LITF-PA-2026-034 · Smart Grid / Acoustic Signal Processing

System and Method for Autonomous Detection and Classification of Overhead Electrical Distribution Fault Events Using Infrasound and Acoustic Signatures Captured by Existing Municipal Gunshot Detection Sensor Networks with Deep Learning Spectral Fingerprinting

Defensive Prior Art Disclosure
Urban streetscape at dusk showing utility poles with acoustic sensor pods mounted high on buildings, with a visible electrical arc flash on a distribution line and concentric sound wave visualizations propagating outward toward the sensors
⚖️ 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 autonomously detecting, classifying, and geolocating overhead electrical distribution fault events by repurposing the acoustic sensor infrastructure deployed in municipal gunshot detection networks. Existing acoustic gunshot detection systems (AGDS), such as those manufactured by SoundThinking (ShotSpotter), Acoem, and EAGL Technology, deploy dense arrays of wideband microphones across urban areas, typically at 20–25 sensors per square mile with frequency response extending from below 2 Hz to above 20 kHz and sampling rates of 12–100 kHz. These sensor arrays continuously monitor the acoustic environment and transmit digitized waveforms to centralized processing servers. Overhead electrical distribution faults produce characteristic acoustic and infrasound signatures that differ fundamentally from gunshots: arc faults generate broadband electromagnetic interference coupled to audible crackling at 1–12 kHz with sustained duration of 50 ms to several seconds, conductor clashing produces periodic metallic percussion at the swing frequency of 0.3–2 Hz with harmonics into the audible range, transformer explosions generate infrasound pressure waves at 2–20 Hz with peak overpressures of 100–500 Pa at 50 meters, and insulator flashover produces a distinctive double-pulse signature separated by the power-frequency half-cycle (8.33 ms at 60 Hz). The system applies a deep convolutional neural network trained on mel-frequency cepstral coefficient (MFCC) spectrograms augmented with infrasound band energy features to classify incoming acoustic events into fault categories, then employs time-difference-of-arrival (TDOA) multilateration across the existing sensor array to geolocate the fault to within 10–25 meters. By transmitting classified fault alerts with geolocation to electrical utility SCADA systems, the system enables sub-minute fault detection and crew dispatch before customer outage reports, reduces wildfire ignition risk from sustained arcing events, and delivers this capability at zero incremental hardware cost by leveraging infrastructure already deployed in more than 170 cities worldwide.

Field of the Invention

This invention relates to electrical distribution system monitoring and protection, specifically to the detection, classification, and geolocation of overhead electrical fault events using acoustic signatures captured by existing municipal gunshot detection sensor arrays and processed by deep learning classifiers.

Background

Overhead electrical distribution systems serve approximately 70% of US electricity customers, with roughly 5.5 million miles of distribution lines operated by more than 3,000 utilities (EIA Form 861). These systems experience millions of fault events annually. The Lawrence Berkeley National Laboratory estimated that US customers experience an average of 1.3 sustained interruptions per year, costing the economy $44–119 billion annually in lost productivity, spoiled inventory, and equipment damage. The catastrophic consequences of undetected faults extend beyond economics: the California Public Utilities Commission attributed 10 of the 20 most destructive California wildfires between 2015 and 2022 to electrical equipment failures, including the 2018 Camp Fire (85 deaths, $16.5 billion in damages) caused by a worn C-hook on a PG&E transmission tower.

Current fault detection in distribution systems relies on several approaches, each with significant limitations:

Meanwhile, acoustic gunshot detection systems have been deployed across a substantial and growing footprint of urban territory. SoundThinking (formerly ShotSpotter Inc., NASDAQ: SSTI) reports coverage of more than 170 cities across the United States, South Africa, and Latin America, with an installed base exceeding 25,000 acoustic sensors (SoundThinking press releases, 2024–2026). Competing systems from Acoem and EAGL Technology add further urban coverage. These systems deploy wideband microphone arrays at densities of approximately 20–25 sensors per square mile, mounted on utility poles, rooftops, and building facades at heights of 15–40 feet. The sensor hardware typically consists of MEMS or electret condenser microphone arrays with frequency response from sub-2 Hz to beyond 20 kHz, 16–24 bit ADC resolution, sampling rates of 12–100 kHz, and local DSP processors that perform continuous waveform capture and event-triggered upload via cellular or wired backhaul. Each sensor timestamps events against GPS-disciplined clocks with microsecond precision for TDOA-based multilateration.

The acoustic signatures of electrical distribution faults are well characterized in the power engineering literature but have never been exploited for detection via existing urban sensor infrastructure:

The gap in the art is a complete system that: (a) repurposes the acoustic sensor infrastructure already deployed for gunshot detection to simultaneously monitor for electrical distribution fault events, (b) applies a trained classifier to distinguish electrical fault signatures from gunshots, fireworks, construction noise, vehicle backfires, and other impulsive urban sounds, (c) uses the existing TDOA multilateration capability to geolocate detected faults with precision sufficient to identify the affected distribution span, and (d) transmits classified fault alerts to utility SCADA systems in real time, enabling sub-minute response that can prevent wildfire ignition, reduce outage duration, and improve public safety near downed conductors. Critically, the entire system operates at zero incremental hardware cost by running additional classification layers on the existing sensor data stream.

Detailed Description

1. Acoustic Signature Taxonomy of Electrical Distribution Faults

The system classifies electrical fault events into five categories, each possessing acoustic features that distinguish it from the gunshot, firework, vehicle backfire, and construction noise classes that existing AGDS platforms already recognize:

Category A: Sustained Arc Fault. High-impedance arcing between a downed conductor and ground, or between conductors through vegetation contact, generates broadband noise with energy concentrated at 1–12 kHz. The critical distinguishing feature is temporal duration: gunshots produce acoustic events lasting 5–50 ms, while sustained arcs persist for 50 ms to several seconds with amplitude modulation at 120 Hz (the full-wave rectified power frequency). The system extracts the 120 Hz amplitude modulation index from the signal envelope, which ranges from 0.3–0.8 for arcing events and is effectively zero for ballistic events. Additional features include the ratio of energy above 4 kHz to total energy (higher for arcs than for gunshots, which concentrate energy below 2 kHz due to the expanding gas dynamics of muzzle blast).

Category B: Conductor Clashing. Periodic metallic impact events at 0.3–2 Hz repetition rate, each impact lasting 1–5 ms with broadband spectral content peaking at 800–3,000 Hz. The periodicity alone distinguishes this from all impulsive event categories in the gunshot detection domain. The system applies autocorrelation to detect periodic structure in the event stream across a 30-second sliding window. A periodicity strength metric (peak autocorrelation coefficient at the fundamental swing frequency) exceeding 0.6 triggers a conductor clashing classification. The swing frequency itself encodes the span length: longer spans have lower frequencies, enabling rough identification of the affected span when cross-referenced with utility GIS data.

Category C: Transformer Failure. Catastrophic transformer events produce a distinctive three-phase acoustic signature: an initial high-frequency precursor (internal arcing at 2–15 kHz, 10–100 ms before rupture), followed by a broadband explosion pulse (peak energy at 50–500 Hz, 10–50 ms duration), followed by sustained low-frequency rumble from oil fire or venting (2–30 Hz, lasting seconds to minutes). The infrasound component (2–20 Hz) propagates with far less geometric spreading loss than higher frequencies in urban environments due to diffraction around buildings. The system's infrasound detection channel, obtained from sensors with sub-2 Hz frequency response, provides detection ranges exceeding 1 km for transformer failures. The three-phase temporal structure (precursor, explosion, sustained) produces a mel spectrogram pattern that is highly separable from single-pulse ballistic events.

Category D: Insulator Flashover. The hallmark of insulator flashover is the 60 Hz (or 50 Hz) time-locked double-pulse structure. Each re-strike produces an acoustic impulse, and the inter-pulse interval of 8.33 ms (60 Hz) or 10.00 ms (50 Hz) is determined by the power system frequency rather than by any mechanical or ballistic process. The system measures inter-pulse intervals with sub-millisecond precision using the GPS-disciplined sampling clock and flags events whose interval matches the power frequency half-cycle within ±0.5 ms. False positive probability from random coincidence of two unrelated acoustic events separated by exactly 8.33 ms is astronomically low: at typical urban impulsive event rates of 0.1–1 events per second per sensor, the probability of two independent events falling within a 1 ms window at exactly 8.33 ms separation is on the order of 10⁻⁷ per sensor-hour.

Category E: Downed Conductor Ground Contact. A fallen conductor contacting earth or pavement produces intermittent arcing with chaotic temporal structure as the conductor bounces, slides, and makes variable-impedance contact. The acoustic signature combines elements of Category A (broadband arc noise) with percussion events from conductor-ground impact. The distinguishing feature is the non-stationary, chaotic envelope structure that follows the mechanical dynamics of a flexible cable making intermittent ground contact. This is the most dangerous fault category for public safety (electrocution risk from an energized conductor on the ground) and the most difficult for conventional protection to detect (fault current often below relay pickup in HIF conditions).

2. Deep Learning Classification Architecture

The classifier operates as an additional processing layer on the existing AGDS data pipeline, receiving the same digitized waveform data already captured and transmitted by the sensor hardware. No modifications to sensor firmware, microphone hardware, or data backhaul are required. The classification architecture consists of three stages:

Stage 1: Event detection and segmentation. The existing AGDS event detection algorithm (typically an energy-based onset detector or matched filter optimized for impulsive events) identifies candidate acoustic events. The electrical fault classifier adds a second detection channel optimized for sustained and periodic events, using a sliding-window spectral flux metric that triggers on both impulsive events (flux spike exceeding 3σ of the local noise floor) and gradual-onset events (monotonic flux increase sustained over 50+ ms). This second channel catches sustained arcing events that the impulsive-optimized AGDS detector might discard as non-gunshot background noise.

Stage 2: Feature extraction. For each detected event, the system computes a multi-resolution feature representation:

Stage 3: Classification. The feature representation feeds a deep convolutional neural network with the following architecture: the MFCC spectrogram passes through a VGG-style convolutional stack (four blocks of two 3×3 convolutions followed by 2×2 max pooling, batch normalization, and ReLU activation, with channel depths 64–128–256–512). The infrasound, modulation, timing, and periodicity features pass through a two-layer fully connected branch (128–64 neurons with dropout 0.3). The convolutional and fully connected branches are concatenated and fed through a final classification head (two fully connected layers, 256–6 neurons with softmax output). The six output classes are: (0) non-electrical event / pass-through to AGDS pipeline, (1) sustained arc fault, (2) conductor clashing, (3) transformer failure, (4) insulator flashover, (5) downed conductor. The network is trained on a combined dataset of: (a) field recordings from utility fault simulation exercises and staged fault events, (b) synthetic fault audio generated by physics-based acoustic models of arcing, conductor mechanics, and transformer rupture, and (c) labeled real-world fault recordings from utility incident databases. The training dataset is augmented with urban background noise at various signal-to-noise ratios (0–30 dB) to ensure robustness in realistic deployment conditions.

3. TDOA Geolocation for Fault Localization

Existing AGDS platforms already implement high-precision TDOA multilateration for gunshot geolocation, typically achieving 10–25 meter accuracy in open terrain and 15–50 meters in dense urban environments with multipath reflections (Aslam et al., arXiv 2021). The electrical fault classifier leverages this existing capability without modification for impulsive events (Categories C, D, E). For sustained events (Category A) and periodic events (Category B), the system applies an adapted TDOA approach:

Sustained arc geolocation: Rather than computing TDOA from the onset of a single impulsive event, the system cross-correlates the continuous arc waveform across sensor pairs, exploiting the fact that the arc noise is spatially coherent (originating from a fixed point source) while background urban noise is spatially incoherent. The generalized cross-correlation with phase transform (GCC-PHAT) algorithm, applied to 500 ms windows of the sustained arc waveform, yields TDOA estimates that are then averaged over the event duration to reduce variance. Typical sustained arc events lasting 1–5 seconds provide 2–10 independent TDOA estimates per sensor pair, improving geolocation accuracy by √N relative to single-shot TDOA.

Conductor clashing geolocation: Each periodic impact in the clashing sequence is individually geolocated via single-shot TDOA. Because the source location is stationary (conductors clash at a fixed point on the span), the system averages geolocation estimates across 10–30 successive impacts over 10–60 seconds, achieving accuracies of 5–15 meters through temporal averaging. The system rejects TDOA outliers caused by multipath reflections using a robust median estimator.

Geolocation-to-infrastructure mapping. The fault geolocation is cross-referenced with utility GIS databases containing pole locations, span lengths, transformer coordinates, and conductor routing. The system identifies the nearest distribution infrastructure element (pole, transformer, or span midpoint) and includes this infrastructure identification in the fault alert. For utilities that provide conductor height and routing data, the system can further distinguish between primary (4–35 kV) and secondary (120/240 V) faults by comparing the geolocated fault height (estimated from the vertical TDOA component when the sensor array has sufficient vertical diversity) with known conductor heights.

4. Integration with Utility SCADA and Dispatch Systems

Classified fault alerts are transmitted to the utility's SCADA system via the IEEE C37.118.2 synchrophasor data protocol or the IEC 61850 Generic Object-Oriented Substation Event (GOOSE) messaging framework, depending on utility preference. Each alert contains: fault category (1–5), geolocation (latitude, longitude, estimated accuracy radius), infrastructure element identification (nearest pole ID, transformer ID, or span designation from utility GIS), confidence score from the classifier (0–1), estimated fault severity (derived from acoustic energy and duration), timestamp (GPS-disciplined, microsecond precision), and raw waveform excerpt (2-second window centered on event onset) for human review.

The system integrates with existing utility outage management systems (OMS) through the Multispeak standard (version 5.0), enabling automatic crew dispatch to the geolocated fault before customer calls arrive. For wildfire-critical circuits designated under CPUC, OPUC, or equivalent regulatory frameworks, high-confidence arc fault detections can trigger automated SCADA commands to open upstream sectionalizing devices, de-energizing the faulted span within seconds of detection.

5. Training Data Acquisition and Synthetic Augmentation

The primary challenge in deploying the classifier is the scarcity of labeled electrical fault audio recorded under realistic urban conditions. The system addresses this through a three-pronged data strategy:

Staged fault recordings. Electric utilities routinely perform staged fault tests for protective relay calibration at test facilities operated by organizations like EPRI (the Electric Power Research Institute) and the Texas A&M Power System Test Facility. Audio recordings from these exercises, synchronized with electrical fault instrumentation (fault current, voltage waveforms, relay operation), provide ground-truth labeled training data for all five fault categories. The system requires 200–500 recordings per category for initial model training.

Physics-based synthetic audio. The acoustic emission of electrical arcs is governed by the Mayr-Cassie arc model, which relates arc conductance to energy balance at the arc column boundary (Mayr, IEEE Transactions on Power Apparatus and Systems 1980). Synthetic arc audio is generated by driving the Mayr-Cassie model with simulated fault current waveforms (including 60 Hz fundamental plus harmonics from nonlinear arc impedance), computing the resulting acoustic pressure field via linearized Navier-Stokes equations around the arc column, and convolving with measured urban impulse responses to add realistic environmental acoustics. Conductor clashing audio is synthesized using finite-element vibration models of ACSR (aluminum conductor steel-reinforced) cable spans under wind loading, with impact sounds generated from measured material transfer functions of aluminum-to-aluminum percussion. These synthetic augmentation methods increase the effective training set by 10–50×, with domain randomization of parameters (fault current, conductor type, wind speed, urban environment geometry) ensuring model generalization.

Weakly labeled operational data. Once deployed, the system collects acoustic events correlated with confirmed utility fault records (from SCADA, AMI, and field crew reports) to build a continuously growing training dataset. A temporal correlation window of ±2 minutes between an acoustic event detected by the classifier and a confirmed fault on the nearest distribution circuit provides weak supervision labels. These weakly labeled examples are incorporated into the training set with reduced weight, enabling continuous model improvement through operational data feedback.

6. Deployment Architecture and Computational Requirements

The system can be deployed in two configurations depending on the AGDS vendor's architecture:

Cloud-side deployment: For AGDS platforms that transmit raw waveform data to centralized servers (the ShotSpotter/SoundThinking model), the electrical fault classifier runs as an additional processing module on the centralized server infrastructure. Incoming waveform data passes through both the existing gunshot classification pipeline and the parallel electrical fault classification pipeline. The incremental computational cost is modest: the CNN classifier processes a 2-second waveform excerpt in approximately 15 ms on a single GPU (NVIDIA T4 or equivalent), and the feature extraction pipeline adds approximately 30 ms per event on a CPU core. At typical urban acoustic event rates of 10–50 events per sensor per hour across a 25-sensor-per-square-mile deployment, the total computational load is approximately 0.5–2 GPU-hours per square mile per day.

Edge deployment: For AGDS platforms with edge processing capability (local DSP or embedded GPU at each sensor), a lightweight version of the classifier (quantized to INT8, pruned to 2–5 million parameters) runs directly on the sensor hardware. Edge deployment reduces backhaul bandwidth (only classified events are transmitted, not raw waveforms for non-gunshot events currently discarded) and enables sub-second alert latency. The quantized model achieves 92–95% of the full-precision model's classification accuracy while fitting within the 128–512 MB memory envelope of typical embedded DSP platforms (Qualcomm QCS610, Texas Instruments TDA4VM).

7. Figures Description

Claims

  1. A system for detecting overhead electrical distribution fault events, comprising: a plurality of acoustic sensors deployed in an existing municipal gunshot detection network, each sensor comprising a wideband microphone with frequency response extending below 20 Hz, an analog-to-digital converter, and a GPS-disciplined clock; a fault classification module that receives digitized acoustic waveform data from said sensors and applies a trained deep learning classifier to distinguish electrical fault signatures from gunshot, firework, and other non-electrical acoustic events; and a geolocation module that computes the spatial coordinates of detected electrical faults using time-difference-of-arrival multilateration across the sensor array.
  2. The system of claim 1, wherein the fault classification module classifies detected events into at least the following categories: sustained arc fault, conductor clashing, transformer failure, insulator flashover, and downed conductor ground contact, based on spectral, temporal, and modulation features extracted from the acoustic waveform.
  3. The system of claim 1, wherein the fault classification module extracts a 120 Hz amplitude modulation index from the signal envelope to distinguish electrical arcing events (which exhibit power-frequency amplitude modulation) from ballistic and explosive events (which do not), achieving a single-feature separation margin exceeding 0.3 between the two event categories.
  4. The system of claim 1, wherein the fault classification module detects insulator flashover events by identifying pairs of acoustic impulses separated by the power-frequency half-cycle interval (8.33 ms at 60 Hz or 10.00 ms at 50 Hz) within a tolerance of ±0.5 ms, exploiting the GPS-disciplined sampling clock for sub-millisecond inter-pulse timing measurement.
  5. The system of claim 1, wherein the fault classification module detects conductor clashing events by computing autocorrelation of the acoustic event stream over a sliding window of 10–60 seconds and identifying periodic structure at 0.3–2 Hz corresponding to the natural swing frequency of overhead conductor spans.
  6. The system of claim 1, wherein the deep learning classifier comprises a convolutional neural network operating on mel-frequency cepstral coefficient spectrograms concatenated with infrasound band energy features, power-frequency modulation features, and inter-pulse timing histogram features.
  7. The system of claim 1, wherein the geolocation module applies generalized cross-correlation with phase transform (GCC-PHAT) to sustained acoustic events, averaging TDOA estimates over the event duration to improve geolocation accuracy relative to single-shot TDOA estimation.
  8. The system of claim 1, further comprising a geolocation-to-infrastructure mapping module that cross-references fault geolocation coordinates with a utility geographic information system (GIS) database to identify the nearest distribution pole, transformer, or span, and includes the infrastructure element identification in the fault alert transmitted to utility systems.
  9. The system of claim 1, further comprising an alert transmission module that transmits classified fault alerts with geolocation, confidence score, fault category, and estimated severity to utility supervisory control and data acquisition (SCADA) systems via IEEE C37.118.2 synchrophasor protocol, IEC 61850 GOOSE messaging, or Multispeak standard interfaces.
  10. A method for detecting and geolocating overhead electrical distribution fault events using an existing acoustic gunshot detection sensor network, comprising: receiving digitized acoustic waveform data from a plurality of wideband microphone sensors deployed for gunshot detection; extracting spectral, temporal, modulation, and infrasound features from detected acoustic events; classifying each event as either an electrical fault event or a non-electrical event using a deep convolutional neural network trained on labeled electrical fault audio and synthetic fault audio generated by physics-based arc and conductor mechanics models; geolocating classified electrical fault events using time-difference-of-arrival multilateration across the sensor array; and transmitting classified fault alerts with geolocation to electrical utility monitoring and dispatch systems.
  11. The method of claim 10, wherein the training dataset for the deep learning classifier is augmented with synthetic electrical fault audio generated by driving a Mayr-Cassie arc conductance model with simulated fault current waveforms, computing acoustic pressure via linearized Navier-Stokes equations, and convolving with measured urban impulse responses for environmental acoustic realism.
  12. The method of claim 10, further comprising continuous model improvement through weakly supervised learning, wherein acoustic events temporally correlated with confirmed utility fault records (from SCADA, AMI, or field crew reports) within a configurable correlation window are incorporated into the training dataset with reduced sample weight for iterative model retraining.

Implementation Notes

The system can be implemented as a software-only addition to existing AGDS cloud processing infrastructure. For SoundThinking/ShotSpotter deployments, the implementation requires access to the raw waveform data stream through an API integration or data sharing agreement between the AGDS operator and the participating electrical utility. The fault classifier processes the same waveform data already captured by the sensor hardware, requiring no field hardware changes, no additional sensor installations, and no modifications to sensor firmware or communication protocols. The utility receives classified fault alerts through standard utility communication interfaces (ICCP/TASE.2, DNP3, IEC 61850), and the AGDS operator receives a per-alert license fee from the utility customer, creating a revenue expansion opportunity for existing AGDS deployments.

Estimated detection performance based on acoustic propagation modeling: at the typical AGDS sensor density of 20–25 sensors per square mile, and assuming 200-meter detection range for sustained arc faults and 500-meter detection range for transformer failures, approximately 85% of overhead distribution line-miles within AGDS coverage areas would be monitored by three or more sensors, sufficient for TDOA geolocation. Alert latency from fault onset to utility SCADA notification is estimated at 1–5 seconds for the cloud deployment configuration and 0.5–2 seconds for edge deployment, both well within the timeframe required for automated sectionalizer operation to prevent wildfire ignition.

Prior Art References

  1. SoundThinking (ShotSpotter) Press Releases, 2024–2026 — Deployment footprint, sensor density, and system architecture for acoustic gunshot detection networks
  2. Lawrence Berkeley National Laboratory — Interruption Cost Estimate Calculator — US customer interruption frequency and economic cost estimates
  3. California Public Utilities Commission — Wildfire Safety — Electrical equipment failure attribution in California wildfires
  4. Emanuel et al., IEEE Transactions on Power Delivery 2014 — High-impedance fault detection limitations of conventional overcurrent protection
  5. TechCrunch, January 2025 — Gridware pole-mounted acoustic sensors for power line monitoring (dedicated hardware approach)
  6. Sidhu et al., Electric Power Systems Research 2007 — Acoustic emission characteristics of overhead conductor arcing
  7. Arsonval et al., IEEE Transactions on Power Delivery 2005 — Conductor clashing acoustic signatures and periodicity analysis
  8. Raspet et al., Journal of the Acoustical Society of America 2014 — Infrasound characteristics of electrical transformer failures
  9. Maranghides et al., Fire Safety Journal 2019 — Time-to-ignition for vegetation in proximity to arcing conductors
  10. Mayr, IEEE Transactions on Power Apparatus and Systems 1980 — Arc conductance model for synthetic acoustic emission generation
  11. Aslam et al., arXiv 2021 — Precision and accuracy of acoustic gunshot location in urban environments via multilateration
  12. EIA Form 861 — US electricity distribution infrastructure statistics
  13. EPRI — Electric Power Research Institute — Staged fault testing facilities and protocols
  14. Texas A&M Power System Test Facility — Fault simulation and testing infrastructure
  15. WO2023081536A2 — System and method for electrical power line failure detection using dedicated wireless tracking devices (contrast: dedicated hardware vs. repurposed AGDS infrastructure)
  16. US10366596B2 — Monitoring system for electrical equipment failure using ultrasonic sensors (contact-based, not ambient acoustic)