• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Sunday, November 22, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Technology

Optical Microcavity and Neural Network Quantify Switchgear Decomposition Gases in Laboratory Testing

by
November 21, 2026
in Technology
Reading Time: 5 mins read
0
Optical Microcavity and Neural Network Quantify Switchgear Decomposition Gases in Laboratory Testing

Optical Microcavity and Neural Network Quantify Switchgear Decomposition Gases in Laboratory Testing

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

Researchers have developed an optical sensing and deep-learning framework that simultaneously determines the concentrations of sulfur dioxide, hydrogen sulfide, and carbon tetrafluoride produced by the decomposition of sulfur hexafluoride in electrical equipment. In gas-insulated switchgear installations across high-voltage and ultra-high-voltage transmission networks, sulfur hexafluoride serves as an insulating and arc-extinguishing medium. When internal partial discharge or overheating faults emerge, the gas breaks down and reacts with trace moisture, oxygen, and solid insulation to generate specific chemical decomposition products. Measuring these decomposition components provides diagnostic insight into internal insulation degradation, but online optical quantitative monitoring has faced persistent obstacles due to overlapping spectral responses, environmental fluctuations, and the computational limits of embedded hardware deployed in field bypass monitoring systems.

To address these measurement challenges, the study combines an all-fiber whispering-gallery-mode optical microcavity probe with a multi-scale residual one-dimensional convolutional neural network termed MSR-1D-CNN. The microcavity design utilizes optical resonance within a silica structure having a measured diameter of approximately 320 micrometers, coupled to a tapered optical fiber with a waist diameter of approximately 2.0 micrometers. When a tunable near-infrared laser continuously scans across the microcavity through the tapered fiber, light confinement within the resonator yields a measured loaded quality factor of approximately 5.2 × 105. Exposure to target gas mixtures induces coupled perturbations in resonance wavelength, linewidth, and resonance depth, transforming complex molecular interactions into measurable one-dimensional transmission spectra.

Spectral excitation and signal acquisition operate at a frequency of 10 Hz, where transmitted optical signals undergo photodetection, amplification, low-pass filtering, and analog-to-digital conversion to produce 200-point spectral frames. Synchronously, dedicated environmental sensors record ambient temperature, relative humidity, and absolute pressure to form an integrated multi-parameter data vector. The raw spectra undergo baseline correction, outlier elimination, and pointwise min-max normalization, which reduces the influence of slow laser power drift and detection gain shifts while preserving essential spectral profile variations. Every ten consecutive spectral frames are averaged without overlap, generating standardized inputs that fuse normalized spectral descriptors with standardized environmental variables calculated strictly from training statistics.

The neural network architecture processes the continuous one-dimensional resonance profiles by addressing spectral variations occurring across disparate spatial extents. Because gas concentration changes cause resonance shifts, linewidth broadening, and asymmetric profile distortions across different spectral intervals, the network incorporates three parallel convolutional branches equipped with kernel sizes of 3, 5, and 7 to extract complementary receptive-field features. To constrain computational overhead for edge devices, each branch utilizes depthwise-separable convolutions that decouple spatial spectral filtering from inter-channel mixing. The concatenated multi-scale outputs are merged through pointwise convolution, while residual shortcut connections preserve lower-level spectral information across three successive residual stages featuring 32, 64, and 96 output channels before global average pooling descriptors enter a 64-neuron regression layer.

Model training and evaluation relied on carefully partitioned experimental data derived from independent gas-preparation batches rather than correlated individual spectral frames. The baseline concentration-grid dataset incorporated 375 independently mixed gas batches spanning 0 to 100 ppm for each component in high-purity sulfur hexafluoride balance gas, yielding 22,500 spectral inputs. These batches were partitioned into 263 training, 56 validation, and 56 internal testing batches, ensuring that temporally correlated spectral frames from any given gas mixture remained confined to a single partition. An auxiliary environmental dataset covering temperatures from 15 to 45 degrees Celsius, relative humidity levels from 30 to 60 percent, and absolute pressures from 0.09 to 0.12 megapascals provided additional batches to train the model against ambient baseline shifts.

During evaluation on the internal test set, the proposed neural network demonstrated quantitative regression performance, achieving a root mean squared error of 1.82 ± 0.05 ppm, a mean absolute error of 1.52 ± 0.04 ppm, and a coefficient of determination of 0.995 ± 0.001 across five random training seeds. When benchmarked against traditional linear and nonlinear multivariate methods under identical preprocessing and partitioning protocols, partial least-squares regression yielded the largest prediction errors, highlighting the limitations of linear latent-variable projections for overlapping optical resonance responses. Support vector regression and conventional one-dimensional convolutional networks improved accuracy, while specialized sequential architectures such as ResNet1D and InceptionTime-R recorded root mean squared errors of 2.34 ppm and 2.12 ppm, respectively, confirming the advantage of combining multi-scale filtering with residual information paths.

Controlled architectural ablation experiments isolated the functional necessity of each structural element within the network. Replacing the parallel multi-scale branches with a single fixed kernel size of 5 increased the overall root mean squared error from 1.82 ppm to 2.20 ppm, verifying that multiple receptive fields capture localized and distributed spectral distortions more effectively. Removing the residual connections raised the root mean squared error to 2.07 ppm, demonstrating that shortcut pathways help retain subtle low-level resonance variations through deeper network layers. Substituting depthwise-separable convolutions with standard convolutions maintained comparable prediction accuracy but increased the parameter count from 0.186 million to 0.412 million, demonstrating that depthwise separation primarily serves computational efficiency rather than raw predictive precision.

To evaluate stability against temporal correlation and real-world batch variations, researchers conducted batch-level aggregations and bootstrap resampling with 1,000 iterations using independent gas batches as the resampling units. The batch-level metrics aligned closely with sample-level findings, confirming that performance was not an artifact of repeated intra-batch frames. In a dedicated generalization trial using 24 non-grid concentration combinations prepared independently on two different dates, the model recorded an average sample-level root mean squared error of 2.14 ppm, with component-specific errors remaining below 2.3 ppm and determination coefficients exceeding 0.990. Hydrogen sulfide exhibited slightly higher inversion errors than sulfur dioxide and carbon tetrafluoride, which authors associated with its comparatively slower desorption recovery dynamics and greater cross-sensitivity.

Environmental robustness assessments revealed that ambient variations introduce spectral perturbations that overlap with gas-induced shifts, necessitating multi-parameter input compensation. When evaluated on held-out environmental conditions completely excluded from model training—specifically 55 degrees Celsius, 70 percent relative humidity, and 0.13 megapascals absolute pressure—the full model achieved root mean squared errors of 2.26, 2.14, and 2.31 ppm, respectively. In contrast, an ablated network using spectral inputs alone produced an overall error of 4.94 ppm across the environmental test dataset. Progressively integrating temperature, relative humidity, and pressure reduced the test error to 4.12, 3.25, and 2.17 ppm, illustrating that environmental fusion mitigates baseline ambiguity while remaining subject to moderate degradation when operating outside calibrated boundaries.

For embedded implementation, the 32-bit floating-point model was deployed onto an STM32H743ZIT6 microcontroller featuring an Arm Cortex-M7 core operating at 480 MHz. In continuous processing benchmarks spanning 10,000 frames following a 100-frame warm-up, the complete preprocessing and model inference pipeline required an average execution latency of 37.8 milliseconds per input, with a 95th-percentile execution latency of 39.6 milliseconds. During a 24-hour continuous stability assessment, the microcontroller executed 864,000 consecutive spectral frames without experiencing program resets or data frame loss, confirming firmware stability, though researchers noted that overall system concentration update rates remain constrained by upstream optical scanning and physical gas-exchange dynamics.

The authors emphasized several experimental boundaries governing the present findings. The concentration-grid experiments utilized 25 ppm as the lowest independently prepared non-zero concentration, meaning the study does not establish a formal limit of detection or limit of quantitation under standard analytical definitions. Furthermore, all evaluations were conducted within controlled laboratory mixing chambers using a single microcavity probe, leaving probe-to-probe manufacturing variability, cross-probe calibration transfer, long-term chemical aging, and performance under complex multi-factor environmental variations to future field investigations on operating gas-insulated switchgear equipment.

Subject of Research: Optical whispering-gallery-mode sensing and convolutional neural network concentration inversion for sulfur hexafluoride decomposition gases

Article Title: GIS gas-component detection based on a WGM gas-sensing microcavity and a one-dimensional convolutional neural network

Article References: Meng, Y., Zhao, J., Xu, Y., Lin, J., Zhang, N., & Wang, J. (2026). GIS gas-component detection based on a WGM gas-sensing microcavity and a one-dimensional convolutional neural network. Journal of Engineering and Applied Science, 73(1), Article 430. https://doi.org/10.1186/s44147-026-01289-9

Image Credits: AI Generated

DOI: 10.1186/s44147-026-01289-9

Keywords: Gas-insulated switchgear, Sulfur hexafluoride decomposition, Whispering gallery mode, Optical microcavity, Convolutional neural network, MSR-1D-CNN, Multicomponent gas sensing, Sulfur dioxide, Hydrogen sulfide, Carbon tetrafluoride, Spectral inversion, Depthwise-separable convolution, Environmental compensation, Cortex-M7 edge inference, Condition monitoring, Spectroscopy

News Source: Cassandra Pierce. (November 21, 2026). Optical Microcavity and Neural Network Quantify Switchgear Decomposition Gases in Laboratory Testing. Scienmag.

Tags: Carbon tetrafluoridecondition monitoringconvolutional neural networkCortex-M7 edge inferenceDepthwise Separable ConvolutionEnvironmental compensationGas-insulated switchgearHydrogen sulfideMSR-1D-CNNMulticomponent gas sensingoptical microcavitySpectral inversionspectroscopysulfur dioxideSulfur hexafluoride decompositionwhispering-gallery-mode
Share12Tweet7Share2ShareShareShare1

Related Posts

Optimized Neural Network Cuts Error in Objective Vocal Music Performance Evaluation

Optimized Neural Network Cuts Error in Objective Vocal Music Performance Evaluation

November 22, 2026
Tensegrity Continuum Robots Reconfigure Morphologies for Cooperative Locomotion and Manipulation

Tensegrity Continuum Robots Reconfigure Morphologies for Cooperative Locomotion and Manipulation

November 22, 2026

Neural Network With Sliding Data Window Enhances Autonomous Ship Path Following

November 22, 2026

Crayfish Optimization Paired with Gaussian Regression Accurately Predicts Mine Blasting Fragmentation

November 22, 2026

POPULAR NEWS

  • RAGE, Not TLR4, Drives Inflammatory Pain in Human Skin

    RAGE, Not TLR4, Drives Inflammatory Pain in Human Skin

    29 shares
    Share 12 Tweet 7
  • Phloem Girdling Drives Rhizosphere Microbial Adaptation to Belowground Carbon Limitation in Tobacco

    29 shares
    Share 12 Tweet 7
  • Direct Integration of GIS and AI in Tuberculosis Studies Remains Rare

    29 shares
    Share 12 Tweet 7
  • Cortical Folding Motifs Reveal Distinct Brain Shape Variations and Asymmetry From Infancy

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Loneliness Strongly Associated With Depression Among Older Adults in Golestan Iran

Self-Powered Neuroprosthesis Harvests Walking Energy to Improve Mobility After Stroke

Comparative Genomics Unveils Evolutionary Lineages and Stress Genes Across Four Vigna Legumes

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 89 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.