A Data-Driven Framework for Predictive Pre-Alarm Fire-Risk Modeling: Sensor Fusion, Calibrated Machine Learning, and Risk-State Representation

Ensuring fire safety in densely populated tall buildings is a critical urban safety concern. Although conventional smoke alarms are tested for response to prescribed fire and nuisance alarms, the alarm is triggered by sensor response criteria and not based on a probabilistic forecast of future risk [1, 2]. Evacuation performance in high-rise and other complex-built environments is affected by building use, the availability of the egress components, human behaviors, and route selection [3, 4].

Smart buildings, which are the subject of the trend and are equipped with massive amounts of sensors, streaming live data, and onboard computer processing, could be used to move the fire detection process from just a reactive type of detection to a more proactive probabilistic evaluation. The state-of-the-art research and multi-sensor studies indicate active research in the areas of early-stage warning, Internet of Things (IoT)-based sensing, and data-fusion-based fire monitoring [5-7]. These methods could help to provide a more comprehensive assessment earlier than relying solely on a sensor threshold; however, any impact on safe egress time must be validated separately, at full-scale and occupant-centered, across all of these approaches. A primary problem is transforming a large number of dynamic sensor streams into valid, meaningful, and actionable risk estimation.

A method to keep a representation of physical systems up to date with sensors has been gaining traction: digital twin technology [8, 9]. In this work, the digital-twin aspect is restricted to a data-driven time-resolved representation of the estimated fire-risk state. Maps calibrated machine learning (ML) results to states and recommendations for decision support. This implementation is not a fully bidirectional cyber-physical digital twin [10], not a full Building Information Modeling (BIM) integration, or even a live building-management-system connectivity [11, 12].

Ensemble tree-based classifiers have been found to perform well on structured tabular data and/or with nonlinear feature interactions [13]. For tabular time-series features, Chen and Guestrin [14] present a new and very expressive model called Extreme Gradient Boosting (XGBoost) that is efficient and can be used to get impressive discriminative performance while providing interpretable feature importance. Some post-hoc calibration methods, like isotonic regression, can enhance the fit between the predicted probabilities and the actual outcomes [15].

1.1 Research gap

Research on early-stage fire warning, IoT-enabled sensing, multi-sensor fire monitoring, and building fire-risk simulation has expanded [5-7, 16, 17], alongside digital-twin research for buildings and infrastructure systems [12, 18]. However, there are crucial gaps that still need to be addressed. First, many prior studies focus on the detection accuracy or warning performance, but do not provide information on the distributions of pre-alarm threshold-crossing-to-alarm intervals at the experiment level, which are key to risk-management interpretation. Second, computational models such as a physics-based approach based on the Fire Dynamics Simulator (FDS) and BIM coupling models are often employed in second workflows related to fire safety, and these models have a great computational demand for repeated or near real-time use [19, 20]. Third, translating the laboratory-scale or bench-scale information into room-scale and operational building-scale challenges is also known [5, 6].

This paper fills these gaps with a proof-of-concept data-driven framework that integrates multi-sensor temporal feature learning, learning under domain shift at the level of probability, and visualization of discrete risk states. This analysis compares descriptive pre-alarm Threshold Crossing timing for independently held-out room-scale National Institute of Standards and Technology (NIST) experiments, and investigates the ability to convert a risk score into an interpretable monitoring state. It is not fully validated in a real-size high-rise building or implemented as a full cyber-physical digital twin.

3.1 Data sources, experimental datasets, and preprocessing

We used data from the publicly available NIST Smoke Alarm Performance dataset, which provides smoke-box and room-scale experimental smoke-alarm performance data relevant to fire development and alarm activation [1]. The smoke-box experiments were used for model development, whereas the room-scale experiments formed the target domain for probability calibration and independent hold-out evaluation under larger-scale experimental conditions.

The supervised target was defined as a fixed-horizon future-alarm indicator. For each time-indexed observation, alarm_after_10s represents the conventional alarm state 10 time steps ahead, y(t) = alarm(t + 10), computed independently within each experiment using a group-wise forward shift of the raw alarm column. Under the nominal one-second sampling interval, this corresponds to an approximately 10-s forecasting horizon. Smoke-box observations for which a future target value was unavailable were excluded, accounting for the reduction from 184,365 raw observations to 179,795 modeling observations. This formulation embeds the forecasting horizon directly within the supervised-learning task.

After construction of the future-alarm target and alignment of the common variables, the smoke-box modeling dataset contained 179,795 time-indexed observations from 457 experiments, including 14,919 positive target observations (8.3%). The room-scale dataset contained 18,700 observations from 18 experiments, including 11,778 positive target observations (63.0%). Seven room-scale experiments (7,034 observations) were reserved for probability calibration, and 11 experiments (11,666 observations) formed the independent hold-out set. Table 1 summarizes the analysis datasets.

Each experiment file was brought into a consistent tabular form, with time in seconds and an experiment ID to uniquely identify observations. To ensure consistency between the training and external validation domains, only variables common to both were considered. A total of four features were selected: (i) transmittance percentage, (ii) smoke obscuration per foot, (iii) microphone (MIC) signal response, and (iv) carbon monoxide (CO) concentration. The binary microphone indicator was removed due to inconsistent information across domains, which would have affected cross-domain validation.

Table 1. Summary of experimental datasets used for model development, calibration, and evaluation

3.2 Temporal feature engineering

Fire development is dynamic; therefore, instantaneous measurements were augmented with causal temporal descriptors computed separately within each experiment. The final 68-predictor specification comprised four raw sensor variables; first differences at lags 1 and 3; short-term slopes over 5 and 10 observations; and rolling mean, minimum, maximum, and standard deviation over trailing windows of 5, 10, and 20 observations. These transformations represent recent signal level, direction, and variability without requiring a recurrent architecture.

The trailing windows of 5, 10, and 20 observations balance responsiveness with stability. Short windows preserve rapid changes, whereas longer windows capture accumulation and attenuate brief fluctuations. The window lengths define local feature context and do not define the separate descriptive threshold-crossing-to-alarm interval.

First-order difference feature:

$\Delta x_i(t)=x_i(t)-x_i(t-1)$        (1)

where, xᵢ(t) = value of sensor variable i at time t; Δxᵢ(t) = first-order temporal difference (rate of change) of sensor i at time t.

Short-term slope feature:

$S_i^w(t)=\frac{x_i(t)-x_i(t-w)}{w}$        (2)

where, w ∈ {5, 10} observations for the short-term slope features; Sᵢⁿ(t) = mean rate of change of sensor i over the preceding w time steps.

Rolling mean feature:

$\mu_i{ }^w(t)=\frac{1}{w} \sum_{k=0}^{w-1} x_i(t-k)$        (3)

where, μᵢⁿ(t) = rolling arithmetic mean of sensor i over window w; k = lag index ranging from 0 (current time) to w−1 (oldest in window).

Rolling minimum, maximum, and standard deviation:

$\begin{aligned} & \min _i{ }^w(t)=\min \left\{x_i(t-k): k=0, \ldots, w-1\right\} \\ & \max _i{ }^w(t)=\max \left\{x_i(t-k): k=0, \ldots, w-1\right\}\end{aligned}$             (4)

where, σᵢⁿ(t) = rolling standard deviation, capturing local signal variability.

3.3 Model development, probability calibration, and evaluation design

The XGBoost classifier was chosen as the main classifier, as it can model nonlinear interactions between temporal predictors in structured multi-sensor data. The final model was trained with 700 boosting estimators, a maximum tree depth of 4, a learning rate of 0.03, a row subsampling rate of 0.85, and a column subsampling rate of 0.85, with a minimum weight of each child set to 5, and a regularization term of 2.0 (L2) and 0.1 (L1). The target distribution in the smoke-box was used to calculate scale_pos_weight to prevent class imbalance. Logistic regression, random forest, linear Support Vector Machine (SVM), and a calibration-derived multi-sensor threshold rule were applied as baseline comparisons with the evaluation design of the experiment.

Features were standardized using a z-score scaler fitted exclusively on the smoke-box training data. The fitted scaler was then applied without refitting to the room-scale calibration and hold-out subsets. The XGBoost trees were not fitted using room-scale observations; they were only fitted using the calibration subset for calibration of the probabilities. Baseline models, preprocessing, and calibration were developed by following scikit-learn compatible workflows [21].

Full configuration details for each comparator are as follows. The logistic regression was computed with the liblinear solver, balanced class weighting, a maximum of 4,000 iterations and default L2 regularization (C = 1.0). Random forest had a total of 500 trees, maximum depth was unlimited, minimum split was 10, minimum leaf was 5 and balanced class weighting was applied. Linear SVM was implemented with C = 1.0, balanced class weighting and max iterations = 30,000. The four per-sensor binary conditions (obscuration ≥ 0.070 %/ft, CO ≥ 0.000, MIC acoustic pressure ≥ 78.42 Pa, transmittance ≤ 99.66%) were combined to an average score, each sensor threshold was determined for every sensor by its 25th percentile (75th percentile for transmittance) of the sensor value in the positive-target calibration-subset observations and the threshold of the rule itself (0.55) was optimized by maximizing the F1-score over 21 evenly spaced candidate thresholds on the calibration subset. Each of the four learned models (logistic regression, random forest, linear SVM, XGBoost) was calibrated on the same seven experiments of the room-scale calibration set, with prefit calibration being applied to all of them. Logistic regression, random forest and XGBoost were prefit calibrated using isotonic regression, while linear SVM was prefit calibrated using sigmoid (Platt) calibration, owing to its non-probabilistic margin-based output and limited number of calibration experiments. Logistic regression, random forest, linear SVM and XGBoost were assessed at a common 0.50 classification threshold, while the sensor-rule baseline was evaluated at its independently determined 0.55 classification threshold.

Smoke-box and room-scale experiments differed in scale, target prevalence (8.3% versus 63.0%), and sensor-response characteristics. Isotonic regression was therefore fitted on the seven-experiment room-scale calibration subset to transform raw XGBoost outputs into calibrated risk scores. Calibration did not alter the fitted XGBoost trees. Because calibration estimates based on finite validation data may be sensitive to sampling variability, the resulting calibration should be interpreted cautiously [22].

The remaining 11 room-scale experiments formed the independent hold-out subset. All final performance metrics were calculated only on these hold-out experiments. The operating classification threshold, θ = 0.50, was specified before hold-out evaluation and was not retrospectively selected from hold-out results.

3.4 Evaluation metrics and exploratory threshold crossing timing analysis

Discrimination was quantified using the area under the Receiver Operating Characteristic–Area Under the Curve (ROC-AUC) and the area under the Precision-Recall–Area Under the Curve (PR-AUC). ROC analysis is widely used to assess classifier discrimination [23], whereas precision-recall analysis is particularly informative for imbalanced classification problems in which the positive class is of primary interest [24]. Threshold-dependent performance was assessed using accuracy, precision, recall, specificity, F1-score, and confusion-matrix counts.

Precision:

$Precision=\frac{T P}{T P+F P}$        (5)

where, TP = true positives (correctly predicted alarm states); P = false positives (non-alarm states incorrectly predicted as alarm).

Recall (Sensitivity):

$Recall=\frac{T P}{T P+F N}$         (6)

where, FN = false negatives (alarm states incorrectly predicted as non-alarm).

F1-score:

$F 1=\frac{2 \times \text { Precision } \text { × } \text { Recall }}{\text { Precision }+ \text { Recall }}$        (7)

Pre-alarm threshold crossing to alarm interval:

$\Delta T_{\text {lead}}=T_{\text {alarm}}-T_{\text {alert}}$       (8)

where, T_alarm = time of conventional alarm activation (seconds from experiment onset); T_alert = first time at which the calibrated risk score crosses the operating threshold; here, the term alert is shorthand for a score-threshold crossing and does not itself denote a correct fixed-horizon prediction.

A positive ΔT_lead indicates that the first threshold crossing occurred before conventional alarm activation, whereas zero or a negative value indicates a first crossing at or after alarm activation. The interval in Eq. (8) is used only as an exploratory descriptive timing quantity and does not alter classification correctness under the fixed-horizon supervised task. Under y(t) = alarm(t + 10), a pre-alarm threshold crossing occurring more than 10 s before conventional alarm activation corresponds to a negative target and is therefore a false-positive classification in the primary supervised evaluation. Accordingly, Eq. (8) is not interpreted as a verified fixed-horizon predictive lead.

3.5 Data-driven risk-state representation

A data-driven risk-state layer was constructed by augmenting each time-indexed room-scale observation with the calibrated model score R(t), a discrete visualization state, and a corresponding decision-support recommendation. This layer will not create synthetic fire scenarios, simulate ignition physics, keep a live connection with a physical building, or be controlled in any way. It is used to show an approximation of risk over time in the experimental data.

The visualization-state boundaries were not adopted as universal probability cut-offs. Only the room-scale calibration subset was used to compare candidate configurations prior to independent hold-out evaluation. Screening included: Representing all four states, avoiding a single-state dominance, ordered progression, maintaining pre-alarm occurrence of the Warning and Critical states, and minimizing unnecessary short-duration switching. The chosen boundaries (0.350, 0.500 and 0.725) were established prior to their application to the independent hold-out set. No hold-out observations or hold-out performance metrics were used for the boundary selection. These are data-derived visualization-state boundaries, and not fire-safety thresholds or regulatory action levels.

The screening procedure searched a candidate grid of 15 lower-boundary values from 0.100 to 0.450 (step 0.025) and 17 upper-boundary values from 0.550 to 0.950 (step 0.025), with the middle boundary fixed at 0.500, yielding 15 × 17 = 255 candidate configurations satisfying lower < 0.500 < upper. Each configuration was evaluated exclusively on the seven-experiment calibration subset and retained as feasible only if it satisfied four constraints: every state occupied at least 5% of calibration observations; no state occupied more than 70%; the Critical state was reached before alarm activation in at least 70% of alarm-producing calibration experiments; and the Warning state preceded or coincided with the Critical state in every experiment. Of the 255 candidates, 135 satisfied all four constraints. Feasible configurations were ranked lexicographically by (1) minimum transitions per 1,000 observations, (2) maximum of the minimum state-occupancy fraction, (3) maximum median state-run duration, and (4) maximum median Critical-state threshold crossing to alarm interval; the top-ranked configuration (0.350, 0.500, 0.725) was retained as the final boundary specification.

Relative to the equally spaced reference configuration, the selected configuration reduced total transitions by 64 (15.8%) and transitions per 1,000 observations from 57.58 to 48.48. It also increased the median state-run duration from 3.0 to 5.0 s and improved the minimum state occupancy from 14.19% to 16.02%, while preserving the critical pre-alarm threshold crossing rate at 0.714 and maintaining pre-alarm Warning-state occurrence. The selected configuration was therefore retained because it improved temporal stability and state representation without sacrificing calibration-subset pre-alarm behavior. The boundary comparison and state mapping are summarized in Tables 2 and 3, respectively.

The resulting state log is a derived time-resolved dataset containing the experiment identifier, timestamp, calibrated risk score, visualization state, and recommended response. It supports analysis of transitions, state occupancy, and the timing of elevated risk relative to conventional alarm activation. The framework is therefore described as digital-twin-oriented rather than as a complete physics-based or cyber-physical digital twin.

Table 2. Calibration-only comparison of the equally spaced reference configuration and the calibration-set-selected risk-state boundaries

Table 3. Calibration-set-selected risk-state visualization mapping based on calibrated score R(t)

The calibrated XGBoost model achieved ROC-AUC = 0.892 and PR-AUC = 0.908 on the independent room-scale hold-out set, indicating substantial discrimination despite the domain difference between smoke-box training and room-scale evaluation. A split and separate target-domain calibration subset is employed to minimize direct information flow between model development and final evaluation, but the number of room-scale experiments is limited, thus limiting the ability to generalize.

The model outputs were calibrated to the room-scale calibration subset by using isotonic calibration. The findings indicate that target domain calibration could be appropriate for cases where representative target data can be found; however, the present study does not generalize the results to operational buildings, and the small calibration sample size restricts general conclusions about transferability.

A calibration-only boundary comparison is used to select the final configuration with a basis of a reference configuration equally spaced. The chosen configuration reduced transition frequency and extended median state runs, and maintained calibration-subset pre-alarm state occurrence. These enhancements are related to temporal stability and state occupancy but are limited to seven calibration experiments and are not definitive state thresholds for other sensors, fuels, or buildings.

At θ = 0.50, the recall, precision, and specificity of XGBoost were 0.924, 0.816, and 0.664, respectively, and the number of false-negative and false-positive rows were 545 and 1,501, respectively. The sensor rule yielded a smaller number of false-positive rows, but a significantly higher number of false-negative rows (1,537) compared with linear SVM (4,236 false-positive rows, specificity 0.051). The comparison is thus not a general superiority of one approach over the others but rather displays different error profiles.

The baseline comparison does not assert that one or the other is better in every situation, but rather helps to be nuanced in interpreting the findings. XGBoost showed the best discrimination and F1-score; linear SVM had the highest recall but more rows of false-positive burden; and the sensor-rule baseline had the best precision, specificity, and overall accuracy, but with more rows of false positives. The trade-off will be specific to each operation, depending on how much the missed risk and nuisance messages will cost the operation, and needs to be determined in advance.

The issue of discrimination was also discussed, because rolling obscuration and CO features also had a role to play. The feature importance analysis is not causal and does not indicate resilience to sensor failure.

This dataset includes the presence of local temporal context, with the relative importance of temporal statistics over instantaneous values. Nevertheless, mechanistic interpretations should be taken with a pinch of salt, as it is not a causal combustion process that can be determined by gain-based feature importance.

The results of the 11 experiments show that, in these controlled room-scale experiments, calibrated-score threshold crossings occurred before conventional alarm activation in 9 experiments, with a median descriptive threshold crossing to alarm interval of 214.0 s. This timing statistic is separate from, and should not be interpreted as, the fixed 10-s supervised forecasting horizon. The width of the confidence interval was broad, and the result is not intended to be extrapolated to gains in evacuation time or operational benefit for occupied buildings.

There was a high degree of fuel-experimental condition heterogeneity, with descriptive threshold-crossing-to-alarm intervals ranging from 17 to 819 s. Long intervals can indicate early elevation of the calibrated score relative to later conventional alarm activation, but they do not establish correct fixed-horizon forecasts or practical usefulness and do not demonstrate avoidance of "alarm fatigue" without non-alarm trials and an independently verified response policy.

The two late-detected synthetic-foam experiments and the huge range in experiments highlight the need for more fuel types, non-fire experiments, more independent experiments, and external validation. It is therefore recommended that further investigation be done before making conclusions at the deployment level for the present findings.