LITF-PA-2026-099 · Energy Infrastructure / Audio Signal Processing / Edge AI

System and Method for Distributed Power Grid Frequency Monitoring and Anomaly Localization Using Electrical Network Frequency Extraction from Consumer IoT Device Audio Streams with Geospatially Indexed Neural State Estimation

Defensive Prior Art Disclosure
Consumer IoT devices in residential homes with electromagnetic coupling lines to power grid infrastructure and frequency heatmap overlay
⚖️ 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 continuous, geospatially dense monitoring of power grid frequency, phase, and stability using the electrical network frequency (ENF) signal naturally embedded in audio recordings captured by consumer Internet of Things (IoT) devices. Every audio-capable device connected to or situated near the alternating current (AC) power grid — smart speakers, security cameras, video doorbells, baby monitors, voice assistants, smart displays — captures a faint but measurable 50/60 Hz mains hum in its microphone signal through electromagnetic coupling and acoustic radiation from nearby transformers, wiring, and appliances. This ENF signal encodes the instantaneous grid frequency at the device's location with sub-millihertz precision extractable through established signal processing techniques. The disclosed system aggregates ENF measurements from millions of geographically distributed consumer devices (with explicit user consent) to construct a real-time, continent-scale power grid frequency map with spatial resolution orders of magnitude finer than existing phasor measurement unit (PMU) networks. An on-device neural ENF extractor (under 200 KB quantized) isolates the mains hum fundamental and harmonics from ambient audio without transmitting raw audio, preserving user privacy. A cloud-based geospatial state estimator fuses the streaming ENF telemetry with grid topology models to detect frequency deviations, localize generation-load imbalances, identify inter-area oscillation modes, detect islanding events, and estimate inertia distribution across the grid — capabilities that currently require dedicated PMU hardware costing $40,000–$100,000 per installation point, with only approximately 2,500 units deployed across the entire North American grid.

Field of the Invention

This invention relates to power systems monitoring and situational awareness, specifically to methods for repurposing the audio capture hardware of consumer IoT devices as a massively distributed grid frequency sensing network through extraction and aggregation of the electrical network frequency signal embedded in ambient audio, with on-device neural signal processing and cloud-based geospatial grid state estimation.

Background

The electric power grid is the largest machine ever built by humans. Keeping it stable requires that generation precisely matches load at every instant — any sustained imbalance causes the grid frequency to deviate from its nominal value (60.000 Hz in North America, 50.000 Hz in Europe and most of Asia). A frequency drop of just 0.5 Hz indicates a generation shortfall of roughly 1–2 GW across the Eastern Interconnection, enough to trigger underfrequency load shedding. The catastrophic Texas grid failure of February 2021 saw frequency drop to 59.302 Hz — 4.8 minutes from a total grid collapse that would have left 26 million people without power for weeks.

Monitoring grid frequency at high spatial resolution is therefore critical for grid operators, yet the existing infrastructure for doing so is remarkably sparse. Phasor measurement units (PMUs) — dedicated devices that measure voltage magnitude and phase angle at the grid's transmission level, GPS-synchronized to sub-microsecond accuracy — represent the gold standard. But PMUs cost $40,000–$100,000 per unit installed, require dedicated communications infrastructure (typically fiber or cellular backhaul), and must be integrated into utility SCADA systems. The North American SynchroPhasor Initiative (NASPI) reports approximately 2,500 PMUs deployed across the three North American interconnections as of 2024, covering a grid serving 370 million people across 8.08 million km². That works out to one measurement point per 3,230 km² — roughly one PMU per county. Distribution-level visibility is essentially nonexistent.

The FNET/GridEye system, developed at the University of Tennessee and Virginia Tech, demonstrated that low-cost frequency disturbance recorders (FDRs) could provide wide-area monitoring at a fraction of PMU cost. Liu et al. (2006, IEEE Power and Energy Magazine) showed that 80 FDR units distributed across the Eastern Interconnection could detect and localize generation trip events within 2–4 seconds. But even FDRs require dedicated hardware ($200–500 per unit), physical installation at wall outlets, and communication provisioning. Scaling to millions of measurement points remains infeasible with purpose-built hardware.

Meanwhile, the audio forensics community has spent two decades proving that consumer microphones are already capturing grid frequency with remarkable fidelity. Electrical network frequency (ENF) analysis — the extraction of mains hum from audio recordings for forensic timestamp authentication and location verification — is an established discipline. Grigoras (2005, International Journal of Speech, Language and the Law) first demonstrated that the ENF signal embedded in audio recordings could be matched against a reference database to authenticate recording timestamps. The ENF signal enters audio recordings through two physical mechanisms: electromagnetic coupling of the 50/60 Hz magnetic field from nearby AC wiring, motors, and transformers into the microphone's coil or circuit traces; and acoustic radiation of mains-frequency vibration from transformers, fluorescent ballasts, and other devices that audibly hum at line frequency. Ojowu et al. (2012, IEEE Transactions on Information Forensics and Security) showed that adaptive techniques could extract ENF from recordings with SNR as low as -30 dB relative to the ambient audio, achieving frequency estimation accuracy of ±2 millihertz over 1-second windows. Bykhovsky and Cohen (2013, IEEE Transactions on Information Forensics and Security) developed a maximum-likelihood ENF estimator exploiting the multi-tone harmonic structure (60 Hz, 120 Hz, 180 Hz, ...) that achieved Cramér-Rao bound performance in many practical scenarios.

The critical gap: two decades of ENF research have focused exclusively on forensic applications — authenticating recordings by comparing their embedded ENF against a known reference. Nobody has proposed inverting the problem. Instead of using a reference grid frequency to validate a recording, the system disclosed here uses distributed recordings to measure the grid frequency itself. The inversion is not trivial. Forensic ENF extraction operates on a single recording after the fact, tolerates seconds to minutes of processing latency, and compares against a known reference obtained from a dedicated probe connected directly to the grid. Grid monitoring requires real-time extraction from millions of simultaneous streams, sub-second latency, and the ability to function without any reference signal — the consumer devices ARE the reference network.

The opportunity is staggering in scale. There are an estimated 500 million smart speakers installed worldwide as of 2025 (Statista). Ring, Nest, Arlo, and other video doorbell and security camera brands have shipped over 100 million units in the United States alone. Every one of these devices contains a microphone that is, right now, passively capturing the local grid frequency in its audio input, discarding it as noise. The system disclosed here turns that noise into signal.

Detailed Description

1. ENF Signal Characteristics in Consumer IoT Audio

The ENF signal present in consumer device audio has specific characteristics that distinguish it from forensic lab recordings and that shape the design of the extraction pipeline.

Coupling mechanisms: In consumer IoT devices, the ENF signal enters the audio path through three primary mechanisms, typically all active simultaneously: (a) electromagnetic induction from the device's own AC power supply into the microphone amplifier circuit traces, producing a stable, high-SNR ENF component (typically -20 to -40 dB relative to full-scale audio at the fundamental, with harmonics at 2f, 3f, 4f extending to 300+ Hz); (b) electromagnetic radiation from nearby household wiring (Romex in walls, extension cords, appliance power cables within 1–3 meters), producing a weaker but spatially specific ENF signal; and (c) acoustic radiation from transformers in wall adapters, HVAC systems, refrigerator compressors, fluorescent lighting ballasts, and dimmer switches, which physically vibrate at line frequency and its harmonics. Battery-powered devices (wireless cameras, portable speakers) receive ENF only through mechanisms (b) and (c), resulting in 10–20 dB lower ENF signal strength but still extractable in most indoor environments where AC-powered equipment is present.

Frequency content: The ENF fundamental at 60.000 Hz (nominal, North America) fluctuates within a normal range of ±0.02 Hz during routine operation, with excursions to ±0.5 Hz during significant generation-load imbalances. Harmonics at 120 Hz, 180 Hz, and 240 Hz carry additional information: the relative amplitudes of these harmonics depend on the waveform distortion of the local AC supply, which varies with the transformer tap ratio, nonlinear loads on the local feeder, and the device's own power supply topology. The harmonic pattern constitutes a local fingerprint that can distinguish between different feeder circuits even when the fundamental frequency is identical across a wide area (because frequency is synchronized across the entire interconnection at timescales above ~15 seconds).

Phase information: While grid frequency is nearly uniform across a synchronous interconnection at timescales above 10–15 seconds, the voltage phase angle varies continuously across the grid, reflecting the flow of real power from generators to loads. Two devices on different feeders in the same city will measure the same frequency but different phase angles. Extracting the absolute phase angle from consumer audio is challenging because the coupling delay through the microphone circuit introduces an unknown, device-specific phase offset. However, the rate of change of phase angle (dφ/dt) is independent of any constant offset and equals the frequency deviation from nominal. More importantly, phase angle differences between pairs of devices can be estimated by cross-correlating their ENF waveforms — the relative phase between two locations is recoverable even when the absolute phase at either location is not.

Temporal resolution: With standard signal processing techniques (STFT with a 1-second Hanning window), the ENF fundamental can be estimated with a frequency resolution of ~1 mHz at a 1-second update rate. For grid monitoring applications, this 1-second cadence exceeds the 30-per-second reporting rate of PMUs but provides sufficient temporal resolution to detect generation trip events (which manifest as frequency nadirs occurring 5–15 seconds after the trip), inter-area oscillation modes (which have periods of 2–10 seconds), and islanding events (which cause immediate frequency divergence from the bulk grid). Faster estimation (100 ms windows) is achievable with reduced frequency precision (~10 mHz), still adequate for detecting major disturbances.

2. On-Device ENF Extraction Neural Network (GridHum)

The system deploys a compact neural network called GridHum on each participating consumer IoT device. GridHum extracts the ENF signal from the raw microphone input and transmits only the extracted frequency, phase, and signal quality metrics — never raw audio — to the aggregation service. This architecture is a hard privacy constraint: no audio content, speech, music, or ambient sound leaves the device.

Input processing: The raw microphone signal (typically sampled at 16 kHz for smart speakers or 8 kHz for security cameras) is bandpass-filtered to isolate a narrow band around the nominal frequency and its first three harmonics: 55–65 Hz, 115–125 Hz, 175–185 Hz, and 235–245 Hz. The bandpass filtering is implemented as a cascade of second-order IIR sections (total: 8 biquad stages) requiring fewer than 50 multiply-accumulate operations per sample. The filtered signal is decimated to 500 Hz (retaining all ENF information while reducing compute by 32×) and segmented into 1-second frames with 50% overlap, yielding 2 ENF estimates per second.

Architecture: GridHum processes each 1-second frame through three stages:

  1. Harmonic feature extractor: A 1D convolutional network (3 layers, 16/32/32 channels, kernel size 7, stride 2) processes the 4-channel bandpass-filtered signal (fundamental + 3 harmonics) into a 64-dimensional feature vector. The multi-harmonic input is critical: in scenarios where the fundamental is obscured by mechanical vibration at exactly 60 Hz (fans, motors), the 120 Hz and 180 Hz harmonics remain uncontaminated and carry the same frequency deviation information. Total parameters: ~12K.
  2. Noise-robust frequency estimator: A 2-layer fully connected network (64 → 32 → 3) outputs three scalar values: estimated frequency deviation from nominal (Δf, in millihertz), estimated rate of change of frequency (ROCOF, in mHz/s), and a confidence score (0–1) indicating the signal quality of the current estimate. The confidence score is trained to predict the estimation error magnitude, enabling the cloud aggregator to weight contributions from high-quality devices more heavily. Total parameters: ~3K.
  3. Phase tracker: A lightweight phase-locked loop (PLL) implemented in fixed-point arithmetic tracks the instantaneous phase of the ENF fundamental across frames, outputting the unwrapped phase angle modulo 2π. The PLL bandwidth is set to 0.5 Hz, fast enough to track grid frequency transients while rejecting phase noise from the microphone circuit. The PLL state (phase, frequency) is preserved across frames, providing continuous phase tracking as long as the device is powered. Total phase tracker state: 12 bytes.

Total model size: ~15K parameters, quantized to INT8: 15 KB. Combined with the IIR filter coefficients and PLL state, the complete GridHum module occupies under 20 KB of device memory. Inference latency: under 2 ms per frame on the low-power DSPs (Tensilica HiFi, ARM Cortex-M4F) found in consumer IoT devices. Additional power draw: under 0.5 mW, negligible compared to the always-on microphone front-end that these devices already operate.

Output telemetry packet: Each 0.5-second cycle, GridHum emits a 32-byte telemetry packet containing: device ID (8 bytes, pseudonymized), timestamp (4 bytes, GPS-derived or NTP-synchronized to ±10 ms), latitude/longitude (8 bytes, from device registration, not GPS — most IoT devices have fixed, known locations), estimated Δf (4 bytes, float32, millihertz), ROCOF (4 bytes, float32, mHz/s), phase angle (2 bytes, uint16, 0–65535 mapping to 0–2π), and confidence (2 bytes, uint16). At 2 packets per second, each device contributes 64 bytes/s or 5.5 KB/day — less bandwidth than a single web page load. Even aggregating 10 million devices, total data inflow is 640 MB/s, manageable for a standard cloud ingest pipeline.

Training: GridHum is trained on a synthetic dataset generated by superimposing simulated ENF signals (with realistic frequency excursion profiles drawn from publicly available grid frequency databases, such as the open dataset published by Rydin Gorjão et al. (2020, Nature Communications) covering 12 synchronous areas) onto real-world ambient audio from the AudioSet corpus (spanning speech, music, environmental sounds, silence). The ENF signal is injected at varying SNR levels (-10 to -50 dB) and with varying harmonic distortion profiles to simulate different device types and coupling conditions. Ground truth frequency is known from the simulation. The model is validated on real recordings from consumer devices with concurrent PMU reference data.

3. Cloud-Based Geospatial Grid State Estimator

The cloud service aggregates streaming ENF telemetry from millions of devices into a real-time model of the grid's frequency and phase state across its geographic footprint.

Spatial aggregation: Devices are grouped into geographic cells using a hierarchical spatial index (H3 hexagonal grid at resolution 7, corresponding to ~5 km² per cell). Within each cell, ENF estimates from all contributing devices are fused using a weighted median filter, where weights are the confidence scores reported by GridHum. The median filter is robust to outliers from malfunctioning devices, devices with strong local interference (e.g., a variable-frequency drive motor near the microphone), and adversarial injection attempts. With typical urban IoT density of 50–200 devices per km², each 5 km² cell aggregates 250–1,000 independent ENF estimates per second, driving the effective frequency estimation noise floor below 0.1 mHz — comparable to a dedicated PMU.

Temporal smoothing: Within each cell, a Kalman filter tracks the frequency state (Δf, ROCOF, and second derivative d²f/dt²) using a constant-jerk kinematic model. The Kalman filter optimally combines the rapid but noisy per-device estimates with the smooth but lagged temporal model, producing frequency estimates with sub-0.1 mHz precision at 2 Hz update rate under typical conditions. During rapid transients (generation trips, line faults), the Kalman filter's innovation sequence triggers an adaptive gain increase that trades precision for responsiveness, tracking the frequency nadir with latency under 1 second.

Grid topology overlay: The geospatial frequency map is overlaid with the known electrical topology of the grid (transmission lines, substations, generator locations, interconnection tie lines). This overlay transforms a purely geographic frequency map into an electrically meaningful one. A frequency depression centered on a geographic cell near a known generator indicates a local generation loss. A frequency gradient aligned with a known transmission corridor indicates a power flow change. An abrupt phase discontinuity at a known substation boundary indicates a switching event or relay operation.

Anomaly detection pipeline: The state estimator feeds a suite of anomaly detectors, each tuned to a specific grid disturbance type:

4. Privacy Architecture

The system is designed around a strict privacy constraint: no audio content ever leaves the device. The GridHum neural network runs entirely on-device and outputs only numerical ENF measurements (frequency, phase, confidence). These measurements carry zero speech, music, or environmental audio content — they encode only the 60 Hz mains hum, which is identical for every device on the same electrical feeder and contains no information about what is happening in the home.

Formal privacy guarantee: The ENF telemetry packet (Δf, ROCOF, phase, confidence) is mathematically a 4-dimensional time series of the local grid frequency, a physical quantity determined entirely by the grid's generation-load balance, not by any activity in the home. Two devices on the same feeder produce statistically identical ENF telemetry regardless of whether one is in a nursery and the other is in a garage. The telemetry is further aggregated at the cell level before any analysis, ensuring that individual device contributions are never examined in isolation.

Differential privacy layer: To protect against theoretical side-channel attacks (e.g., inferring that a specific device is present or absent from the aggregate by selectively withdrawing and observing the aggregate change), the system adds calibrated Laplacian noise (ε = 1.0) to each device's reported Δf before transmission. This noise (σ ≈ 2 mHz) is negligible compared to the measurement noise of individual devices (~5 mHz) but provides formal (ε, δ)-differential privacy guarantees for individual device contributions to the aggregate.

Location privacy: Device locations are registered at the H3 resolution 7 cell level (~5 km²), not at the exact address. The device ID is a pseudonymous token that rotates monthly. The system operator cannot map a device ID to a specific household.

5. Calibration and Validation

Self-calibration: The system exploits the physical coherence of grid frequency to self-calibrate without any external reference. Within any geographic cell, all devices should report the same frequency (within measurement noise). Systematic bias in a device's frequency estimate (e.g., due to a clock drift in its audio sampling hardware) manifests as a constant offset from the cell median and is automatically corrected by the aggregator. A device that consistently reports frequency 5 mHz above its neighbors has its estimates shifted down by 5 mHz — the grid itself serves as the calibration reference.

Cross-validation against PMUs: In areas where PMUs are installed, the consumer audio network's estimates can be continuously compared against the PMU reference. Early deployments would use this cross-validation to quantify estimation accuracy and refine the GridHum model. As confidence in the consumer network grows, the system can extend coverage into areas with no PMU presence — the distribution grid, rural areas, developing nations — where grid visibility is most needed.

Intentional test events: Grid operators routinely perform switching operations and generation dispatch changes that produce known frequency perturbations. By correlating the consumer network's observed frequency response with the operator's known actions, the system validates its detection sensitivity and localization accuracy under controlled conditions.

6. Deployment Scenarios

7. Figures Description

Claims

  1. A system for monitoring the state of an electrical power grid, comprising: a plurality of consumer IoT devices, each containing a microphone capturing ambient audio in which an electrical network frequency (ENF) signal is embedded through electromagnetic coupling or acoustic radiation from the AC power grid; an on-device signal processing module on each device that extracts the instantaneous grid frequency, rate of change of frequency, and phase angle from the captured audio without transmitting raw audio; and a cloud-based aggregation service that receives ENF telemetry from the plurality of devices, groups measurements by geographic cell, and produces a spatially resolved real-time estimate of the power grid's frequency state across its geographic footprint.
  2. The system of claim 1, wherein the on-device signal processing module comprises a bandpass filter isolating the ENF fundamental frequency and at least two harmonics, a neural network that processes the multi-harmonic filtered signal to estimate frequency deviation from nominal with sub-millihertz precision, and a phase-locked loop that tracks the instantaneous phase of the ENF fundamental across consecutive frames, with total module size not exceeding 200 KB and inference latency under 5 ms per frame.
  3. The system of claim 1, wherein the cloud-based aggregation service groups device measurements using a hierarchical hexagonal spatial index, applies a weighted median filter within each cell using device-reported confidence scores as weights, and tracks the per-cell frequency state using a Kalman filter with a kinematic frequency model that adapts its gain during rapid transients to trade precision for responsiveness.
  4. The system of claim 1, further comprising a generation trip detector that identifies sudden frequency drops exceeding a threshold magnitude observed coherently across multiple geographic cells, estimates the geographic location of the generation loss from the centroid of the earliest-detecting cells, and estimates the magnitude of lost generation from the frequency nadir depth using the grid's known frequency response characteristic.
  5. The system of claim 1, further comprising an inter-area oscillation detector that applies spatial Fourier decomposition to the geographic frequency map to extract electromechanical oscillation mode shapes, frequencies, and damping ratios, and generates early warning alerts when a mode's damping ratio falls below a configurable threshold.
  6. The system of claim 1, further comprising an islanding detector that identifies geographic clusters of devices whose frequency trajectories diverge from surrounding devices, indicating electrical isolation of a portion of the distribution grid, and estimates the boundary of the islanded region from the geographic extent of the divergent cluster.
  7. The system of claim 1, wherein the on-device module processes ENF harmonics (at 2f, 3f, and 4f of the nominal grid frequency) in addition to the fundamental, extracting harmonic distortion ratios that encode distribution-level power quality information, enabling detection of local events including transformer tap changes, capacitor bank switching, and the connection or disconnection of large nonlinear loads, which are invisible to transmission-level PMU monitoring.
  8. The system of claim 1, wherein the aggregation service exploits the physical coherence of grid frequency across nearby devices to self-calibrate individual device measurements without any external reference signal, detecting and correcting systematic frequency estimation bias in individual devices by comparing each device's estimates against the robust median of its geographic cell.
  9. A method for privacy-preserving grid monitoring using consumer device audio, comprising: capturing ambient audio on a consumer IoT device; processing the audio entirely on-device through a neural ENF extraction pipeline that isolates the grid frequency signal from all other audio content; transmitting only numerical ENF measurements (frequency deviation, rate of change, phase angle, confidence score) to a remote service; and adding calibrated differential privacy noise to the frequency measurements before transmission, ensuring that individual device contributions cannot be distinguished from the aggregate through participation inference attacks.
  10. The system of claim 1, wherein device locations are registered at a coarse geographic resolution (cell area of at least 1 km²) and device identifiers are pseudonymous tokens that rotate periodically, preventing the system operator from mapping device telemetry to specific households while maintaining sufficient spatial resolution for grid state estimation.
  11. A method for estimating the spatially distributed rotational inertia of a power grid using the system of claim 1, comprising: measuring the rate of change of frequency (ROCOF) across all geographic cells during known or detected generation disturbance events; computing the ratio of disturbance magnitude to local ROCOF for each cell; and mapping the resulting inertia estimates to grid topology to identify regions where declining inertia from renewable generation displacement increases vulnerability to frequency instability.
  12. The system of claim 1, wherein battery-powered consumer devices that receive the ENF signal only through environmental electromagnetic radiation and acoustic coupling (without direct AC power connection) contribute to the monitoring network, with the aggregation service assigning lower confidence weights to these devices' measurements to account for their typically weaker ENF signal strength.

Prior Art References

  1. Grigoras, International Journal of Speech, Language and the Law 2005 — Digital audio recording analysis: the electric network frequency criterion. Foundational work establishing ENF extraction from audio recordings for forensic timestamp authentication.
  2. Ojowu et al., IEEE Trans. Inform. Forensics and Security 2012 — ENF extraction from digital recordings using adaptive techniques and frequency tracking. Demonstrated ENF extraction at -30 dB SNR with ±2 mHz accuracy.
  3. Bykhovsky and Cohen, IEEE Trans. Inform. Forensics and Security 2013 — ENF maximum-likelihood estimation via a multitone harmonic model. Achieved Cramér-Rao bound performance exploiting multi-harmonic structure.
  4. Rydin Gorjão et al., Nature Communications 2020 — Open database analysis of scaling and spatio-temporal properties of power grid frequencies. Open dataset of grid frequency measurements across 12 synchronous areas.
  5. Liu et al., IEEE Power and Energy Magazine 2006 — Wide-area monitoring system (FNET/GridEye) using low-cost frequency disturbance recorders. Demonstrated generation trip detection with 80 FDR units across the Eastern Interconnection.
  6. Lab11/GridWatch, University of Michigan — Grid monitoring system using smartphone accelerometers and microphones to detect power outages. Binary outage detection only, no continuous frequency monitoring or grid state estimation.
  7. Norouzian et al., Scientific Reports 2024 — Intra-grid location estimation of smartphone videos using ENF extraction and improved super-pixel technique. ENF-based video forensic localization within a grid.
  8. Wikipedia: Electrical network frequency analysis — Overview of ENF forensics history, including Metropolitan Police (UK) ENF database maintained since 2005 and Bavarian Police database since 2010.
  9. Derviskadic et al., IEEE Trans. Power Systems — PMU measurement architecture and cost analysis. PMU installation costs $40K–$100K per unit including communications infrastructure.
  10. North American SynchroPhasor Initiative (NASPI) — Industry consortium tracking PMU deployment across North American interconnections. Approximately 2,500 PMUs deployed as of 2024.
  11. Milano et al., Electric Power Systems Research 2021 — Foundations and challenges of low-inertia systems. Analysis of declining grid inertia from renewable displacement and implications for frequency stability.
  12. Cui et al., IEEE Trans. Smart Grid 2018 — Machine learning-based anomaly detection for power system frequency. Neural approaches to grid disturbance classification from frequency time series.