Headset-mounted threat detection device using acoustic and radio-frequency sensors with machine learning and doppleraugmented localization

Abstract

A head-worn threat detection device comprises at least one acoustic sensor and at least one radio frequency sensor with a fixed spatial separation, a processor executing a machine learning model trained to detect and classify threats based on acoustic and radio frequency data, a localization module using sensor spacing and antenna directionality to localize threats, and an alert output. The localization module is further configured to determine at least one of direction and approximate distance by performing Doppler analysis on at least one of the acoustic data and the radio frequency data. Real-time alerts may be provided via auditory output of the headset.

Description

HEADSET-MOUNTED THREAT DETECTION DEVICE USING ACOUSTIC AND RADIO-FREQUENCY SENSORS WITH MACHINE LEARNING AND DOPPLER- AUGMENTED LOCALIZATIONDESCRIPTIONTechnical Field

[0001] The present disclosure relates to wearable threat detection and alerting. In particular, it relates to a headset-mounted or head-worn device that detects unmanned aerial vehicles (UAVs) or other defined threats using acoustic sensing and radio-frequency (RF) sensing, and that determines at least one of direction and approximate distance using a localization pipeline including Doppler analysis, and provides real-time user alerts.Background Art

[0002] Small UAVs have become widely available and may be used for benign purposes as well as for malicious surveillance, targeting assistance, or delivery of hazardous payloads. Portable detection systems are therefore desirable for individual users and small teams.

[0003] Many existing counter-UAV detection solutions are vehicle-mounted or basemounted, or require non-wearable sensors, limiting suitability for dismounted users. Other solutions may rely on a single sensing modality, which can reduce robustness in cluttered environments.

[0004] Documents such as US2023 / 0168951 Al, US2022 / 165165 Al, US2022 / 0036741 Al, and US2023 / 401943 Al describe various approaches to UAV detection, identification, tracking, and / or user notification, including wearable implementations and the use of RF signals and / or acoustic sensing.

[0005] There remains a need for a compact head-worn detection device that improves localization performance in realistic field conditions, while maintaining low power consumption and providing clear, immediate user alerts.Summary of the Invention

[0006] According to a first aspect, there is provided a headset-mounted threat detection device comprising a first head-worn module and a second head-worn module arranged with a fixed spatial separation, at least one acoustic sensor and at least one RF sensor disposed on at least one of the first head-worn module and the second head-worn module, a processor executing a machine learning model trained to detect and classify threats based on acoustic and RF data, a localization module using the fixed spatial separation and antenna directionality to localize threats, and a warning system configured to provide auditory alerts via the headset when a threat is detected.

[0007] In embodiments, the localization module is further configured to determine at least one of direction and approximate distance of detected threats by performing Doppler analysis on at least one of the acoustic data and the RF data, thereby augmenting localization with Doppler-derived information.

[0008] In embodiments, the device supports different head-worn form factors including overear headsets, helmet-mounted configurations, and earbud configurations. The device may be implemented as an integrated headset or as an add-on module that retrofits an existing headset.

[0009] In embodiments, the warning system provides user-distinguishable alerts, such as different tones, pulse rates, or spatialized audio cues, to indicate threat type, confidence, direction, and / or range class.

[0010] According to a second aspect, there is provided a method of operating a head-worn threat detection device, comprising acquiring acoustic data and RF data, detecting and classifying a threat using a machine learning model, estimating a threat bearing and / or distance using triangulation and Doppler analysis, and generating an alert for the user.

[0011] The disclosed device and method can improve user situational awareness and reduce reaction time in environments where line-of-sight detection is limited, such as at night, in fog, or in complex terrain.Brief Description of the Drawings

[0012] Figure 1 shows example head-worn form factors, including an over-ear headset, a helmet-mounted configuration, and an earbud configuration, with illustrative placement of acoustic sensors and RF antenna elements.

[0013] Figure 2 is a functional block diagram of a headset-mounted threat detection device including sensing, processing, localization, and alerting subsystems.

[0014] Figure 3 illustrates an example sensor geometry and reference frame used for localization, including a left-side sensor module and a right-side sensor module with a fixed baseline, and optional directional antenna characteristics.

[0015] Figure 4 is an example processing flow for threat detection and Doppler-augmented localization.Detailed Description of Embodiments

[0016] The following description is provided to enable a person skilled in the art to make and use the invention. Unless stated otherwise, features described in one embodiment may be used in combination with features of other embodiments. The scope of protection is defined by the claims.Definitions and Terminology

[0017] In this document, the term "head-worn" includes headsets, helmet-mounted accessories, and ear-worn devices such as earbuds.

[0018] The term "threat" includes a UAV and may include other moving hazards defined by a user or by a signature library, such as other aerial vehicles or projectiles, depending on implementation.

[0019] The term "localize" includes estimating at least one of a bearing (direction) and approximate distance or distance class. A distance class may be coarse (for example near / mid / far) while still providing operational value.[0019A] In embodiments, the fixed spatial separation is a baseline distance between the first head-worn module and the second head-worn module positioned on opposite sides of awearer’s head. The baseline distance may be known by design, by calibration, or by retrieving a stored baseline parameter associated with the headset form factor.[0019B] In embodiments, Doppler-derived information is used as a constraint or weighting term within the localization pipeline. For example, candidate bearing and / or distance-class hypotheses derived from time / phase / amplitude differences may be weighted or rejected based on whether the hypotheses are consistent with an estimated approach / recede indicator or range-rate trend from Doppler analysis over successive observation windows.

[0020] "Doppler analysis" includes determining a Doppler-derived parameter from a received acoustic or RF signal, such as a frequency-shift component, a range-rate estimate, or an approach / recede indicator.

[0021] "Machine learning model" includes classifiers and regressors implemented on embedded processors, including neural networks, support vector machines, decision trees, or hybrid models.Reference Signs List

[0022] 10 Head-worn device (headset / helmet / earbud)

[0023] 12 Left sensor module

[0024] 14 Right sensor module

[0025] 16 Acoustic sensor / microphone

[0026] 18 RF sensor / receiver

[0027] 20 Antenna element (omnidirectional or directional)

[0028] 22 Processor

[0029] 24 Machine learning model

[0030] 26 Triangulation / localization module

[0031] 28 Doppler analysis module (implemented in software and / or hardware within 22 / 26)

[0032] 30 Warning / alert output (speaker / earpiece / haptic output)

[0033] 32 Power supply (battery)

[0034] 34 Data interface (wired and / or wireless update interface)

[0035] 36 Signature database / library

[0036] 38 User input (optional)

[0037] 40 Environmental noise suppression / filtering (optional)System Overview

[0038] The device 10 is a head-worn platform that places sensors at an elevated position relative to the user's body, which can improve exposure to acoustic and RF signals. In an over-ear headset embodiment, the left module 12 and right module 14 may be located near respective earcups. In an earbud embodiment, the modules may be integrated into respective earbuds or into a connecting band.

[0039] Each module may include one or more acoustic sensors 16 and / or one or more RF sensors 18 coupled to one or more antenna elements 20. The RF sensing may target frequency bands relevant to known UAV control links, telemetry links, video downlinks, Remote ID transmissions, or other RF emissions, depending on jurisdiction and application.

[0040] The processor 22 acquires acoustic data and RF data, optionally performs preprocessing such as filtering and noise suppression 40, and executes a machine learning model 24 configured to detect and classify threats by recognizing patterns in one or both data domains. The model may be trained on labeled examples of threat and non-threat signals.

[0041] When a threat is detected or suspected, a localization module 26 estimates at least one of a bearing (direction) and approximate distance. The localization uses the known geometry of the sensor baseline (fixed spatial separation between modules 12 and 14) and may use directionality of antenna elements 20. In embodiments, the localization further uses Doppler analysis 28 on the acoustic and / or RF data to obtain Doppler-derived information that augments localization.

[0042] The warning output 30 provides an alert to the user. In a headset embodiment, the alert may be audio delivered through the headset speakers. In other embodiments, the alert may include haptic output and / or visual output. The alert may encode direction, threat type, confidence, and / or a range class.Sensing Subsystem

[0043] The acoustic sensing may be implemented using one or more microphones positioned on each side of the head-worn device. Multiple microphones may be used per side to improve directionality, wind noise reduction, and robustness against occlusion. Acoustic preprocessing may include band-pass filtering, spectral whitening, beamforming, and suppression of wind and speech components.

[0044] The RF sensing may be implemented using an RF receiver coupled to one or more antenna elements. The antenna elements may be omnidirectional, directional, or implemented as phased arrays. The RF receiver may include one or more front-end filters, amplifiers, and frequency conversion stages appropriate for target bands. In some embodiments, the receiver monitors multiple bands sequentially or in parallel.

[0045] In some embodiments, the RF sensing is configured to detect characteristic emissions associated with UAV operation, such as command-and-control link characteristics, telemetry patterns, or other periodic transmissions. In some embodiments, the RF sensing includes detection of Remote ID transmissions where legally permissible and where such signals are present.

[0046] The acoustic and RF subsystems may be time-synchronized by the processor such that features extracted from each domain are aligned in time for fusion and localization.Acoustic Signal Processing

[0047] In some embodiments, acoustic preprocessing includes an initial voice suppression stage to reduce sensitivity to user speech and nearby friendly voices. Voice suppression may be implemented using spectral subtraction, adaptive filtering, or a classifier trained to distinguish human speech from UAV signatures.

[0048] In some embodiments, wind noise reduction is performed using multi -mi crophone correlation and / or a wind-noise model. Wind noise reduction can improve detection performance in outdoor environments where wind is present.

[0049] In some embodiments, acoustic feature extraction includes computing a short-time Fourier transform, mel-frequency filterbank energies, cepstral coefficients, harmonic-to-noise ratios, and modulation spectra. The specific features may be selected based on the targeted threat types.

[0050] In some embodiments, the device uses a two-stage acoustic detector, where a lightweight detector screens for candidate events and a higher-accuracy model is executed only for candidates.

[0051] In some embodiments, the acoustic pipeline includes a directionality stage that uses multiple microphones per side to improve separation of external sound sources from internal leakage and reflections.RF Signal Processing

[0052] In some embodiments, RF preprocessing includes energy detection, spectral occupancy estimation, and identification of intermittent burst patterns. The device may compute features such as bandwidth, burst periodicity, symbol rate indicators, and channel hopping behavior where applicable.

[0053] In some embodiments, the RF receiver operates in a scanning schedule across multiple bands. The schedule may be adapted based on observed activity to focus on bands with higher likelihood of threat emissions.

[0054] In some embodiments, the device uses an RF front-end with selectable filters and gain stages to maintain sensitivity while limiting overload from strong nearby transmitters.

[0055] In some embodiments, chip antennas or antenna boosters are used to achieve compact multiband performance. Antenna response characteristics may be stored and used by the localization module to interpret received power differences as directional cues.

[0056] In some embodiments, RF sensing includes rejecting known friendly emissions by applying a whitelist or by recognizing known modulation patterns.Threat Detection and Classification

[0057] The processor 22 executes the machine learning model 24 to distinguish threats from background sources. In one embodiment, acoustic features include spectral peaks, harmonic structure, modulation patterns, and temporal envelopes characteristic of rotorcraft. RF features may include band occupancy patterns, modulation signatures, burst timing, and other characteristics indicative of UAV links.

[0058] In one embodiment, the machine learning model outputs a threat classification label (for example UAV / non-UAV, or a threat category) and a confidence value. The confidence value may be used to gate localization and alert generation, thereby reducing false alerts.

[0059] In some embodiments, the processor maintains separate classifiers for acoustic and RF data and fuses their outputs. Fusion may be rule-based or learned. For example, an RF detection may raise the confidence of an acoustic detection when temporal correlation exists.

[0060] In some embodiments, the device operates in a continuous monitoring mode with a low-power duty cycle. When a candidate detection is identified, the processor may temporarily increase sampling rate or processing intensity to refine localization before issuing an alert.Model Training and Deployment

[0061] In some embodiments, the machine learning model is trained offline using curated datasets that include UAV signatures and representative background environments. The trained model parameters are then deployed to the device.

[0062] In some embodiments, the device supports field updates to model parameters via the data interface. Updates may include new classes, revised thresholds, or improved feature extraction settings.

[0063] In some embodiments, training emphasizes generalization across different UAV models and environmental conditions, such as different wind, urban reflections, and different head-worn mounting geometries.Fusion, Confidence, and False- Alert Reduction

[0064] In some embodiments, the processor combines acoustic-domain and RF-domain detections using a fusion logic that outputs a unified threat score. The fusion logic may include rules (for example requiring temporal coincidence) and / or a learned fusion model.

[0065] In some embodiments, the fusion logic includes a persistence criterion, such that an alert is issued only when a threat score exceeds a threshold for a minimum duration, reducing spurious alerts.

[0066] In some embodiments, the device maintains an adaptive noise floor estimate for acoustic and RF domains and adjusts thresholds accordingly. This can improve robustness as environmental conditions change.

[0067] In some embodiments, the device distinguishes between a threat confidence threshold for generating an initial alert and a higher threshold for generating a high-urgency alert.

[0068] In some embodiments, the device suppresses repeated alerts for the same tracked threat unless a confidence or urgency metric increases.Localization Using Baseline Geometry and Antenna Directionality

[0069] The localization module 26 may utilize the known baseline between modules 12 and 14. In one embodiment, the module estimates a bearing by comparing arrival times, phases, or amplitudes of received signals at the left and right modules. The baseline provides a fixed geometry that can be calibrated at manufacture and / or during operation.

[0070] Antenna directionality may be used as an additional localization cue. For example, if a directional antenna pattern is known, relative received power as a function of head orientation can constrain a bearing estimate. In some embodiments, a phased array provides electronically steerable reception patterns that enable improved direction estimation.

[0071] In some embodiments, the localization module uses multiple cues (acoustic interaural time differences, acoustic level differences, RF power differences, and antenna response) and computes a fused bearing estimate.Doppler-Augmented Localization

[0072] Doppler analysis 28 may include estimating a frequency shift component in a received acoustic or RF signal that is attributable to relative motion between the threat and the device 10. The Doppler-derived information may be used as an additional constraint or feature for localization and / or for estimating whether a threat is approaching or receding.

[0073] In one embodiment, the localization module 26 determines a bearing using sensor spacing and / or antenna directionality and determines an approximate distance or distance class using a combination of signal strength features, directionality, temporal features, and Doppler-derived information.

[0074] In one embodiment, the localization module 26 computes a time difference of arrival (TDOA) or phase difference between signals captured at modules 12 and 14, and uses the known baseline distance to estimate a bearing. Doppler-derived information may be used to improve stability of the estimate over time and to reject inconsistent hypotheses.

[0075] In one embodiment, the Doppler-derived information is extracted from an acoustic signature associated with UAV propulsion (for example rotor or propeller harmonics) and used to refine an estimated approach rate. In another embodiment, Doppler-derived information is extracted from RF emissions and used similarly. In some embodiments, Doppler-derived information is extracted from both acoustic and RF data and combined, for example using confidence weighting.

[0076] In one embodiment, the Doppler-derived information is used to discriminate between stationary interferers and moving threats. For example, an approaching UAV may exhibit a consistent Doppler trend over a short window, whereas many stationary sources will not.

[0077] In one embodiment, the localization module evaluates multiple candidate bearings and selects a hypothesis that best fits both baseline-derived cues and Doppler-derived cues over time. This can improve localization performance in environments with reflections or intermittent RF reception.

[0078] In one embodiment, Doppler-derived information is used to estimate a time-to-closest-approach metric. The alert logic may use this metric to prioritize urgency even when distance is uncertain.Doppler Estimation Techniques

[0079] In an acoustic embodiment, Doppler analysis may be performed by tracking the frequency of one or more spectral peaks associated with propulsion harmonics and estimating a frequency shift over a time window. The estimate may be stabilized by averaging over multiple harmonics.

[0080] In an RF embodiment, Doppler analysis may be performed by estimating a carrier frequency offset or a drift of an RF feature over a time window. The device may compensate for oscillator drift using calibration or by referencing known signals.

[0081] In some embodiments, the Doppler analysis outputs a qualitative indicator (approaching, receding, or unknown) rather than a precise numeric velocity, thereby reducing computation while still improving alert relevance.

[0082] In some embodiments, the Doppler-derived information is used to improve range class estimation by increasing confidence that a threat is closing when both signal intensity and Doppler trend indicate approach.Tracking Over Time

[0083] In some embodiments, the localization module maintains a track state over successive time windows, including an estimated bearing, an optional distance class, and an optional approach / recede state. The track state may be updated using a filter such as a Kalman filter or an exponential smoother.

[0084] In some embodiments, the track state is used to suppress duplicate alerts and to provide a consistent directional cue to the user even when measurements are intermittent.

[0085] In some embodiments, Doppler-derived information contributes to track update by indicating whether the threat is closing, which may be used to prioritize urgency and to predict near-future bearing changes.

[0086] In some embodiments, the device supports tracking of multiple threats by maintaining multiple track hypotheses and associating detections to tracks using gating based on bearing and temporal proximity.Calibration and Self-Test

[0087] In some embodiments, the device performs calibration to account for manufacturing tolerances, antenna pattern variation, and microphone sensitivity differences. Calibration may include storing per-unit gain / phase offsets and applying correction during localization.

[0088] In some embodiments, the device performs a self-test at power-up or periodically. The self-test may include verifying sensor connectivity, measuring noise floors, and verifying that the processor and alert output are functional.

[0089] In retrofit embodiments, calibration may include a user-guided procedure to account for the mounting geometry of the add-on module relative to the headset.Head Motion and Opportunistic Scanning

[0090] In some embodiments, the device exploits natural head motion to sample different antenna orientations. As a user turns the head, changes in RF received power relative to known antenna patterns can provide additional directional constraints.

[0091] In some embodiments, the device uses inertial information from an external source or internal sensors (if present) to associate measurements with head orientation. Alternatively, head orientation can be inferred from relative changes in received signals.

[0092] In some embodiments, the device guides the user with a subtle audio cue to perform a short scan motion when confidence is high but bearing uncertainty remains large.Security and Update Integrity

[0093] In some embodiments, the device verifies authenticity of software updates and signature library updates using cryptographic signatures. This can reduce the risk that a malicious update degrades detection performance.

[0094] In some embodiments, the device includes a rollback mechanism to revert to a previous software or signature state if an update fails validation or results in abnormal behavior.

[0095] In some embodiments, event logs are stored in a tamper-evident format to support later review.Power Management and Updates

[0096] Low-power operation may be achieved by duty-cycling sensors, using event-driven processing, and selecting computationally efficient models. The device may run in a low-power scan mode and switch to a high-fidelity mode when a candidate threat is detected.

[0097] Firmware and signature library updates may be delivered via wired or wireless interfaces. Update integrity may be protected using cryptographic signatures, and the device may verify authenticity before applying updates.

[0098] In some embodiments, the device logs detection events, including classification outputs and localization estimates, for later review and model refinement.Machine Learning and Signature Libraries

[0099] The machine learning model 24 may operate on features extracted from the acoustic and / or RF signals, including spectral features, temporal modulation features, and other characteristics indicative of UAV operation. The model may be optimized for low-power execution on embedded hardware.

[0100] In some embodiments, the device maintains a signature library 36 comprising reference patterns for known threats and non-threats. The library may be updated to include new signatures. In some embodiments, the library may be synchronized or exchanged with external signature sets, for example commercial or professional signature repositories, via the data interface 34.

[0101] In some embodiments, adaptive learning is supported, where user feedback via optional user input 38 and / or post-event confirmation can be used to refine thresholds or update the signature library 36. Adaptive learning may be constrained to preserve safety and to reduce false positives.Alerting and User Interface

[0102] The warning output 30 may provide an alert as soon as the processor 22 determines that a threat confidence exceeds a threshold. The alert may be a tone, a sequence of tones, a pulse train, or a spatialized audio cue indicating a bearing (for example left / right panning or interaural level differences).

[0103] In one embodiment, different alert patterns correspond to different threat types, confidence levels, or range classes. In one embodiment, alert intensity or repetition rate increases when Doppler-derived information indicates that the threat is approaching.

[0104] In some embodiments, the device may provide a quiet mode, a training mode, or a logging mode. Logged events may be stored locally and / or transferred via data interface 34 for analysis and model improvement.

[0105] In some embodiments, the user may select profiles optimized for different environments (for example urban, rural, maritime, or battlefield). A profile may adjust thresholds, frequency bands, or alert styles.

[0106] In some embodiments, the device may provide a silent mode where alerts are primarily haptic, and a high-noise mode where alerts are louder or use distinctive patterns.Form Factors and Retrofit Implementations

[0107] The device 10 may be manufactured as an integrated headset, integrated helmet accessory, or integrated earbuds. In a retrofit embodiment, the system is configured to attach to or integrate with an existing headset that already includes microphones, speakers, and a power supply. Additional RF sensing hardware may be added as a module, and existing headset components may be used for alert output and / or acoustic acquisition.

[0108] In some embodiments, phased arrays may be used for enhanced directionality, off-axis noise suppression, and more precise triangulation. In some embodiments, chip antennas (virtual antennas / antenna boosters) may be used to facilitate multiband reception and to reduce physical size and cost of the RF reception unit.

[0109] The power supply 32 may be a rechargeable battery. Charging may be provided via a wired interface (for example USB-C) and / or via wireless charging. Firmware and signature library updates may be provided via data interface 34.Additional Embodiments and Variations

[0110] The device may be configured for different threat definitions by selecting different signature libraries and model parameters. For example, an industrial security configuration may focus on small quadcopter signatures, while another configuration may include other aerial hazards.[Oil 1] In some embodiments, the warning output includes both auditory and haptic cues. For example, a vibration motor may indicate direction by left / right placement or by pattern selection.

[0112] In some embodiments, the processor is implemented as a microcontroller, a system-on-chip, or a dedicated digital signal processor. Portions of the localization pipeline may be implemented in dedicated hardware accelerators to reduce power consumption.

[0113] In some embodiments, the device supports a training mode in which alerts are suppressed but detections are logged, enabling collection of field data for improving models and signature libraries.

[0114] In some embodiments, the device supports configurable privacy settings, such that raw data is not exported and only derived event summaries are shared externally.Examples

[0115] Example 1 (Over-ear headset): The device is integrated into an over-ear headset with one microphone and one RF receiver per side. The processor executes a low-power classifier to detect rotorcraft signatures in acoustic data and correlates candidate detections with RF band activity. When a threat is detected, the localization module estimates a left-right bearing and uses Doppler-derived information to indicate approach. The headset plays a spatialized tone toward the side of the threat, with repetition rate increasing as approach is detected.

[0116] Example 2 (Helmet-mounted retrofit): A retrofit module is attached to a helmetmounted hearing protection headset. The retrofit module reuses the headset speakers for alerts and adds an RF receiver and antenna booster. Calibration is performed once after installation to account for mounting geometry. The device provides alerts based on combined sensing and provides a coarse distance class (near / mid / far).

[0117] Example 3 (Earbud configuration): The device is implemented as earbuds with a fixed baseline provided by an interconnecting band. Acoustic and RF sensors are integrated into the earbuds. The device provides discrete alert patterns corresponding to threat confidence and bearing.Industrial Applicability

[0118] The disclosed device and method are applicable to personal security, protective services, critical infrastructure monitoring, law enforcement, military operations, and emergency response. The compact head-worn form factor enables deployment where larger counter-UAV systems are impractical.Additional Implementation ExamplesMachine Learning Implementation Examples

[0119] In some embodiments, the machine learning model comprises a compact classifier suitable for embedded execution, such as a small neural network operating on a timefrequency representation of acoustic data, a one-dimensional convolutional model operating on an audio envelope, or a gradient-boosted decision tree operating on hand-crafted features. The model may be selected to balance detection performance and power consumption.

[0120] Feature extraction may include windowing the acoustic signal into frames, computing a spectrum (e.g., via FFT), and generating features such as dominant harmonic frequencies, harmonic spacing, spectral centroid, spectral flux, modulation depth, and temporal persistence. For RF signals, feature extraction may include band-energy features, burst timing statistics, modulation indicators, channel occupancy, and temporal correlation with acoustic events.

[0121] In one embodiment, fusion is performed as late fusion in which the processor computes an acoustic-domain score and an RF-domain score and combines them using weighted averaging or rule-based gating. For example, the processor may require that an RF score exceed a threshold during a time window centered on an acoustic detection to increase confidence and reduce false positives.

[0122] In some embodiments, the model is optimized for embedded deployment using quantization (for example 8-bit integer quantization), pruning, or distillation. Inference may be executed periodically according to a duty cycle, and may be accelerated using a DSP block or a hardware accelerator where available.Localization and Triangulation Implementation Examples

[0123] In one acoustic embodiment, a bearing estimate may be derived from a time difference of arrival (TDOA) At between left and right sensors separated by a known baseline b, where an approximate relation sin(9)~c- At / b may be used for small baselines relative to range, with c being the speed of sound. The processor may estimate At by cross-correlation of band-limited signals and may refine 9 by combining interaural level differences and temporal persistence over successive windows.

[0124] In one RF embodiment, directionality constraints may be applied by comparing received power across known antenna response patterns, optionally leveraging head motion to sample multiple orientations. Approximate distance may be represented as a distance class derived from one or more of received-signal-strength indicators, persistence, and Doppler-derived approach / recede indicators, rather than a precise geometric range.

[0125] Doppler-derived information may be used as a consistency constraint in multihypothesis localization. For example, among candidate bearings consistent with baseline-derived cues, the processor may prefer hypotheses that produce a Doppler trend over time consistent with an approaching or receding trajectory and may reject hypotheses that are inconsistent across successive windows.Power Management Implementation Examples

[0126] In some embodiments, the device operates in a scan mode in which sensors are duty-cycled and feature extraction is performed at reduced cadence to detect candidate events. Upon candidate detection, the device enters a confirm mode for a limited interval in which sampling rate, RF scan bandwidth, and / or processing intensity are increased to refine classification and localization before generating an alert.

[0127] Power reduction techniques may include selectively enabling RF front-end components only during scan windows, using wake-on-sound triggers for acoustic capture, and using event-driven scheduling in which the processor remains in a low-power state between inference windows. Inference cadence and window length may be configured based on environment and desired reaction time.

[0128] In some embodiments, storage and radio interfaces used for updates are powered only during authenticated update sessions. Model and firmware updates may be cryptographically verified prior to installation, and the device may maintain a rollback image to reduce risk of failure while minimizing time spent in high-power radio operation.

Claims

AMENDED CLAIMSreceived by the International Bureau on 10 July 2026 (10.07.2026)

1. A headset- mo unted unmanned aerial vehicle (UAV) threat detection device, comprising:a first head-worn module and a second head- worn module arranged on opposite sides of a wearer's head with a fixed spatial separation;at least one acoustic sensor disposed in or coupled to each of the first head- worn module and the second head-worn module;at least one radio frequency sensor including or coupled to at least one antenna having a known antenna directionality;a processor with an embedded machine learning model trained to detect and classify UAV threats based on acoustic data from the acoustic sensors and radio frequency data from the radio frequency sensor;a localization module; anda warning system configured to provide auditory alerts via the headset upon detecting a UAV threat,wherein the localization module is configured to determine a bearing and at least one of an approximate distance and a distance class of a detected UAV threat by:(i) determining a time difference of arrival between acoustic signals captured by the acoustic sensors at the fixed spatial separation;(ii) applying the known antenna directionality to the radio frequency data;(iii) estimating a Doppler frequency- shift component in at least one of the acoustic data and the radio frequency data; and(iv) using the estimated Doppler frequency- shift component as an input, constraint, or weighting factor together with the time difference of arrival and the known antenna directionality to refine localization of the detected UAV threat.

2. The device of claim 1, wherein the machine learning model is optimized for low-power consumption.

3. The device of claim 1, wherein the distance class is selected from a plurality of predefined distance classes.

4. The device of claim 1, further comprising adaptive learning configured to improve detection accuracy over time by adapting at least one of (i) a detection threshold, (ii) a noise floor estimate, and (iii) a feature normalization parameter, based on at least one of user feedback and logged detection events.

5. The device of claim 1, wherein the at least one acoustic sensor is configured to detect UAV- specific noise including rotor and propeller sounds.

6. The device of claim 1, wherein the warning system supports user- selectable auditory signals to differentiate between at least one of threat types, distance classes, and directions.

7. The device of claim 1, wherein the device is configured to be retrofitted to an existing headset by utilizing at least one of existing microphones, an existing power supply, and existing speakers, and by adding at least one additional sensor module.

8. The device of claim 1, further comprising at least one of a plurality of acoustic sensors and a plurality of antenna elements arranged to operate as a phased array to provide enhanced directionality, off-axis noise suppression, or more precise source localization.

9. The device of claim 1, wherein the at least one antenna comprises a chip antenna configured to facilitate multiband reception and to reduce physical size of a radio frequency reception unit.

10. The device of claim 1, wherein the estimated Doppler frequency- shift component is estimated from a spectral peak, a propulsion harmonic, a carrier frequency offset, or a drift of a radio frequency feature over a time window.

11. The device of claim 1, wherein the Doppler frequency-shift component is estimated in both the acoustic data and the radio frequency data, and the estimated Doppler frequency- shift components are combined in the localization module.

12. The device of claim 1, wherein the localization module determines an approach rate or receding rate of the detected UAV threat based on the Doppler frequency-shift component and modifies an alert output based on the approach rate or receding rate.

13. The device of claim 1, wherein Doppler-derived information from the estimated Doppler frequency- shift component is used to increase stability of the bearing over successive time windows.

14. The device of claim 1, wherein the processor applies a confidence score to the detected UAV threat based on combined acoustic features and radio frequency features, and wherein the warning system provides differentiated alerts based on the confidence score.

15. The device of claim 1, wherein the processor applies noise suppression and filtering prior to executing the machine learning model.

16. The device of claim 1, wherein the device is configured as at least one of an over-ear headset, a helmet-mounted headset, and an earbud configuration.

17. The device of claim 1, further comprising a signature library storing reference signatures of threats and non-threats, wherein the processor updates the signature library based on detected events.

18. The device of claim 17, wherein the signature library is configured to be synchronized or exchanged with an external signature set via a data interface.

19. The device of claim 1, wherein the processor executes a first-stage detector and, upon identifying a candidate event, executes a second- stage detector with higher accuracy.

20. The device of claim 1, wherein the processor correlates temporally aligned acoustic features and radio frequency features to increase a threat confidence value.

21. The device of claim 1, wherein the processor maintains an adaptive noise floor estimate and adjusts at least one detection threshold based on the adaptive noise floor estimate.

22. The device of claim 1, wherein the radio frequency sensor is configured to monitor at least one of a control link band, a telemetry band, and a video downlink band associated with UAV operation.

23. The device of claim 1, wherein the warning system provides spatialized audio cues indicating a left-right bearing of the detected UAV threat.

24. The device of claim 1, wherein the warning system provides a first alert pattern when the localization module estimates that the detected UAV threat is within a first distance class and a second alert pattern when the localization module estimates that the detected UAV threat is within a second distance class different from the first distance class.

25. The device of claim 1, wherein the processor logs detection events and stores at least one of acoustic data, radio frequency data, and localization estimates for later transfer via a data interface.

26. The device of claim 1, wherein the processor is configured to receive software or signature updates via a wired interface.

27. The device of claim 1, wherein the processor is configured to receive software or signature updates via a wireless interface.

28. The device of claim 1, wherein the device comprises a rechargeable battery power supply and is configured for at least one of wired charging and wireless charging.

29. The device of claim 1, further comprising a calibration routine configured to store at least one of microphone gain offsets, antenna pattern parameters, and baseline geometry parameters and to apply the stored parameters during localization.

30. The device of claim 1, wherein the processor is configured to operate in a low-power scan mode and to switch to a higher-fidelity mode upon a candidate detection.

31. The device of claim 1, wherein the localization module computes a time-to-closest-approach metric based on Doppler-derived information and uses the metric to prioritize alert urgency.

32. The device of claim 1, wherein the device verifies authenticity of at least one of software updates and signature library updates using cryptographic signatures.

33. The device of claim 1, wherein the device performs a self-test routine to verify at least one of sensor connectivity and alert output functionality.

34. The device of claim 1, wherein the warning system is further configured to provide a haptic alert in addition to an auditory alert.

35. The device of claim 1, wherein the localization module maintains a track state over successive time windows and updates the track state using Doppler-derived information.

36. A method of operating a head- worn unmanned aerial vehicle (UAV) threat detection device, the method comprising:acquiring acoustic data from acoustic sensors disposed in or coupled to a first head-worn module and a second head- worn module arranged on opposite sides of a wearer's head with a fixed spatial separation;acquiring radio frequency data from at least one radio frequency sensor including or coupled to at least one antenna having a known antenna directionality;detecting and classifying a UAV threat based on the acoustic data and the radio frequency data using a machine learning model executed by a processor;estimating a bearing and at least one of an approximate distance and a distance class of the UAV threat by:(i) determining a time difference of arrival between acoustic signals captured by the acoustic sensors at the fixed spatial separation;(ii) applying the known antenna directionality to the radio frequency data;(iii) estimating a Doppler frequency- shift component in at least one of the acoustic data and the radio frequency data; and(iv) using the estimated Doppler frequency- shift component as an input, constraint, or weighting factor together with the time difference of arrival and the known antenna directionality to refine localization of the UAV threat; andproviding an auditory alert to a user via an output of the head- worn device.

37. The method of claim 36, wherein determining the time difference of arrival comprises cross-correlating band-limited acoustic signals captured at the fixed spatial separation.

38. The method of claim 36, wherein estimating comprises estimating an approach rate based on the Doppler frequency-shift component andincreasing an alert repetition rate when the approach rate indicates approach.

39. The method of claim 36, further comprising updating a threat signature library based on detected events.

40. The method of claim 36, wherein providing the auditory alert comprises providing a spatialized audio cue indicating a bearing of the UAV threat.

41. The method of claim 36, further comprising performing a calibration routine that applies stored gain or phase offsets to at least one of acoustic and radio frequency signals prior to localization.STATEMENT UNDER ARTICLE 19 (1)The claims have been amended to define more precisely the localization architecture of the head-worn UAV threat detection device. In particular, amended independent claims 1 and 36 now specify that localization is refined using a combination of: (i) a time difference of arrival between acoustic signals captured at the fixed head-worn spatial separation, (ii) known antenna directionality applied to radio frequency data, and (iii) an estimated Doppler frequency-shift component used as an input, constraint, or weighting factor in the localization process.The amendments clarify that the claimed invention is not merely the presence of acoustic and radio frequency sensing in a headset, but a particular cooperative use of baseline-derived acoustic timing information, antenna-directionality information, and Doppler-derived information to determine bearing and approximate distance or distance class of a UAV threat.The amended claims remain fully supported by the application as filed, including the disclosure of head-worn modules having fixed spatial separation, acoustic and RF sensing, antenna directionality, Doppler analysis, TDOA-based bearing estimation, Doppler-derived consistency and track updating over successive time windows, and auditory alerting through the headset. No amendment is made to the description or drawings.

Images