The investigation began with field sampling rather than laboratory guesswork. The researchers collected glass insulator samples from 220 kV jumper and suspension strings operating for years in metallurgical environments, alongside samples from 500 kV and 110 kV lines exposed to chemical dust, tar, and metal dust. Macroscopic inspection already told a story: contaminant accumulation was visible across the surfaces, but distribution was uneven. On the upper surfaces of the insulator sheds, contamination was relatively sparse and spread fairly uniformly. On the lower surfaces, particles concentrated in the grooves, with both the total amount and the unevenness of distribution exceeding what appeared above. This asymmetry matters because the lower grooves are precisely where moisture retention and surface conductivity combine to trigger pollution flashover.

To characterize the particles themselves, the team turned to scanning electron microscopy and X-ray spectrometry. The electron micrographs revealed contaminant particle diameters ranging from no less than 2 micrometers to no more than 20 micrometers, a size regime small enough to disperse widely across a surface yet large enough to distort local electric fields. X-ray analysis of eight subsamples drawn from different insulators and locations showed that although the elemental composition varied from sample to sample, metallic elements were detected in every single subsample, with total metallic content exceeding 1 percent in each case. Among the detected metals, iron stood out with a consistently high proportion, identifying it as the representative contaminant for the region. Earlier literature had already flagged aluminum and iron among insoluble metallic contaminants as notable reducers of pollution flashover voltage, so the field findings pointed squarely at iron as the species to study.

With iron confirmed as the target, the researchers constructed a laboratory platform to isolate its effects. Because glass insulators possess an axisymmetric structure, each shed unit approximates a fan-shaped cross-section, so the team used an isosceles triangular glass plate, 45 centimeters tall with a 30 centimeter base and 8 millimeter thickness, as a physically analogous specimen. Conductive tape at the apex and base allowed voltage application across both ends. Artificial contamination was prepared from sodium chloride as the soluble component and a mixture of diatomite and iron powder as the insoluble fraction. The non-soluble deposit density was held at 5 milligrams per square centimeter and the equivalent salt deposit density at 0.05 milligrams per square centimeter across all specimens, while the iron powder proportion within the insoluble fraction was varied from 10 to 60 percent. After air-drying for 24 hours, the plates were placed in a 1.2 meter acrylic tank where a humidifier introduced micron-scale water droplets to simulate saturated humidity, and specimens were left to absorb moisture for 15 minutes before voltage was applied stepwise.

The measurements delivered a striking result. With no metal in the contamination, the leakage current displayed a purely sinusoidal waveform, with high-frequency pulses appearing only when the electric field approached the critical breakdown threshold. When metal was present, the waveform changed at the same applied voltage: the AC component distorted and high-frequency pulses emerged superimposed near the AC voltage peaks. The mechanism is electrical field distortion. Metallic particles scattered across the pollution layer create localized field concentrations that promote partial discharge activity even under sub-critical field conditions, causing micro-discharges to initiate prematurely. Quantitatively, the pulse parameters told the story. Under conventional contamination, a typical pulse showed a rise time of 1.49 microseconds, a fall time of 2.68 milliseconds, and a full width at half maximum of 33.62 microseconds. With metal present, all of these compressed dramatically: rise time fell to 0.86 microseconds, fall time to 27.36 microseconds, and FWHM to 9.06 microseconds. The metal content also lowered the inception voltage at which pulses first appeared.

As the iron proportion climbed, the pulse statistics shifted in ways that could be exploited diagnostically. Within a one-second interval, the total number of pulses increased consistently with rising metal content, while the applied voltage had a comparatively weaker effect on pulse count. The maximum positive pulse amplitude also grew with metal proportion, and at high metallic content the amplitude began to vary sharply with voltage, possibly signaling a transition in discharge mode. Negative pulse amplitudes showed no clear consistent trend except under exceptionally high metal content and high voltage, where they increased significantly. Perhaps most dramatically, the pulse discharge quantity, calculated by combining average pulse amplitude and average pulse duration, revealed a threshold effect. At a metallic proportion of 30 percent, raising the voltage from 8 to 20 kilovolts increased discharge quantity only slightly, from 3.9 to 4.3 microcoulombs. At 60 percent metal, the same voltage increase drove discharge quantity from 5.0 to 248.5 microcoulombs, a nearly fiftyfold surge that indicates severe compromise of surface insulation once metal content crosses a critical level.

The root mean square value of the leakage current added a final layer of complexity. Under fixed soluble and non-soluble densities, the RMS current increased nonlinearly with metallic proportion, and the trend and magnitude were comparable to those observed when equivalent salt deposit density rose. This is a crucial finding because metallic contamination, although classified as non-soluble, behaves differently from typical non-soluble pollutants such as diatomite, which affect insulation indirectly through moisture retention. Metal particles directly alter surface conductivity, similar to soluble salts. Consequently, RMS-based single-parameter monitoring cannot distinguish metallic from salt-driven contamination and may lead to substantial misinterpretation. The researchers argue that combined time- and frequency-domain analysis, incorporating pulse onset voltage, pulse repetition rate, and waveform steepness, offers a reliable approach to differentiate metallic pollution from conventional contamination.

To automate this diagnosis, the team built a deep learning model they call ResNet18-SA, a ResNet-18 architecture augmented with a channel self-attention mechanism. The pipeline begins with one-second windows of raw leakage current data, amplitude-normalized to eliminate voltage-level influence. The signal passes through an initial 7 by 7 convolutional layer and a 3 by 3 max pooling layer, then through eight deep residual modules, each containing two 3 by 3 convolutional layers with skip connections that mitigate gradient vanishing. At the deep stage of the network, the self-attention mechanism adaptively recalibrates feature channel weights, enhancing sensitivity to key pulse features while suppressing irrelevant noise. A global average pooling layer compresses spatial dimensions, and a fully connected layer with softmax classification maps the learned features to three contamination grades: light from 0 to 20 percent, moderate from 20 to 40 percent, and heavy from 40 to 60 percent metal.

Training used 480 leakage current waveform images split 7:3 between training and validation sets, with a learning rate of 1 times 10 to the minus 4 and a batch size of 8. By the 1000th iteration the model converged, achieving 96 percent validation accuracy and reducing loss to 0.1. The confusion matrix on the validation set showed correct classification across all three contamination levels. Comparative benchmarks underscored the architecture’s value: a CNN-SA model reached only 71.23 percent accuracy, ResNet18 alone improved to 96.58 percent, and ResNet18-SA achieved 100 percent accuracy with only a slight increase in training time, from 1141 to 1161 seconds. The residual structure provides a stable foundation for extracting deep waveform features, while the self-attention module enables selective focus on high-frequency pulse regions and suppression of background noise, a synergy the authors say is precisely why the combined model outperforms both alternatives.

The implications extend beyond the laboratory. Insulators near industrial, mining, and coastal areas face the highest contamination risk, and pollution increases surface conductivity, raising the likelihood of partial discharges and dry band arcing that can cascade into grid-wide failures. Because metallic contamination mimics salt contamination in RMS measurements, grids in metallurgical regions may be underestimating their true risk with conventional monitoring. The ResNet18-SA model, reading pulsed current signals with robust accuracy, offers an automated path to insulator condition assessment that does not require visual inspection or power outages. The authors acknowledge room for improvement, noting that the model does not yet fully account for complex temporal dynamics or environmental interference, and they plan to introduce optimization algorithms such as genetic algorithms to automatically search for optimal hyperparameters. Still, the core achievement stands: a self-attention neural network that transforms the subtle electrical signatures of iron dust into a reliable, quantifiable verdict on insulator health.

Subject of Research: Research on insulator contamination component identification based on neural network

Article Title: Research on insulator contamination component identification based on neural network

Article References: Luo, Y., Liu, J., Wang, Y., Mei, Z., & Liu, X. (2025). Research on insulator contamination component identification based on neural network. Frontiers in Electronics, 6, Article 1680502. https://doi.org/10.3389/felec.2025.1680502

Image Credits: AI Generated

DOI: 10.3389/felec.2025.1680502

Keywords: Research, insulator, contamination, component, identification, based, neural, network, scientific research