LITF-PA-2026-160 · Severe Weather / Distributed Sensing / Edge AI

System and Method for Tornado Early Warning Using Distributed Smartphone Barometric Infrasound Sensing with Edge-Deployed Vortex Signature Classification and Multi-Device Triangulation

Defensive Prior Art Disclosure
Smartphone detecting infrasound pressure waves radiating from a distant tornado across plains at dusk
Prior Art Notice: This document is published as defensive prior art under 35 U.S.C. § 102(a)(1). The inventions described herein are dedicated to the public domain as of the publication date above. This disclosure is intended to prevent the patenting of these concepts by any party.

Abstract

Disclosed is a system and method for tornado early warning that repurposes the barometric pressure sensors in participating consumer smartphones as a distributed infrasound array. Tornadoes radiate infrasound with a fundamental frequency in the 0.5 to 10 Hz band, detectable tens of kilometers from the vortex, in some documented cases minutes before touchdown. During National Weather Service severe weather watches, participating phones switch their barometers into a high-rate sampling mode (25 to 100 Hz), run an on-device spectral classifier that distinguishes tornadic vortex signatures (a fundamental tone with linearly spaced overtones consistent with vortex-core resonance) from non-tornadic convection, wind buffeting, and indoor pressure artifacts, and transmit compact event reports containing spectral features, timestamps, and coarse location to a fusion service. The fusion service cross-correlates detections across multiple phones to estimate source bearing and range via time-difference-of-arrival, gates alerts against radar-derived rotation tracks to suppress false alarms, and issues targeted warnings to phones in the projected path. The system converts the existing installed base of barometer-equipped smartphones into a dense, zero-hardware-cost tornado sensor network.

Technical Field

This invention relates to severe weather detection and public warning systems, specifically to the use of distributed consumer smartphone barometric sensors for infrasound-based tornado detection, on-device acoustic signature classification, multi-sensor source localization, and targeted alert dissemination.

Background

Tornadoes kill an average of 70 to 80 people per year in the United States and cause approximately $3 billion in annual damage (NOAA Storm Prediction Center long-term averages). Warning performance remains limited: for the five-year period ending 2018, the national average tornado warning lead time was 8.6 minutes with a probability of detection of 59% and a false alarm rate of 70% (Elbing et al., NOAA repository). Radar-based detection struggles with tornadoes that form quickly, occur beyond effective radar range or below the radar horizon, or develop in regions with sparse low-level radar coverage such as the southeastern United States.

Tornado infrasound is a well-documented physical phenomenon. Tornadoes emit infrasound, defined as acoustic energy below 20 Hz, with a fundamental frequency consistently reported in the 0.5 to 10 Hz range (Bedard, 2005; reviewed in Allen et al., AMT 2022), where smaller vortex core diameters produce higher fundamental frequencies. Elbing et al. (2019) recorded a small Oklahoma tornado at 18.7 km range and observed a 75 dB spectral peak near 8.3 Hz, 18 dB above pre-tornado levels, with linearly spaced overtones at approximately 18, 29, 36, and 44 Hz; elevated infrasound began approximately 7 to 8 minutes before the verified tornado report. Infrasound array deployments during the spring 2018 severe weather season in the southeastern United States obtained accurate bearings on tornadoes at ranges exceeding 100 km, with the dominant band of coherent infrasound between 2 and 6 Hz (JASA, 2024). Weak atmospheric absorption at these frequencies and downward-refracting atmospheric ducts allow long-range passive detection. The empirical frequency-to-core-diameter relationship fn = (4n+5)c/4d (Abdullah, 1966) provides a physical basis for estimating vortex size from the measured fundamental.

Consumer smartphones increasingly carry barometric pressure sensors (Bosch BMP280/BMP380/BMP390 series, STMicro LPS22HH, and equivalents) originally included for altitude and indoor positioning. These MEMS sensors support output data rates from 25 Hz to over 100 Hz, sufficient to resolve the 0.5 to 10 Hz tornadic infrasound band. Prior work has demonstrated the sensing principle without building a warning system:

  • RedVox Infrasound Recorder: A smartphone application that records infrasonic pressure via the internal microphone and barometer and streams recordings to a cloud server for geophysical research (Google Play listing). Sandia National Laboratories evaluated a Samsung S10 running the RedVox app as a low-cost infrasound sensor package (Slad and Merchant, 2021). RedVox is a passive data-collection tool; it performs no automated tornado classification, no multi-device triangulation, and no alerting.
  • PressureNet (University of Washington): A crowdsourced smartphone pressure network for improving short-term numerical weather forecasting (Mass et al., 2013). PressureNet samples at synoptic rates (minutes between readings) for mesoscale data assimilation. It does not sample at infrasound rates, does not detect acoustic signatures, and does not address tornadoes.
  • Dedicated infrasound arrays: Research arrays (e.g., the 2018 southeastern US deployments) use purpose-built microbarometers costing thousands of dollars per station with station spacing of tens of kilometers. They demonstrate the detection physics but cannot scale to the density needed for neighborhood-level warning.

The gap in the art is a complete automated warning system that: (a) opportunistically switches phone barometers into infrasound-rate sampling during severe weather threats; (b) classifies tornadic vortex signatures on-device in real time, rejecting wind noise and indoor artifacts; (c) fuses detections across many phones to localize the vortex by time-difference-of-arrival; (d) cross-checks against radar rotation data to control the false alarm rate; and (e) delivers targeted alerts to people in the projected damage path. No existing system combines these elements.

Detailed Description

1. Threat-Gated High-Rate Barometric Sampling

Continuous 50 Hz barometer sampling on every phone would drain batteries and generate mostly useless data. The system therefore operates in two modes. In standby mode, the phone samples pressure at 1 Hz or less for altitude and weather purposes, as it already does. When the phone's location falls inside a National Weather Service tornado watch polygon, severe thunderstorm watch with tornado-possible tagging, or a convective outlook with elevated tornado probability, the system switches to surveillance mode: the barometer is driven at 25 to 100 Hz output data rate for the duration of the watch plus a configurable buffer (default 30 minutes after watch expiration).

Surveillance mode is further duty-cycled by a low-cost trigger: the phone computes short-term RMS pressure fluctuation energy in the 0.5 to 15 Hz band at 1 Hz cadence using an integer-arithmetic IIR filter chain (power cost under 1 mW on a modern SoC). Full spectral analysis engages only when band-limited energy exceeds an adaptive threshold set at 6 dB above the trailing 10-minute noise floor. Indoor phones behind closed windows still observe infrasound; building attenuation at 2 to 8 Hz is typically 5 to 15 dB, which the adaptive threshold absorbs because it is referenced to each phone's own noise floor.

2. On-Device Vortex Signature Classifier

When the energy trigger fires, the phone captures a 60-second rolling window of pressure data and computes a 0.25 Hz-resolution power spectral density via Welch's method (8-second segments, 50% overlap, Hann window). A two-stage classifier then evaluates the spectrum:

Stage 1: Tonal comb detector. The detector searches the 0.5 to 12 Hz band for a fundamental peak exceeding 10 dB above the local spectral background, then tests for overtones at integer-related multiples consistent with the linear overtone spacing reported for tornadic vortices (Elbing et al. observed overtones scaling approximately linearly with mode number). A harmonic comb score is computed as the geometric mean of the signal-to-background ratios at the fundamental and the first three predicted overtone frequencies. Non-tornadic convection produces broadband infrasound without a stable tonal comb; this stage rejects it.

Stage 2: Vortex vs. clutter neural classifier. A compact 1D convolutional network (3 convolutional layers, 32/64/64 channels, kernel size 7, followed by global average pooling and a 2-layer MLP; approximately 45,000 parameters, 180 KB quantized INT8) takes the 0.5 to 50 Hz log-magnitude spectrum plus the comb score and outputs P(tornadic vortex) versus P(clutter). Clutter classes in training include: gust-front broadband rumble, HVAC compressor cycling (narrowband 7 to 30 Hz with strong 60 Hz mains harmonics, a reliable reject feature), road/rail traffic, aircraft, door slams and indoor transients, and elevated wind buffeting. Training data combines published tornado infrasound recordings (Elbing et al. 2019 spectra, Frazier et al. 2014 Oklahoma recordings), the RedVox community archive of storm recordings, and synthetic vortex signatures generated by filtering broadband noise through the Abdullah resonance model with randomized core diameters (30 to 1500 m) and ranges (5 to 150 km) with atmospheric absorption applied. Target operating point: 90% recall on tornadic signatures at under 0.5 false triggers per phone per watch.

3. Accelerometer-Coherent Wind Noise Cancellation

The dominant noise source for a phone barometer is wind buffeting on the handset and its immediate surroundings, which produces large 0.1 to 5 Hz pressure fluctuations that overlap the tornadic band. The phone's accelerometer and gyroscope observe the mechanical component of this buffeting. An adaptive LMS filter (32 taps, updated at the 25 Hz sensor rate) models the transfer function from the 3-axis acceleration magnitude signal to the pressure signal and subtracts the coherent component. Because true infrasound from a distant vortex produces negligible phone acceleration while local wind produces strongly correlated acceleration and pressure, the canceller is expected to suppress wind-induced pressure variance by 8 to 14 dB based on the coherence differential between local buffeting and distant sources, while attenuating genuine distant infrasound by less than 1 dB. Phones reporting sustained accelerometer RMS above a motion threshold (phone in a moving vehicle, in a pocket during running) are down-weighted in fusion rather than discarded, since vehicle-mounted phones still contribute.

4. Multi-Device Coherence and TDOA Triangulation

A single phone detection is suggestive; a coherent detection across many phones is actionable. When a phone's classifier output exceeds 0.8, it transmits an event report to the fusion service containing: a 64-bin spectral magnitude vector (0.5 to 16 Hz, 0.25 Hz resolution, 8-bit log-compressed), the classifier score, GPS coordinates truncated to approximately 100 m precision, a timestamp disciplined to the cellular network clock (typical accuracy better than 100 ms; NTP-disciplined where cellular time is unavailable), and a device capability descriptor (barometer model, sampling rate, indoor/outdoor inference from ambient light and WiFi context).

The fusion service clusters event reports in 5-minute sliding windows. For clusters of 3 or more phones within 60 km, it computes pairwise cross-correlations of the reported spectral vectors and of short raw waveform snippets (2-second, 25 Hz, uploaded only for phones in a cluster to bound bandwidth). Coherent clusters, defined as mean pairwise spectral correlation above 0.6 with consistent fundamental frequency within 0.5 Hz, proceed to localization. Time-difference-of-arrival is estimated from cross-correlation peak lags of the waveform snippets; with 25 Hz sampling the raw TDOA quantization is 40 ms (approximately 13.6 m range resolution at 343 m/s sound speed), refined to sub-sample precision by parabolic interpolation of the correlation peak. Hyperbolic multilateration over the phone positions yields a maximum-likelihood source location with an uncertainty ellipse. Bearing-only fallback uses the dominant eigenvector of the cluster's spatial covariance when TDOA geometry is degenerate (e.g., all phones along one road).

5. Radar Fusion Gating

To control the false alarm rate that has historically plagued tornado warnings, infrasound-derived source locations are cross-checked against operational radar data before public alerting. The fusion service queries the MRMS (Multi-Radar Multi-Sensor) rotation track product or equivalent azimuthal shear data for the cluster's location and time window. Alert tiers are assigned as follows:

  • Tier 1 (confirmed): Coherent infrasound cluster with localized source AND coincident radar-indicated rotation (azimuthal shear exceeding 0.01 s-1) within 5 km and 10 minutes. Triggers immediate targeted alert.
  • Tier 2 (probable): Coherent infrasound cluster with stable tonal comb persisting over 3 minutes but no radar rotation (possible below-horizon or rapidly forming vortex). Triggers advisory notification to emergency managers and a low-urgency public heads-up within 15 km.
  • Tier 3 (single-sensor): High-confidence single-phone detection with no corroboration. Logged for forecaster review; no public alert.

This gating directly addresses the documented 70% false alarm rate of the current warning system by requiring two independent physical observations (acoustic vortex signature plus radar rotation) before the highest-urgency alert.

6. Projected-Path Targeted Alerting

Upon Tier 1 confirmation, the system projects a damage-path polygon from the localized source using storm motion vectors from radar (default: 15-minute extrapolation with lateral uncertainty growing at 20% of forward distance, minimum half-width 1.5 km). Push alerts are delivered to participating phones inside the polygon and to the Wireless Emergency Alert interface for all phones in the affected counties. The alert payload includes estimated vortex location, estimated core diameter class derived from the measured fundamental via the Abdullah relation (small: under 100 m; medium: 100 to 500 m; large: over 500 m), detection confidence, and protective-action guidance. Alerts are refreshed or cancelled as the cluster evolves; a cluster that loses spectral coherence for 5 minutes triggers an all-clear for its polygon.

7. Privacy-Preserving Architecture

No audio-band data ever leaves the phone: the barometer cannot capture speech in any case, and only spectral feature vectors and short infrasound waveform snippets (below 16 Hz, containing no intelligible content) are transmitted, and only after a local detection. Locations are truncated to 100 m. Device identifiers are rotated daily. The on-device classifier runs entirely within the phone's sensor hub or application processor; no raw pressure time series is uploaded except the 2-second snippets used for TDOA within confirmed clusters, which are discarded after correlation.

8. Figures Description

  • Figure 1: System architecture showing participating smartphones in surveillance mode within a tornado watch polygon, event reports flowing to the fusion service, radar data ingest, and targeted alert polygons pushed back to phones in the projected path.
  • Figure 2: Representative power spectral densities comparing a tornadic vortex signature (fundamental near 8 Hz with linearly spaced overtones) against gust-front broadband rumble, HVAC narrowband interference with mains harmonics, and wind-buffeting noise before and after accelerometer-coherent cancellation.
  • Figure 3: TDOA multilateration geometry for a five-phone cluster, showing hyperbolic loci, the maximum-likelihood source location with uncertainty ellipse, and the projected damage-path alert polygon.
  • Figure 4: Alert tier decision flowchart from single-phone trigger through coherence clustering, TDOA localization, radar gating, and tiered alert dissemination.

Claims

  1. A system for tornado early warning, comprising: a plurality of consumer smartphones each containing a barometric pressure sensor; a surveillance mode that switches the barometric pressure sensor into high-rate sampling at 25 to 100 Hz when the smartphone's location falls within a severe weather watch polygon; an on-device spectral classifier that evaluates pressure spectra in the 0.5 to 50 Hz band for tornadic vortex signatures; and a fusion service that aggregates event reports from multiple smartphones and issues targeted alerts.
  2. The system of claim 1, wherein surveillance mode is gated by a low-power integer-arithmetic band-energy trigger that engages full spectral analysis only when 0.5 to 15 Hz pressure fluctuation energy exceeds an adaptive threshold referenced to each phone's trailing noise floor.
  3. The system of claim 1, wherein the on-device spectral classifier comprises a tonal comb detector that identifies a fundamental peak in the 0.5 to 12 Hz band with linearly spaced overtones, and a neural classifier distinguishing tornadic vortex spectra from clutter classes including gust-front rumble, HVAC cycling identified by mains-frequency harmonics, traffic, and indoor transients.
  4. The system of claim 1, further comprising an accelerometer-coherent wind noise canceller that adaptively subtracts the component of the pressure signal correlated with phone acceleration, suppressing local wind buffeting while preserving distant infrasound.
  5. The system of claim 1, wherein the fusion service clusters event reports in sliding time windows, computes pairwise spectral cross-correlation across reporting phones, and admits only coherent clusters with consistent fundamental frequency to localization.
  6. The system of claim 5, further comprising time-difference-of-arrival source localization via sub-sample-interpolated cross-correlation of short infrasound waveform snippets uploaded by phones within a coherent cluster, producing a maximum-likelihood vortex location with uncertainty ellipse.
  7. The system of claim 1, further comprising radar fusion gating that cross-checks infrasound-derived source locations against radar azimuthal shear data and assigns alert tiers requiring coincident radar-indicated rotation before issuing highest-urgency public alerts.
  8. The system of claim 1, further comprising projected-path alerting that extrapolates a damage-path polygon from the localized source using radar storm motion vectors and delivers alerts to phones within the polygon.
  9. The system of claim 1, wherein the on-device classifier estimates vortex core diameter class from the measured fundamental frequency via a vortex resonance frequency-to-diameter relationship and includes the estimate in the alert payload.
  10. The system of claim 1, wherein no audio-band or speech-capable data leaves the smartphone, transmitted event reports contain only spectral feature vectors with location truncated to approximately 100 m precision, and device identifiers rotate at least daily.
  11. A method for tornado early warning using consumer smartphones, comprising: switching smartphone barometers into 25 to 100 Hz sampling upon entry of the phone's location into a severe weather watch polygon; detecting band-limited infrasound energy exceeding an adaptive per-device threshold; classifying the resulting pressure spectrum on-device as tornadic vortex or clutter using a tonal comb detector and a neural classifier; cancelling wind buffeting via accelerometer-coherent adaptive filtering; transmitting compact event reports for high-confidence detections; clustering reports across devices by spectral coherence; localizing the vortex by time-difference-of-arrival; gating alerts against radar rotation data; and disseminating projected-path warnings to phones in the forecast damage area.
  12. The method of claim 11, further comprising downgrading single-sensor detections without cross-device corroboration to forecaster-review logging with no public alert, and escalating coherent multi-sensor clusters lacking radar corroboration to advisory-tier notification.

Implementation Notes

Deployable as an OS-level emergency-alert feature or as a background service within an existing weather application. Barometer hardware requirements: MEMS pressure sensor with 25 Hz or higher output data rate and RMS noise below 1 Pa in the 0.5 to 15 Hz band after averaging; satisfied by Bosch BMP380/BMP390, STMicro LPS22HH, and most flagship-phone barometers shipped since approximately 2018. Estimated 60 to 70% of smartphones in the US tornado belt carry a barometer.

Power budget in surveillance mode: barometer at 50 Hz draws under 1 mW; the integer band-energy trigger adds negligible load; full 60-second spectral classification at approximately 50 MFLOPS runs in under 200 ms on a mid-range SoC and executes only on trigger, expected fewer than 10 times per watch per phone. Total surveillance-mode incremental drain is under 2% of battery per watch day.

Network density requirement: reliable TDOA localization needs 3 or more reporting phones within approximately 60 km of the vortex. At 5% participation among barometer-equipped phones in a typical tornado-belt county (population 50,000, roughly 40,000 smartphones, 25,000 with barometers), approximately 1,250 participating phones per county provide ample density; even 1% participation (250 phones) satisfies the geometric requirement for most county-scale events.

Known limitations: phones inside well-sealed buildings experience 5 to 15 dB infrasound attenuation, reducing single-phone sensitivity but not cluster coherence; the system provides no protection against tornadoes whose infrasound is masked by extreme local wind noise at every nearby phone simultaneously; and detection range is asymmetric, favoring downwind geometries where atmospheric ducting enhances propagation. The radar gating tier is specifically designed so these limitations degrade to missed advisories rather than false public alarms.

Prior Art References

  1. Allen et al., Atmospheric Measurement Techniques 15, 2022: Infrasound measurement system for real-time tornado measurements; 0.5 to 10 Hz fundamental band (Bedard, 2005)
  2. Elbing et al., JASA 146, 2019: Infrasound from an Oklahoma tornado: 8.3 Hz peak 18 dB above background, overtones linear with mode number, elevated signal 7 to 8 minutes before verification
  3. Elbing et al., NOAA repository: Warning statistics: 8.6 min average lead time, 59% POD, 70% FAR
  4. JASA 2024: Southeastern US infrasound arrays: bearings on EF-0 to EF-2 tornadoes beyond 100 km, dominant coherent band 2 to 6 Hz
  5. AMS 2020: Infrasound propagation for tornado monitoring: downwind ducting efficiency, upwind attenuation
  6. RedVox Infrasound Recorder: Smartphone infrasound recording app (passive data collection, no classification or alerting)
  7. Slad and Merchant, Sandia 2021: Evaluation of Samsung S10 with RedVox as low-cost infrasound sensor
  8. Mass et al., University of Washington, 2013: PressureNet crowdsourced smartphone pressure for weather forecasting
  9. Abdullah, A.J., 1966: Vortex resonance relation fn = (4n+5)c/4d relating infrasound frequency to core diameter
  10. Frazier et al., 2014: High-fidelity acoustic recordings (0.2 to 500 Hz) of three Oklahoma tornadoes; beamforming for long-duration tornado tracking