Real-time time-resolved optical diagnostics with LSTM neural network and electrical signals

Abstract Optical diagnostics are crucial for revealing physical mechanisms in discharge plasmas and monitoring essential process parameters. However, existing optical measurement methods are often expensive, complex, and time-consuming, restricting their applicability in large-scale and real-time industrial applications. To address these challenges, this study introduces an innovative data-driven approach for real-time time-resolved plasma optical diagnostics. Leveraging advanced long short-term memory (LSTM) neural networks, this method accurately and efficiently predicts key dynamic optical parameters in nanosecond pulsed spark discharge (NPSD), including emission spectra and grayscale intensity, directly from easily accessible electrical signals measured by standard oscilloscopes, such as discharge voltage and current. Through feature extraction and data mapping, the trained LSTM network models demonstrate high predictive accuracy, with R 2 values of 0.955 76 for imaging average grayscale and 0.788 33 for emission spectral intensity at six wavelengths, validated through leave-one-out cross-validation across 30 experimental groups. Additionally, with a prediction time of less than 1 s per dataset group, this approach demonstrates the potential for efficient and precise optical diagnostics for industrial applications demanding real-time monitoring and time-resolved parameter analysis, such as plasma combustion optimization, pollution control, and advanced materials processing. By integrating artificial intelligence technology with plasma diagnostics, this work provides novel perspectives and valuable insights for advancing optical parameter diagnostics and control technology over plasma-based applications.