The quantitative diagnosis of low-temperature plasmas (LTPs) is crucial in materials science and engineering. Optical emission spectroscopy (OES) provides real-time monitoring capabilities but struggles with high-dimensional, nonlinear data. This paper presents a two-stage model for spectral transition recognition. First, an autoencoder reduces high-dimensional OES data to low-dimensional latent features. Second, a multi-layer perceptron (MLP) identifies transitions using Bayesian optimization (BO) to tune hyperparameters. Experimental results demonstrate the model's superior recognition accuracy and reduced computational complexity, offering a robust solution for real-time plasma diagnostics.