Session: 6.1.2 - Student Competition

Paper Number: 109005

109005 - Audio-Based Classification of Swirl Combustion Regimes Using Deep Learning 

Deep learning-based prediction of combustion systems is of growing interest in the combustion and power energy conversion systems. Such predictions have potential to detect the onset of different combustion regimes and early predictions of combustion instability and increased levels of noise emission during gas turbine operation. Advanced artificial intelligence-based combustion control system is conceptualized with similar deep learning classification algorithms. Simultaneous image (high-speed chemiluminescence imaging) and audio feature-based recognition of swirl assisted combustion will be reported in this paper using a lab-scale swirl combustor at a heat release intensity of 5.72 MW/m3-atm. The main focus is to classify conventional swirl-assisted combustion from low-oxygen concentration reaction zones leading to distributed combustion regime using methane as the fuel. Previous studies in this area were performed and reported at the Combustion Laboratory, University of Maryland, to classify swirl-assisted combustion and distributed combustion using convolutional neural network (CNN) based on image features that included flame shape, chemiluminescence intensity, and stand-off distance from the burner exit. However, flame images may not always classify combustion regimes due to highly complex, non-linear dynamics, and geometrical dependence of combustion systems. Hence, a combined image and audio-based CNN is conceptualized and utilized in this research. Two simultaneous CNN models were developed using flame chemiluminescence images and acoustic signatures obtained using microphone signals in Google Colab platform (Tensorflow). Results were verified using the Teachable Machine software which is based on MobileNet that aims to efficiently develop light weight deep learning models for mobile applications using depth wise separable convolutions. While MobileNet focuses on small model development, the Teachable Machine utilizes transfer learning technique. Transfer learning is efficient in transferring the knowledge from a pre-trained model to another model, which is generally of higher complexity and it requires training using new available data. Transfer learning eliminates the requirement of initiating a fresh learning scheme every time with different datasets. The mathematical background of such learning was explained as D = {X, P(X)} with X as the feature space and P(X) as marginal probability such that, P(X) = {x₁, x₂, ..., x_n} for x_i  X. With such domain definition, a task T can be defined as T = {y, P(y|X)} = {y, η}, where η is the objective function which can be denoted by P(y|X) as well. Now, if Dt, Tt represent the target domain and respective target task then D_s, Ts be the source domain with respective source tasks. The transfer learning aims at learning the target probability distribution P(y_t|X_t) in Dt with information gained from Ds and T_s (D_s ≠ D_t or T_s ≠ T_t). The broader goal is to develop a deep learning classification model using both image and audio signatures collected from the swirl combustor. Detailed results obtained using chemiluminescent and acoustic signals, representing different operational conditons,will be presented in the full paper.

Presenting Author: Rishi Roy University of Maryland

Presenting Author Biography: Graduate Student