Practical Application and Validation of Enhanced Acoustic Leak Detection through Data Augmentation

Leak detection in industrial plants is critical for safety and environmental protection. Traditional methods have limitations, especially in noisy environments. Background noise is a significant challenge in industrial leak detection. Training with real industrial noises on-site is impractical but crucial for effective detection. Background noise affects the Signal-to-Noise Ratio (SNR), which is critical for successfully implementing an acoustic leak detection system. Lower SNR means background noise can easily mask leak signals, complicating detection. Strategies to address this include transforming acoustic signals using Fourier Transformation and analyzing patterns through autocovariance and autocorrelation. Additionally, CNNs trained on leak sounds without background noise perform poorly in noisy environments, whereas those trained with background noises show improved accuracy. Real-world industrial environments, where background noises are louder and more varied, require system adaptations to each specific setting.

A previous not published paper (Quick et al., (2024)) developed an enhanced acoustic leak detection system using data augmentation to overcome background noise challenges, employing convolutional neural networks (CNN) and a mixup technique. The developed system uses two datasets: one for leak noises recorded in a laboratory and another for background noises recorded on-site. These datasets are combined using a mixup technique, creating realistic training data that mirrors the acoustic environment of industrial settings. The method leverages Short-Time Fourier Transformation (STFT) spectrograms and trains a CNN to distinguish between leak and non-leak sounds. The proposed method was tested using the IDMT-ISA-Compressed-Air (IICA) dataset, demonstrating detection accuracies of 92.4% for ventleaks and 86.4% for tubeleaks, outperforming other methods. The depth of the network architecture significantly impacted detection accuracy, with deeper networks like ResNet18 providing better results. The method showed improved performance over existing approaches, particularly in noisy environments. While laboratory tests have shown promising results, real industrial environments present unique challenges that cannot be fully replicated in controlled settings. Background noises in actual industrial plants are more varied and complex, which can affect the performance of the detection system. Therefore, it is essential to test and validate the enhanced leak detection method in real-world industrial plants to ensure its reliability and robustness in diverse operational conditions.

To validate the proposed methodology, a series of tests will be conducted at industrial plants. A specialized test facility has been constructed to generate industrial leak noises. This facility includes a pressure reservoir capable of reaching up to 7 bar and is equipped with two G.R.A.S. 146AE 1/2" microphones positioned up to 8 meters from the leak source. Additionally, the facility features a Coriolis mass flow measurement tool and interchangeable nozzles to simulate different leak geometries. These nozzles are 3D printed using stereolithography to ensure minimal tolerance.Background noises were recorded over two measurement days at Industrial Plant X and Industrial Plant Y. Furthermore, leaks were artificially generated and recorded at these plants. The same leak geometries were also reproduced under laboratory conditions for comparison.

The gathered laboratory leak data will be mixed with the background noises recorded from the industrial plants. These augmented superpositional data will undergo bandpass filtering and transformation into spectrograms via Short-Time Fourier Transformation (STFT). The input size of the spectrograms will be (384, 384) for the convolutional neural network used for training and validation. Subsequently, the recorded leak data from the industrial plants will be utilized to test the network's accuracy, recall, and precision, verifying the methodology. The findings from these tests will be presented at this conference.