underwater explosion, feature extraction, deep learning, key signal recognition, Naval architecture. Shipbuilding. Marine engineering, VM1-989

The amount of data collected from underwater explosion tests is enormous, which is mixed with a large amount of useless data. To protect the data from the effects of the explosion, it is crucial to prioritize the recognition and storage of key data during the test. In response to this, a key signal recognition model that combined feature extraction methods with deep learning models was proposed to improve the accuracy of key signal recognition. Firstly, different preprocessing methods for removing trend components from underwater explosion acceleration signals were studied. Existing test results demonstrated that wavelet packet decomposition, empirical mode decomposition, and high-pass filtering could significantly enhance the model’s recognition performance. Secondly, to make the extracted features more conducive to distinguishing between explosion and non-explosion segments, a feature extraction method based on the inter-class variance ratio for underwater explosion acceleration signals was proposed. According to the actual measured underwater explosion acceleration signal data, it was found that compared to Log Mel features, the proposed features improved classification accuracy by approximately 4.92% using the K-means method. Finally, the ECAPA-TDNN model incorporating the SE-Res2Block module was introduced, ensuring better recognition accuracy. With the proposed features as input, the recognition accuracy reached 99.31%.