frac{\\tilde{Q} a \\Delta t}{\\lambda^2} e^{-d_i^2/\\lambda^2}$. The algorithm is a Sequential Monte Carlo particle filter with a Metropolis-Hastings perturbation step that carries a belief distribution over the source position $\\mathbf{r}_s$ and the two unknown parameters $\\tilde{Q}$ and $\\lambda$, updating it with Bayes' rule at every measurement round.","core_discovery":"On its own terms, the paper's discovery is that a sequential Monte Carlo algorithm, which jointly infers the source position and two environmental parameters (effective emission rate $\\tilde{Q}$ and diffusion length scale $\\lambda$) from the Poisson hit counts of nine static sensors, localizes a water vapor source accurately in both simulations and a real room. In numerical experiments with known ground truth, the localization error averaged over one thousand runs settles at about 10% of the sensor spacing. In the real experiment, repeated ten times at each of two source positions, the average error is 10-20% of the 60 cm spacing by the end of the run, with convergence typically reached before the full 20 minutes. The authors conclude that, despite the oversimplified environment model, the combination of the Gaussian plume likelihood and Monte Carlo sampling is sufficient for practical indoor source localization in the tested configurations.","pith_inferences":["A likely failure mode is a ventilated room: because the model has no mean wind, a steady cross-draft would break the circular symmetry of the Gaussian plume, biasing the estimate downwind; this can be tested directly with a desk fan.","The two tested source positions both lie roughly along the horizontal symmetry axis of the sensor cross; the paper does not test sources near the arena boundary, where fewer sensors would see the plume and the posterior may become multimodal.","The Poisson hit discretization suggests a direct recipe for other sparse chemical sensors: calibrate the per-sensor background, threshold to hits, and reuse the same likelihood, which would let an e-nose array adopt the method without new algorithm development.","The claim 'consistently lower than the sensor distance' rests on 20 real trials total (two positions, ten repetitions); whether it holds for arbitrary source positions, room sizes, and source temperatures remains open."],"forward_implications":["Source localization becomes feasible with hardware costing under 10 USD per sensing chip and drawing less than 1.4 mW, since the algorithm needs only discrete hit counts from a static sensor array.","The method converges online, typically within 1000 seconds, fast enough for real-time leak response and continuous indoor air quality monitoring.","Because the likelihood is a Poisson model of sparse detections, the same algorithm should carry over to turbulent environments where gas signals are intermittent, not just the near-stationary plume tested here.","The algorithm simultaneously estimates the environmental parameters $\\tilde{Q}$ and $\\lambda$, so it does not need prior knowledge of emission rate or room diffusivity, removing a common practical obstacle.","The Sensibus single-wire protocol lets nine sensors share one cable and take synchronized readings, making the hardware simple to deploy on ceilings, walls, or mobile platforms."],"supporting_citations":[{"why":"provides the analytical Gaussian plume concentration profile used as the environment model.","marker":"[29]"},{"why":"supplies a dispersion-model reference for the simplified environment.","marker":"[30]"},{"why":"introduces the discretized Poisson detection model that turns concentrations into sensor hit counts.","marker":"[39]"},{"why":"gives the Smoluchowski relation linking mean hit rate to concentration in Eq. (4).","marker":"[40]"},{"why":"provides the sequential Monte Carlo and MCMC perturbation machinery used for online Bayesian inference.","marker":"[26]"},{"why":"supplies the resampling and effective sample size steps that prevent particle degeneracy.","marker":"[27]"},{"why":"offers a prior example of Monte Carlo gas source localization on real atmospheric data, framing the novelty.","marker":"[44]"},{"why":"another recent demonstration of Monte Carlo gas source search with real measurements, used as a comparison point.","marker":"[45]"}],"fun_headline_variants":["Nine IoT sensors locate gas within 20% of spacing","Bayesian Monte Carlo pins gas source with 9 sensors","Sensor array + Bayesian inference: gas source in minutes","Gas source localization via Monte Carlo Bayesian grid","9 smart sensors + Bayesian logic beat gas source hunt"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the time-averaged concentration field at the sensor plane is a symmetric Gaussian plume described by a single diffusion length scale, with Poisson-distributed sensor hits; if real room convection or plume buoyancy distorts this shape, the inferred source position could be biased despite the success in the two tested configurations.","fun_headline_variants_meta":{"raw":{"variants":["Nine IoT sensors locate gas within 20% of spacing","Bayesian Monte Carlo pins gas source with 9 sensors","Sensor array + Bayesian inference: gas source in minutes","Gas source localization via Monte Carlo Bayesian grid","9 smart sensors + Bayesian logic beat gas source hunt"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000506,"raw_usage":{"total_tokens":2432,"prompt_tokens":870,"completion_tokens":1562,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":1486}},"tokens_in":486,"tokens_out":1562,"duration_ms":13712,"temperature":1.0,"reasoning_tokens":1486,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:37:52.165004+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical nine-sensor array and SMC algorithm in a room with a steady cross-draft, for example a household fan producing a 0.5 m/s wind across the plume, and measure the average localization error against the known source position; if the error systematically exceeds the 60 cm sensor spacing or is biased downwind, the Gaussian plume assumption fails. A second check would be to compare the measured time-averaged concentration profile along the sensor line with the exponential $e^{-d^2/\\lambda^2}$ shape of Eq. (2); a mismatch predicts exactly where the method will break.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the analytical Gaussian plume concentration profile used as the environment model."},{"cited_title":"Hutchinson, H","cited_arxiv_id":null,"evidence_quote":"supplies a dispersion-model reference for the simplified environment."},{"cited_title":"Ristic, A","cited_arxiv_id":null,"evidence_quote":"introduces the discretized Poisson detection model that turns concentrations into sensor hit counts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"gives the Smoluchowski relation linking mean hit rate to concentration in Eq. (4)."},{"cited_title":"Johannesson, B","cited_arxiv_id":null,"evidence_quote":"provides the sequential Monte Carlo and MCMC perturbation machinery used for online Bayesian inference."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"offers a prior example of Monte Carlo gas source localization on real atmospheric data, framing the novelty."},{"cited_title":"Asenov, M","cited_arxiv_id":null,"evidence_quote":"another recent demonstration of Monte Carlo gas source search with real measurements, used as a comparison point."}],"review_version":1}