Abstract Marine engineering has boomed and many deep learning-based methods have been proposed for underwater acoustic target recognition. However, most of these methods are dedicated to develop more complex convolutional neural networks to achieve better performance. This results in these models being unable to be deployed to low cost and miniaturized automatic underwater vehicles. A novel lightweight attentional residual neural network with data augmentation and feature fusion is proposed in this paper. Mel Frequency Cepstral Coefficient (MFCC), delta-MFCC and delta–delta MFCC features are extracted in the time dimension for fusion to obtain the fusion feature. The SpecAugment data augmentation is also used to enhance the randomness and diversity of features by masking in time and frequency dimensions randomly. Shuffle attention in the residual blocks is introduced to enhance the representation of features. The lightweight model is evaluated and compared by using several metrics on ShipsEar and DeepShip datasets. The proposed lightweight model only requires 1.628 M parameters for the trained model. This work shows that the proposed method requires small memory storage, while it achieved comparative performance.

This article presents a neural network model based on edge intelligence for seismic event detection. We implemented the model in hardware using a field-programmable gate array (FPGA) to achieve in-situ detection of seismic events at acquisition nodes or edge nodes. We designed and implemented the model, focusing on its suitability for hardware implementation on FPGA, employing an encoder—decoder structure. The encoder incorporates reparameterization and depthwise separable convolutions. During training, a multibranch structure was employed, which was then converted to an equivalent single-branch structure during inference to reduce model complexity and parameters. The features extracted by the encoder were further learned by the bi-directional long short-term memory (Bi-LSTM) network and then fed into the decoder for classification. We evaluated the model using the stanford earthquake data set (STEAD) and observed a 70% reduction in parameters while achieving comparable detection performance to EQTransformer. Furthermore, the model structure is well-suited for hardware implementation on FPGA. Applying this model to edge devices for seismic event detection can effectively minimize redundant data transmission and enable in-situ quality control.

The identification and determination of seismic phase arrival times is a critical task in seismic data processing. In recent years, methods based on convolutional neural networks (CNN) and Transformer models have been widely applied in this field. While CNNs offer potential in seismic feature extraction, modeling limitations restrict their efficiency for detailed waveform analysis. Conversely, Transformer models, though powerful, are constrained by their quadratic computational complexity. Recent research has shown that the state space model (SSM) represented by Mamba can effectively simulate long-range interactions while maintaining linear computational complexity. Inspired by this, we propose the seismic phase picking mamba (SPPMamba) model. We have designed a novel Conv-SSM module that combines CNN layers’ local feature extraction capabilities with SSM’s long-range dependency-modeling abilities. This enables the model to effectively identify and utilize the temporal variations of seismic signals, enhancing the model’s analytical capabilities for dynamic seismic feature characteristics. To validate the performance of SPPMamba, we conducted experiments on public seismic datasets. Experimental results show that SPPMamba demonstrates superior performance in seismic phase picking. This study aims to lay a research foundation for developing seismic data processing algorithms based on SSM.