A ppb-Level Ammonia Detection System Based on TD-FD(HHT)–CNN-TL Deep Learning Model

The recent exponential growth in the utilization of the amino fertilizer has resulted in the significant release of ammonia (<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\text {NH}_{{3}}$ </tex-math></inline-formula>) during its hydrolysis, causing serious safety risks or environmental pollution, thereby accurate <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\text {NH}_{{3}}$ </tex-math></inline-formula> detection is imperative under this condition. In this article, a novel dual-branch time-domain (TD)–frequency-domain (FD) Hilbert-Huang transform (HHT)–convolutional neural network (CNN)–transfer learning (TL) deep learning model combined with tunable diode laser absorption spectroscopy (TDLAS) is proposed to detect ammonia, where (FD) branch information obtained by HHT are extracted through 2-D CNN and TD branch features are extracted by the corresponding 1D-CNN. Subsequently, this dual-branch deep learning model is initially trained by plenty of simulation datasets, and then the pretrained model parameters are further fine-tuned by TL with finite experimental datasets. The final test results show that the system provides reliable <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\text {NH}_{{3}}$ </tex-math></inline-formula> detection with an average absolute error of 0.76 ppm and achieves the minimum detection limit of 0.03 ppm via the Allan variance analysis. Hence, it can be concluded that the system holds significant potential for <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\text {NH}_{{3}}$ </tex-math></inline-formula> detection applications, and the proposed model introduces an innovative idea for designing neural networks in subsequent spectral signal analysis.