PaperlayerView exact revision on arXiv

Source-linked AI summary

ReCo: When to Relocate Sensor Kits under Deployment Constraints -- A NILM Case Study

Haokun Chen, Yu Tong, Yehai Chen

arXiv:2609.35075v1cs.LG

TL;DR

With few kits, a deadline, and downtime at each move, field collectors must decide when to leave a site. ReCo forecasts new regime coverage for NILM and weighs staying against moving after downtime; in replayed deployments it outperformed fixed-dwell and count-based schedules across settings and improved both F1 and MAE over the threshold rule.

  • Problem

    Field data collection must decide when to stay or relocate despite limited kits, a deadline, and downtime at each move.

    Source: §I · p. 1

  • Method

    ReCo counts new target-and-background regimes, forecasts each home's future gain, and compares staying with moving after downtime.

    Source: §I · p. 2

  • Results

    ReCo outperformed fixed-dwell and count-based schedules in every setting and improved both F1 and MAE over the threshold rule in all four settings.

    Source: §VI · p. 7, §VIII · p. 8

  • Takeaways & Limitations

    ReCo's advantage reflects allocating collection days to more valuable homes and periods, rather than collecting more days alone.

    Source: §VI · p. 7, §VIII · p. 8

    The evaluation replays recorded data rather than field deployments, so installation, site access, and equipment failures are not reflected.

    Source: §VII · p. 7

Abstract

from arXiv · show

Many sensing tasks obtain training labels only by deploying instruments in the field. With a limited number of sensor kits, a collection deadline, and measurement downtime at every move, the collector must repeatedly decide whether to stay at the current site or relocate. We study this decision in non-intrusive load monitoring (NILM), which estimates the power drawn by individual appliances from a home's main meter and is trained on data from homes temporarily fitted with appliance-level sub-meters. In NILM, appliance usage varies with the appliance, season and climate, and the value of new data depends on how diverse the combinations of target operation and background load are. To address this, we propose a constraint-based relocation framework and instantiate it for NILM as ReCo (Relocation by Coverage gain). ReCo counts new operating regimes in a joint target-background feature space, forecasts each home's future gain from the data collected so far, and each night weighs the gain of staying against the gain of moving elsewhere after the downtime. In replayed deployments on the Plegma dataset under two kit counts and two downtime costs, ReCo outperforms fixed-dwell and count-based schedules and a threshold rule using the same metric in every setting. Its advantage is not explained by collecting more days alone and reflects allocating the days to more valuable homes and periods.

I. INTRODUCTION

Field data collection must balance limited kits, a deadline, and relocation downtime; in NILM, a useful stay-or-move decision must account for how target-appliance operation combines with background load. The paper frames this decision as constrained kit reuse and instantiates it with a regime-coverage gain for NILM.

I. INTRODUCTION

  • A limited kit count, collection deadline, and relocation downtime make the collector's nightly choice whether to stay or move a constrained sequential decision.Staying can spend time on redundant data, while moving can incur downtime and leave an informative site; the balance shifts as the deadline approaches.Source: §I · pp. 1–2
  • NILM estimates appliance power from a home's main meter, using appliance-level sub-meter data for training; useful data depend on diverse combinations of target operation and background load.Appliance usage varies by appliance, season, and climate, so the yield of a fixed stay can differ across homes and periods.Source: §I · p. 1

II. RELATED WORK

  • Prior NILM work allocates permanent meters, studies single-home collection or data scaling, or uses stopping criteria; these do not address reusable kits under kit, time, and relocation-cost budgets.Related work also considers sensor movement for environmental monitoring and offline sensor allocation, but not this dwell-time decision under relocation downtime.Source: §I · p. 1, §II · p. 2

I. INTRODUCTION

  • The framework compares the expected gain of staying one more day with relocating after downtime, adapting dwell times to site yield and the remaining budget.Its gain measure is replaceable; the NILM instantiation counts previously unseen target-appliance and background-feature regimes without training a model.Source: §I · p. 2
  • The empirical study compares ReCo with fixed-dwell, count-based, and threshold schedules under two kit counts and two downtime costs, alongside metric substitutions, ablations, and diversity scenarios.The diversity scenarios vary the downstream model, target appliance, or dataset.Source: §I · p. 2
  • The framework is intended for field-deployed tasks with synchronous labels and cross-site generalization, but the coverage metric requires redesign and validation for each domain.The reported empirical conclusions are limited to NILM data.Source: §I · p. 2

III. PROBLEM FORMULATION

The deployment formulation constrains each kit's visits by kit availability, a collection deadline, and downtime for relocation. Schedule comparisons are meaningful only under the same constraint settings.

A. Deployment constraints

  • Each campaign has K sensor kits, a deadline T, and relocation downtime c as its three deployment constraints.A kit may record at most K sites simultaneously across the campaign.Source: A. Deployment constraints · p. 2
  • The kit limit K bounds the number of sites recorded simultaneously.The campaign is characterized by K sensor kits.Source: A. Deployment constraints · p. 2
  • All collection must finish by the T-day deadline.The deadline is one of the campaign's three constraints.Source: A. Deployment constraints · p. 2
  • A kit incurs c days of downtime each time it moves to another site.Relocation downtime is a campaign constraint.Source: A. Deployment constraints · p. 2
  • A kit's schedule can include n visits, with dwell times d1 through dn, and a site may be revisited.The formulation expresses the constraints for each kit's visit sequence.Source: A. Deployment constraints · p. 2
  • The first installation incurs no downtime.Downtime is incurred when the kit relocates, not for its initial installation.Source: A. Deployment constraints · p. 2
  • Schedules are compared only under the same K, T, and c; changing K or c changes the budget or calendar, so absolute values are not comparable across settings.The paper does not rank schedules across different K or c settings.Source: A. Deployment constraints · p. 2

B. Objective

The objective is to choose a feasible schedule that maximizes expected downstream performance on evaluation sites. ReCo approximates that post-training objective with a nightly, data-side stay-or-switch decision, evaluated in replayed deployments.

B. Objective

  • The schedule objective is expected downstream metric performance on evaluation sites not used for collection, subject to the per-kit deployment constraint.The objective is evaluated after training, so the nightly rule uses data-side gain as a heuristic proxy.Source: B. Objective · p. 2

C. Replay evaluation protocol

  • The replay protocol cuts alternative site-and-day schedules from real recordings and completes them within the same calendar window.Schedules differ only in when they relocate; holding the window fixed does not rule out date effects because collected days differ.Source: C. Replay evaluation protocol · p. 3

IV. METHOD

  • ReCo processes candidate-home windows collected so far each evening and compares the expected gain of staying with the gain of moving after downtime.Its three steps are counting new regimes, forecasting each home's remaining gain, and making the stay-or-switch comparison.Source: §IV · p. 3

A. New-regime accounting

ReCo scores windows as new regimes when their target-operation and background features differ sufficiently from previously seen windows. It forecasts each home's remaining gain from usable-window, activity, and novelty estimates, with novelty declining as coverage accumulates.

A. New-regime accounting

  • Off windows are represented by background power and time of day, while run windows use run characteristics and background level; each is tested against previously seen windows in its feature space.Off windows use a fixed-radius match, whereas run windows use per-feature tolerances; newly counted regimes join the seen set immediately, and the two spaces are counted separately.Source: A. New-regime accounting · p. 3

B. Per-home forecasts of q, λ and p

  • ReCo estimates each home's usable-window share, active windows per day, and the probability that an active window is new relative to the global seen set.The underlying proportions use scheduled, valid, and active windows as trials; estimates are selected each evening using history available before that decision.Source: B. Per-home forecasts of q, λ and p · p. 3
  • The forecast gain combines expected new run regimes per day with weighted new off regimes.The run term uses usable share, active windows, and novelty rate; the off term uses its daily rate and weight.Source: B. Per-home forecasts of q, λ and p · p. 3
  • Predicted run novelty declines as coverage grows, while predicted off-regime yield declines with accumulated off exposure.The current evening's estimates are re-estimated nightly, and the decay applies to future days.Source: B. Per-home forecasts of q, λ and p · pp. 3–4

C. Stay or switch

ReCo compares the expected gain from one more day at the current home with the average gain from moving after downtime and staying at a feasible duration. Its forecast for an unvisited home is built from other homes’ replayed data and future-gain estimates.

C. Stay or switch

  • Each evening, ReCo compares tomorrow’s expected gain from staying with the next home’s average gain over downtime plus a feasible stay, then moves only when moving is better.The candidate stay length maximizes cumulative gain divided by downtime plus stay length, subject to the remaining budget.Source: C. Stay or switch · p. 4
  • For an unvisited home, ReCo constructs a default gain curve by replaying first stays from visited homes and forecasting beyond observed days.Replay excludes the donor’s later data, evaluation homes, and future data; donors receive equal weight day by day.Source: C. Stay or switch · p. 4
  • For a previously visited next home, ReCo forecasts its gain from that home’s own history; longer downtime lowers the average switch gain and can keep the kit in place longer.The donor replay scores new run regimes plus ω times new off regimes, so off regimes affect both stay and switch values.Source: C. Stay or switch · p. 4

V. EXPERIMENTS

The experiments use replayed deployments to evaluate washing-machine NILM under shared collection and training conditions. The setup covers the data protocol, downstream model, and implementation settings.

A. Data and evaluation protocol

  • Eleven usable Plegma homes are divided into four folds, each with two evaluation homes and nine candidate homes, and collection runs for 120 days.Evaluation homes are excluded from collection, routing, forecasting, tuning, and training; settings use three seeds and shared random visiting orders.Source: A. Data and evaluation protocol · p. 4

B. Features and downstream model

  • All methods use the same temporal convolutional disaggregation model, inputs, training steps, on-threshold, and seeds, so only the collection calendar differs.The dilated residual TCN has a gated power head and about 3.2M parameters, mapping 721 samples to the central 121 points.Source: B. Features and downstream model · p. 4

C. Implementation settings

  • The NILM instance uses the washing machine’s implementation settings, with event and regime settings fixed before evaluation.Table I identifies the implementation-settings scope, but the supplied passage does not enumerate its values.Source: C. Implementation settings · p. 4, Table I · p. 5

D. Baselines

ReCo is evaluated against fixed-dwell, count-based, and threshold schedules, then compared with a per-fold best baseline and an alternative gain metric. Across the main settings it outperforms every baseline, while calendar and alternative-metric results qualify that advantage.

D. Baselines

  • Fixed-7/14, Count-5/10, and a threshold of 1.00 new regimes per day provide calendar, count, and same-metric switching references.The threshold rule averages gain over the last three valid dwell days and waits until three such days are available.Source: D. Baselines · p. 5

E. Main results

  • In all four settings, ReCo achieves higher on-state F1 and lower MAE than every baseline; it beats the per-fold best in F1 in all four and in MAE in three.The per-fold best is selected after the fact separately for each metric and is not available in a real deployment.Source: E. Main results · p. 5, Figure 3 · p. 5, Table II · p. 6

F. Collection calendars

  • ReCo relocates less often than Fixed-7, Count-5, and the threshold rule, while collecting the most valid windows per collected day in every setting.Its advantage is not explained by collecting more days alone: Fixed-14 and Count-10 collect about as many or more device-days and similar or more windows, yet are less accurate.Source: F. Collection calendars · p. 5, Table III · p. 6

G. Alternative gain metric

  • With the same constraints and folds, Surprise yields lower F1 than ReCo in three settings and higher MAE in all four.Surprise turns Bayesian surprise of aggregate-signal step events into a stay-or-switch gain metric.Source: G. Alternative gain metric · pp. 5–6

H. Component ablation

The ablation compares ReCo with removals of availability correction, home novelty, and adaptive shrinkage or forgetting. Removing home novelty produces the largest F1 decline in most reported cells.

H. Component ablation

  • Removing per-home novelty causes the largest F1 drop in three of four cells, by 0.090–0.097, with MAE increases of 1.57–1.72 W.This ablation replaces the per-home novelty rate with a pooled cross-home rate.Source: H. Component ablation · p. 6
  • Removing availability correction lowers F1 by 0.057–0.084 in all cells, while MAE rises in three cells but falls by 0.45 W at K=1, c=3.The removal sets forecast availability to one and leaves history handling unchanged.Source: H. Component ablation · p. 6
  • Fixing shrinkage and forgetting has the smallest F1 effect, with declines of 0.005–0.053, but consistently raises MAE by 0.85–0.97 W.The fixed settings are no longer reselected by forecast loss.Source: H. Component ablation · p. 6

I. Diversity scenarios

The diversity scenarios tested whether ReCo’s advantage carried across changes to the model, target appliance, and dataset. ReCo’s F1 exceeded the strongest baseline in ten of twelve combinations, with lower MAE in both remaining combinations.

I. Diversity scenarios

  • The scenarios changed one factor at a time: the downstream model, target appliance, or dataset.They used SGN, an air conditioner, or REFIT with the washing machine as target.Source: §I · p. 6
  • Comparisons used the best of Fixed-14, Count-10, and the threshold rule separately for F1 and MAE in each fold.Results were averaged over three seeds and four folds; reported differences were mean ± SD over folds.Source: §I · p. 7
  • ReCo’s F1 exceeded the strongest baseline in ten of twelve combinations; in the other two, both on REFIT, it was lower by about 0.045 while MAE remained lower.Its MAE was higher in one case: SGN, K=1, c=3.Source: §I · p. 7

VI. DISCUSSION

The discussion attributes ReCo’s results to both regime coverage and its stay-or-switch decision, while noting that their benefits vary by setting. Ablations identify per-home novelty and valid-data availability as important modeling choices.

VI. DISCUSSION

  • Coverage alone does not guarantee a better calendar: at K=1, Fixed-14 is slightly ahead of the threshold rule in F1, and also in MAE when c=3.The threshold rule collects the fewest valid windows, yet its F1 exceeds Fixed-7 and Count-5 in every setting.Source: §VI · p. 7
  • ReCo improves both F1 and MAE in all four settings, but the margin varies and is smallest for F1 at K=2, c=3.The improvement over the threshold rule reflects both per-home forecasts and the cost-aware comparison, not downtime alone.Source: §VI · p. 7
  • Per-home novelty modeling gave the most consistent ablation benefit because homes differ in how quickly they exhaust new regimes.Ignoring valid-data availability distorts the comparison between staying and switching; online shrinkage and forgetting adaptation had a smaller, mainly MAE, effect.Source: §VI · p. 7
  • Future work could improve the gain measure, select the next home actively, coordinate kits, handle multiple targets, model monetary and travel costs, and test field deployments.Suggested gain improvements include weighting toward the final metric, using uncertainty or expected error reduction, or learning from past campaigns.Source: §VI · p. 7

VII. LIMITATIONS

The evaluation replays recorded NILM data rather than deploying kits in the field, and the forecast simplifies dependencies and is less reliable on first visits with little history. The conclusion reports ReCo’s performance in replayed deployments and describes its coverage-based, per-home gain forecast.

VII. LIMITATIONS

  • The evaluation replays recorded data rather than field deployments, so installation, site access, and equipment failures are not reflected.The multiplicative forecast also simplifies factor dependence and is less reliable on a first visit with little history.Source: §VII · p. 7
  • The empirical work focuses on NILM; applying the framework to other sensing-based collection tasks remains to be explored.The framework is intended for field deployment with synchronous labels and a need to generalize across sites, but each domain requires a redesigned and validated coverage metric.Source: §VII · p. 7

VIII. CONCLUSION

  • The framework compares staying with moving after downtime, treating the gain measure as a replaceable input; ReCo instantiates it for NILM with per-home forecasts of new target-background regimes.The conclusion says the collection days were allocated to more valuable homes and periods.Source: §VIII · p. 8
  • In replayed deployments on real household data, ReCo outperformed fixed-dwell and count-based schedules under every kit-count and downtime combination, and stayed at or near the hindsight-best baseline.The reported advantage largely carried over when the model, target appliance, or dataset changed.Source: §VIII · p. 8
Loading 2609.35075v1…