Multivariate time series prediction is a critical problem that is encountered in many fields, and recurrent neural network (RNN)‐based approaches have been widely used to address this problem. However, traditional RNN‐based approaches for predicting multivariate time series are still facing challenges, as time series are often related to each other and historical observations in real‐world applications. To address this limitation, this paper proposes a spatiotemporal self‐attention mechanism‐based LSTNet, which is a multivariate time series forecasting model. The proposed model leverages two self‐attention strategies, spatial and temporal self‐attention, to focus on the most relevant information among time series. The spatial self‐attention is used to discover the dependences between variables, while temporal attention is employed to capture the relationship among historical observations. Moreover, a standard deviation term is added to the objective function to track multivariate time series effectively. To evaluate the proposed method’s performance, extensive experiments are conducted on multiple benchmarked datasets. The experimental results show that the proposed method outperforms several baseline methods significantly. Therefore, the proposed spatiotemporal self‐attention‐based LSTNet is a promising approach for predicting multivariate time series.
In the field of fault diagnosis for factory machinery systems, the development of deep learning methods has been hindered by the challenge of acquiring fault data, highlighting the need to extract noise robust features from limited labeled data. In this paper, a light and efficient complex-domain acoustic feature extraction method, complex-domain principal frequency convolution network (CPFCN) is proposed for fault diagnosis in rotating machinery, which consists of a principal frequency filter (PFF) and stacked convolution network (SCN). The PFF filters out non-principal frequency noise to focus on the predominant frequency. The SCN is designed to effectively extract the amplitude and phase features, which can fully leverage the complex-domain information within the acoustic data. The experimental results show that the proposed CPFCN have 33% increasing in accuracy while 87% reduction in training time and 41% reduction in feature extraction time. Additionally, the proposed framework has improved the accuracy by 59% on the dataset with noise compared to the best-performing method in the experimental study, achieving stronger noise robustness in the case of limited samples.
4 November 2024 in IEEE Transactions on Knowledge and Data Engineering
Short-term time series forecasting is pivotal in various scientific and industrial fields. Recent advancements in deep learning-based technologies have significantly improved the efficiency and accuracy of short-term time series modeling. Despite advancements, current time short-term series forecasting methods typically emphasize modeling dependencies across time stamps but frequently overlook inter-variable dependencies, which is crucial for multivariate forecasting. We propose a multi patterns memory model discovering various dependency patterns for short-term multivariate time series forecasting to fill the gap. The proposed model is structured around two key components: the short-term memory block and the long-term memory block. These networks are distinctively characterized by their use of asymmetric convolution, each tailored to process the various spatial-temporal dependencies among data. Experimental results show that the proposed model demonstrates competitive performance over the other time series forecasting methods across five benchmark datasets, likely thanks to the asymmetric structure, which can effectively extract the underlying various spatial-temporal dependencies among data.
Health state assessment is critical for mechanical equipment’s smooth and healthy operation. This paper proposes a novel approach for health state assessment based on acoustic signals during the process of machinery running. It consists of multi-domain feature (MF) extraction and comprehensive health indicator (CHI) construction. MF is extracted from various acoustic features, including time and frequency (TF) features, mel-frequency cepstral coefficients, and gammatone frequency cepstral coefficients. The stacked long short-term memory (LSTM) is used to extract the high-level features of the MF, which are then input to the downstream PCA to obtain the LSTM-PCA health indicator (LP-HI). Parallelly, the MF is fed into the self-organizing mapping (SOM) model to calculate the minimum quantization error (MQE) as SOM-MQE health indicator (SM-HI). These two indicators are fused using weighted fusion and nonlinear mapping to calculate CHI. The experimental results on air compressor dataset show a 25.8% reduction in evaluation error compared with SOTA results in this paper. The proposed nonlinear mapping function furthermore reduces fitting error on HI by 38.9%. These demonstrate the effectiveness and superiority of the proposed method in machinery health state assessment.