Abstract In addressing the challenge of ships being prone to occlusion in multi‐target situations during ship target detection, leading to missed and false detections, this paper proposes an enhanced ship detection algorithm for YOLOv7‐tiny. The proposed method incorporates several key modifications. Firstly, it introduces the Convolutional Block Attention Module in the Backbone section of the original model, emphasizing position information while attending to channel features to enhance the network's ability to extract crucial information. Secondly, it replaces standard convolution with GSConv convolution in the Neck section, preserving detailed information and reducing computational load. Subsequently, the lightweight operator Content‐Aware ReAssembly of Features is employed to replace the original nearest‐neighbour interpolation, mitigating the loss of feature information during the up‐sampling process. Finally, the localization loss function, SIOU Loss, is utilized to calculate loss, expedite training convergence, and enhance detection accuracy. The research results indicate that the precision of the improved model is 91.2%, [email protected] is 94.5%, and the F1‐score is 90.7%. These values are 3.7%, 5.5%, and 4.2% higher than those of the original YOLOv7‐tiny model, respectively. The improved model effectively enhances detection accuracy. Additionally, the improved model achieves an FPS of 145.4, meeting real‐time requirements.

With the advancements in technologies such as 5G, Unmanned Aerial Vehicles (UAVs) have exhibited their potential in various application scenarios, including wireless coverage, search operations, and disaster response. In this paper, we consider the utilization of a group of UAVs as aerial base stations (BS) to collect data from IoT sensor devices. The objective is to maximize the volume of collected data while simultaneously enhancing the geographical fairness among these points of interest, all within the constraints of limited energy resources. Therefore, we propose a deep reinforcement learning (DRL) method based on Graph Attention Networks (GAT), referred to as “GADRL”. GADRL utilizes graph convolutional neural networks to extract spatial correlations among multiple UAVs and makes decisions in a distributed manner under the guidance of DRL. Furthermore, we employ Long Short-Term Memory to establish memory units for storing and utilizing historical information. Numerical results demonstrate that GADRL consistently outperforms four baseline methods, validating its computational efficiency.

The automatic diagnosis of heart sounds is particularly important for cardiologists. However, the existing diagnostic methods still have a large space to be improved, In this paper, we proposed a novel method for heart sound classification. Our method consists of two stages. In the first stage, we preprocessed the heart sound signal, including two steps of denoising and downsampling, to reduce the noise and decrease the complexity of processing. In the second stage, we classify the processed signal, including framing and input network, and finally output three types of results. Our method was validated on the CirCor DigiScope Phonocardiogram Dataset. The result shows the F1 score reached 0.922 and is better compared to other networks’ results.

Cardiovascular disease is the most important cause of death in the world. In the early stage of cardiovascular disease, arrhythmia is often accompanied. Therefore, it is of great significance to carry out safe and effective arrhythmia detection for prevention and diagnosis of cardiovascular diseases. With the increase in widely available digital ECG data and the algorithmic paradigm of deep learning, multi-class arrhythmia classification based on automatic feature extraction of ECG has become increasingly attractive. However, the majority of studies cannot accept varied-length ECG signals and have limited performance in detecting multi-class arrhythmias. In this study, we propose a multi-label classification network based on the multi-scale feature fusion module for arrhythmia in 12-lead varied-length ECG. Our model utilizes the complementary power between different structures, which include group convolution block, Spatial Pyramid Pooling (SPP) layer, and multi-scale feature fusion module. The proposed method can extract features from every lead separately for 12-lead varied-length records, and effectively achieve multi-scale feature extraction and cross-scale information complementarity of ECG by integrating multiple convolution kernels with different receptive fields. The experimental results show that our model achieved an overall classification F1 score of 83.3%. Combining all these excellent features, this model offers a solution for varied-length signal processing problems.

The detection of consecutive cardiac cycles plays an important role in the daily monitoring of cardiovascular diseases. The Seismocardiogram (SCG) signal measures the cardiac-induced vibrations and is suitable for continuously tracking since it can be measured in a non-invasive fashion. In this paper, a beat-to-beat heart rate detection method based on low-frequency SCG signal using BiLSTM network is proposed. The regression model to predict Electrocardiogram (ECG) from SCG was trained on CEBS dataset and a 5-fold cross-validation method was adopted. The RMSE of the predicted ECG signal and ECG signal was 0.0367. The sensitivity and precision achieved 0.97 and 0.98 separately and the Spearman Correlation was 0.98. This work shows promise for the proposed SCG based methodology to be used as an alternative to ECG for continuous beat-to-beat heart rate monitoring.

Electrocardiogram (ECG) signal is a common diagnostic basis for heart disease. Reliable waveform detection method for ECG has important applications in clinical diagnosis. This work proposes a new method for P- and T- wave detection based on auxiliary waveform. The algorithm includes three parts: preprocessing, construction of auxiliary waveforms and P- and T-wave position detection. First, the ECG signal is preprocessed to remove baseline drift and high-frequency noise. Then, the auxiliary waveforms is constructed in the P- and T-wave searching windows which are defined based on the QRS wave position. Next, on the basis of the prominence of the auxiliary waveform, each lead of the 12-lead ECG signal is weighted and the final auxiliary waveform is constructed. Finally the P/T-wave positions are detected according to the auxiliary waveform memory. The performance of the proposed algorithm has been evaluated on a 12-lead manually annotated database. Standard deviations of 26.55 and 30.92 ms were obtained for the onset and offset of T-wave respectively, the standard deviation of the onset and offset of P-wave is 13.25 and 15.7 ms. The algorithm we proposed can provide reference for following research.

Atrial fibrillation and atrial flutter are two more common rapid atrial arrhythmias, they are independent of each other and can transform each other. Because the clinical characteristics of the two diseases are similar, it is easy to cause misdiagnosis. Therefore, effective diagnosis of distinguishing atrial fibrillation and atrial flutter can reduce the damage suffered by patients during treatment. In this paper, we propose a kind of based on residual network and multi-scale feature fusion structure of neural network to automatically classify atrial fibrillation, atrial flutter and other ECG signals. The ECG signals used are derived from clinical data labeled by expert diagnosis. In addition, we used several popular classification methods proposed in the past for comparative experiments to verify the classification performance of our models. The experimental results show that the total F1, recall and precision of the proposed algorithm are 0.923, 0.916 and 0.93. The algorithm we propose has the potential to provide reference for subsequent research.

Electrocardiogram (ECG), as an important method for diagnosing cardiovascular diseases, can record the heart activity over a period of time. However, most of the current studies on ECG classification focus on the single scale information and ignore the complementary information between different scales. Therefore, this paper proposed an end-to-end multi-scale fusion convolutional neural network (CNN) for heartbeat classification. In this method, multiple convolution kernels of different reception domains are used to extract unique features of different scales, and the extracted multiple scale features are fused, which could effectively capture disease patterns and suppress noise interference. At the same time, attention module is used to select features to improve model performance. Improve efficiency with residual module. Finally, we obtained 34, 983 heartbeats from the Physikalisch-Technische Bundesanstalt (PTB) dataset to validate the model performance. The overall Fl-score is 99. 69%, and the Fl-score of each single class is more than 99. 35%, which is better than the existing algorithms. It can be described as a reference for future research.

Wearable electrocardiogram (ECG) monitoring systems were effective ways to diagnose intermittent cardio diseases. However, the Electrode Motion Artifact (EMA) and the Muscle Artifact (MA) destroy the incipient shape of ECG signals, decrease the accuracy of diagnostic results. Here, we proposed a novel method for detecting destroyed segments of ECG signals. The method was based on a deep learning network, and its performance was evaluated on a synthetic dataset of the MIT-BIH arrhythmia database. Its practicability was tested with three R-peak detection algorithms. By removing the destroyed segments in ECG signals, the sensitive and positive prediction of these R-peak detection algorithms were promoted significantly.

In this work, arrhythmia appearing with the presence of abnormal heart electrical activity is efficiently recognized and classified. A novel method is proposed for accurate recognition and classification of cardiac arrhythmias. Firstly, P-QRS-T waves is segmented from ECG waveform; secondly, morphological features are extracted from P-QRS-T waves, and ECG segment features are extracted from the selected ECG segment by using PCA and dynamic time warping(DTW); finally, SVM is applied to the features and automatic diagnosis results is presented. ECG data set used is derived from the MIT-BIH in which ECG signals are divided into the four classes: normal beats(N), supraventricular ectopic beats (SVEBs), ventricular ectopic beats (VEBs) and fusion of ventricular and normal (F). Our proposed method can distinguish N, SVEBs, VEBs and F with an accuracy of 97.80 percent. The sensitivities for the classes N, SVEBs, VEBs and F are 99.27, 87.47, 94.71, and 73.88 percent and the positive predictivities are 98.48, 95.25, 95.22 and 86.09 percent respectively. The detection sensitivity of SVEBs and VEBs has a better performance by combining proposed features than by using the ECG morphology or ECG segment features separately. The proposed method is compared with four selected peer algorithms and delivers solid results.