Flow regime identification from acoustic sensors in pipe flow using deep neural networks

In contemporary industrial processes, precise flow signal measurement is crucial for several reasons, such as enabling condition-based monitoring of pipeline health, preventing damage to expensive materials/equipment, minimizing environmental pollution, and ensuring the safety of operational personnel. In this work, a novel method for flow regime identification using advanced signal denoising techniques and deep neural networks is proposed. The time series data is collected from passive acoustic sensors that capture flow-generated acoustic signals on pipes within an experimental flow loop. The data is obtained under various flow conditions, with particular emphasis on slug flows, given their inherently high-noise nature. The denoising methods used here are an amalgamation of whitening procedures, autocorrelation filtering, bandpass filtering, and highly sophisticated deep neural networks. Notably, the first three methods are ingeniously adapted to facilitate the training of deep neural networks, eliminating the necessity for field data during the training process. Subsequently, the application of deep neural networks proves highly effective in refining signals associated with slug flows amidst extreme noise. This refinement allows for accurate determination of the slug regime type and its associated frequencies. Our approach to deep neural networks is particularly intriguing, as it leverages time–frequency segments for data denoising and incorporates artifical noise during training. This stands in contrast to existing neural network methodologies that predominantly focus on mapping inputs to outputs. Consequently, this technique introduces a cheap and fast method for flow lines monitoring where high levels of process noise are present, compared to existing extensive other approaches such as fiber optics monitoring or intrusive sensing. The experimental results demonstrate that the proposed technique benchmarking advanced methods, including the independent component analysis (ICA) and the eigenvalue decomposition (EGD) based approaches. In the presence of strong random noise, the presented methodology achieved an SNR of 30.2 dB, whereas the EGD and ICA methods attained SNR values of 15.3 dB and 17.5 dB, respectively, highlighting the superiority of the proposed approach.