Partial Discharge Detection

PD events generate ultrasonic pressure waves in the 20–500 kHz range as the discharge creates a microscopic explosion within the insulation. Piezoelectric sensors mounted on the transformer tank wall detect these mechanical vibrations and triangulate the discharge source using time-of-flight calculations across multiple sensor positions.

• Frequency range: 20–500 kHz, above audible noise
• Localization accuracy: ±10 cm with 4+ sensor arrays
• Key advantage: Immune to electrical interference from substation equipment
• Limitation: Signal attenuation through oil-paper barriers complicates internal winding PD detection

PD pulses generate electromagnetic waves in the 300 MHz–3 GHz range due to the extremely fast rise time of the discharge current, typically under 1 nanosecond. UHF antennas installed inside the transformer tank or through dielectric windows capture these signals, providing exceptional signal-to-noise ratio and immunity to external corona.

• Frequency range: 300 MHz–3 GHz
• Rise time: < 1 ns for typical void discharges
• Key advantage: Highly sensitive to low-level PD in complex insulation structures
• Limitation: Requires internal sensor installation or dedicated dielectric windows

The apparent charge of a PD event is measured in picocoulombs (pC) using a coupling capacitor and detection impedance connected to the bushing tap or high-voltage conductor. This is the only method standardized under IEC 60270 for quantitative PD magnitude assessment, providing calibrated measurements traceable to national standards.

• Measured quantity: Apparent charge in pC
• Sensitivity threshold: Typically < 5 pC in shielded environments
• Key advantage: Quantifiable, repeatable, and legally recognized for acceptance testing
• Limitation: Susceptible to external noise in energized substations; requires offline or shielded conditions

PD in oil-paper insulation decomposes the dielectric fluid, generating characteristic fault gases—primarily hydrogen (H₂) and methane (CH₄) —detectable through Dissolved Gas Analysis (DGA). A sharp rise in hydrogen concentration with low levels of acetylene strongly indicates PD activity rather than thermal faults or arcing.

• Key gas: Hydrogen (H₂), with methane (CH₄) as secondary indicator
• Diagnostic ratio: High H₂/CH₄ ratio with negligible C₂H₂ suggests PD
• Key advantage: Integrates with existing DGA monitoring infrastructure
• Limitation: Slow response; gas accumulation lags behind PD onset by hours to days

PRPD mapping plots PD pulse magnitude and repetition rate against the 50/60 Hz AC phase angle, creating distinctive visual patterns that fingerprint specific defect types. Void discharges in solid insulation produce symmetric patterns near voltage zero-crossings, while surface discharges show asymmetric distributions near voltage peaks.

• Pattern types: Void (symmetric), surface (asymmetric), corona (peak-aligned)
• Analysis method: Statistical operators like skewness and kurtosis quantify pattern shapes
• Key advantage: Enables defect classification without physical inspection
• Limitation: Requires phase-reference voltage signal synchronized with PD measurements

PD emits ultraviolet and visible light photons during the ionization and recombination process within the discharge channel. Fiber-optic sensors inserted into transformer oil channels or embedded within windings detect this optical emission directly, offering complete immunity to electromagnetic interference.

• Detection spectrum: UV (200–400 nm) and visible light
• Sensor type: Fluorescent optical fibers or photomultiplier tubes
• Key advantage: Zero EMI susceptibility; ideal for HVDC converter transformers
• Limitation: Requires direct line-of-sight to the discharge site; invasive installation

An unsupervised neural network architecture trained exclusively on normal transformer operational data. When partial discharge begins, the reconstruction error between input and output increases significantly, flagging the anomaly without requiring labeled fault data.

• Training data: Uses historical sensor readings during healthy operation periods
• Detection mechanism: High reconstruction error signals deviation from learned normal patterns
• Advantage: Identifies novel PD signatures not present in training datasets
• Implementation: Deployed on edge AI hardware within substation gateways for real-time inference

An off-line diagnostic test that compares the transfer function of a transformer winding over a wide frequency range. While FRA primarily detects mechanical deformation, severe partial discharge activity that erodes winding insulation can alter the capacitive and inductive signature of the winding.

• Sweep range: Typically 20 Hz to 2 MHz
• Comparison methods: Time-based, type-based, and phase-based comparisons identify deviations
• Sensitivity: Detects winding displacement, core movement, and turn-to-turn short circuits
• Complementary role: Confirms physical damage resulting from prolonged PD activity

A set of methods applied to transformer fault classification models to provide asset managers with interpretable feature attributions justifying specific maintenance alerts. When a model flags PD activity, XAI techniques reveal which sensor inputs most influenced the decision.

• SHAP values: Quantify each feature's contribution to a specific prediction
• LIME: Generates local surrogate models to explain individual classifications
• Use case: Validates that PD alerts are based on genuine acoustic or electrical signatures, not sensor noise
• Trust building: Enables engineers to verify model reasoning before committing to costly outages