Electrocardiogram (ECG) signals frequently encounter diverse types of noise, such as baseline wander (BW), electrode motion (EM) artifacts, muscle artifact (MA), and others. These noises often occur in combination during the actual data acquisition process, resulting in erroneous or perplexing interpretations for cardiologists. To suppress random mixed noise (RMN) in ECG with less distortion, we propose a Transformer-based Convolutional Denoising AutoEncoder model (TCDAE) in this study. The encoder of TCDAE is composed of three stacked gated convolutional layers and a Transformer encoder block with a point-wise multi-head self-attention module. To obtain minimal distortion in both time and frequency domains, we also propose a frequency weighted Huber loss function in training phase to better approximate the original signals. The TCDAE model is trained and tested on the QT Database (QTDB) and MIT-BIH Noise Stress Test Database (NSTDB), with the training data and testing data coming from different records. All the metrics perform the most robust in overall noise and separate noise intervals for RMN removal compared with the baseline methods. We also conduct generalization tests on the Icentia11k database where the TCDAE outperforms the state-of-the-art models, with a 55% reduction of the false positives in R peak detection after denoising. The TCDAE model approximates the short-term and long-term characteristics of ECG signals and has higher stability even under extreme RMN corruption. The memory consumption and inference speed of TCDAE are also feasible for its deployment in clinical applications.

In recent years, there has been a surge in applying deep learning methods for signal quality assessment (SQA) and denoising of ECG signals. However, solely focusing on denoising can lead to situations where models trained on severely contaminated signals perform poorly on higher-quality signals. Likewise, carrying out SQA in isolation may be constrained by the number of training data and susceptible to label noise. In this study, we leverage the correlation between denoising and SQA and propose a multi-task learning network (MTL-NET) that addresses both tasks in parallel to mitigate the challenges. MTL-NET employs a stacked bidirectional LSTM network, combined with an attention module, as a shared backbone to extract common features for both SQA and denoising tasks. Weighted loss functions are utilized to balance the performance of these two tasks. Within MTL-NET, we classify ECG signal quality into three categories, and compare MTL-NET’s performance with baseline methods dedicated solely to denoising or three-class SQA. The results demonstrate that MTL-NET outperforms other denoising methods across all metrics obviously, including the corresponding single-task denoising model. Specifically, MTL-NET significantly enhances the denoising performance of the "intermediate" quality category and the SQA performance in identifying ‘unacceptable’ signal segments. It provides a novel perspective on ECG signal preprocessing by implementing a balance rule through loss functions for both tasks within a model. With 0.28M parameters and 27.35M FLOPs, the MTL-NET is not only efficient but also exhibits excellent generalization performance on an independent dataset, making it suitable for deployment in resource-constrained environments.

This study addresses the challenges of predicting ship heave motion in real time, which is essential for mitigating sensor–actuator delays in high-performance active compensation control. Traditional methods often fall short due to training on specific sea conditions, and they lack real-time prediction capabilities. To overcome these limitations, this study introduces a multi-step prediction model based on a Seq2Seq framework, training with heave data taken from various sea conditions. The model features a long-term encoder with attention-enhanced Bi-LSTM, a short-term encoder with Gated CNN, and a decoder composed of multiple fully connected layers. The long-term encoder and short-term encoder are designed to maximize the extraction of global characteristics and multi-scale short-term features of heave data, respectively. An optimized Huber loss function is used to improve the fitting performance in peak and valley regions. The experimental results demonstrate that this model outperforms baseline methods across all metrics, providing precise predictions for high-sampling-rate real-time applications. Trained on simulated sea conditions and fine-tuned through transfer learning on actual ship data, the proposed model shows strong generalization with prediction errors smaller than 0.02 m. Based on both results from the regular test and the generalization test, the model’s predictive performance is shown to meet the necessary criteria for active heave compensation control.

Motion interference is a major issue in the process of acquiring electrocardiogram (ECG) signals. It introduces noise and artifacts into the ECG signal, causing signal distortion and deformation, thus affecting the accuracy of subsequent analyses and diagnoses. This study introduces a deep learning-based model for motion artifact removal in ECG signals, namely MSCT, which combines Multi-Scale Convolution with Transformer encoder to extract the local and global features of ECG signals and motion artifacts. The comparative experiments are conducted on the MIT-BIH arrythmia database which is contaminated with the motion artifact from MIT-BIH noise stress test database. Our MSCT model shows significantly higher Cosine Similarity and output signal-to-noise ratio (SNR), and lower Root Mean Square Error (RMSE) and Multi-Scale Entropy Based Weighted Percentage Root-mean-square Difference (MSEWPRD), compared with that of the three previous methods on the motion artifact-corrupted data with various input SNR (-6 dB, 0 dB and 6 dB). Especially, when the input SNR is -6 dB, the proposed MSCT achieves the results 6.15±2.83%, 0.25±0.14 mV, 0.91±0.08, 9.20±3.74 dB for MSEWPRD, RMSE, Cosine Similarity and SNR, respectively. All the four metrics, including quantitative and qualitative metrics, have demonstrated the superiority and robust of the MSCT model proposed in this study. Furthermore, the time-domain and frequency-domain plots of a randomly selected denoising segment at -6 dB effectively showcase the superior performance of our denoising method.

Multivariate time series (MTS) classification is essential in industries, such as healthcare and manufacturing, where it helps extract key features from complex data for decision-making and predictions. However, current MTS classification methods often struggle with high-dimensional data and limited labeled samples, leading to poor performance. To address this, we propose a semisupervised MTS classification method via heterogeneous graph representation. This method provides a new perspective on modeling MTS, which integrates various additional information and captures their relationships. First, we employ a contrast temporal self-attention module to obtain sparse MTS representations. Then, we use soft dynamic time warping to model similarities and construct a similarity graph. Next, we learn shapelets and incorporate both subject features and shapelets into the graph, refining it into a heterogeneous graph. Finally, we apply a dual-level graph attention network for node classification task. Experiments on 19 MTS datasets show that our method outperforms existing state-of-the-art methods, demonstrating its effectiveness for MTS classification task.