Time Series Information Research Articles

A time series and deep fusion framework for rotating machinery fault diagnosis

For shafting rotation equipment, in the fault diagnosis based on vibration analysis, the sampling signal is a waveform of a short moment, which is quasi-static information at the corresponding moment. A single quasi-static piece of information does not necessarily contain obvious fault information, however, in a longer period much larger than the sampling period, the time series information composed of multiple quasi-static combinations may well-characterize the fault. Traditional deep learning algorithms often focus on fault features in quasi-static information, ignoring time series features, which results in low diagnostic accuracy. To solve the above problems, this paper firstly presents a fault diagnosis framework of time series and deep fusion network (TDFN). Then, a method to extract time series information and quasi-static information based on expert experience knowledge is proposed, and finally studies the method of using this framework to train the network and fuse the two types of features. Different from existing deep learning methods, TDFN includes time series blocks, depth blocks and fusion blocks. This paper uses laboratory data and industrial rotor data for verification, the results show that the overall accuracy of TDFN reached 93.8% and it does not experience a drop in accuracy after fusion. The diagnosis accuracy is higher than traditional networks, providing guiding suggestions for the management and operation of on-site industrial equipment.

A deep learning-based approach for predicting oil production: A case study in the United States

The accuracy of oil production predictions is crucial in the field of petroleum engineering. However, due to the time series characteristics of oil production and the complex relationship among multiple influencing factors, traditional methods, and time series prediction techniques have limitations in fully extracting time series features and exploring the internal relationships between variables. Deep learning techniques possess unique advantages in solving nonlinear problems and time-series problems but require a significant amount of data and can only conduct short-term oil production predictions due to the nonlinear and chaotic nature of yield. Therefore, this work aims to establish a model that can overcome these limitations. A modified GRU (M-GRU) is proposed to extract time series features of variables and external economic information. Then, the autoregressive (AR) model is integrated into importing priori knowledge to enhance the model's capability in extracting time series information. In addition, feature mapping is used to improve convergence performance. Finally, A loss function that can dynamically adjust the weights is proposed, which can enhance the model's ability to fit from error-prone samples. By validating the proposed model using oil production data, it has been proven that the model can predict oil production accurately and outperform other models with a correlation coefficient reaching 0.99874. Further tests show that the model can provide accurate prediction results even with limited sample sizes and during coincidental events such as financial crises and COVID-19 epidemics, providing strong support for decision-making for reservoir engineers.

Research on Multistate Variable Prediction of a Master‐Slave Electric‐Hydraulic Hybrid Vehicle Based on CNN‐LSTM Model with Attention Mechanism

This paper utilizes a convolutional neural network and long short‐term memory with attention mechanism (CLA) prediction model for time series forecasting. An innovative master‐slave electric‐hydraulic hybrid vehicle (MSEHV) is introduced. It simplifies the drivetrain structure by utilizing planetary gears for energy transfer and coupling to transform electrical, mechanical, and hydraulic energy. This paper conducts simulation verification using existing control strategies. It combines the velocity time series information with the hydraulic accumulator pressure time series information to create a unified time series. Some of the more popular machine learning and neural network algorithms are considered, and the mean square error, mean absolute error, and coefficient of determination evaluation metrics are calculated for validation. The research presents a developed battery state of charge regression prediction model based on genetic algorithm‐backpropagation by combining the predicted results of composite time series with the current discharge depth of the battery. The validation results indicate that the CLA time series prediction model proposed in this paper for MSEHV has higher prediction accuracy compared to other prediction models.

Point and interval forecasting of ultra-short-term carbon price in China

Accurate carbon price prediction is a reference that allows market participants to make decisions. This study adopts a total of 1,857 trading days of data from April 2, 2014, to June 15, 2022, in the Hubei carbon market, one of the first and largest pilot carbon markets in China for carbon price prediction. We propose a new framework based on the GA-VMD-CNN-BiLSTM-Attention hybrid model: a genetic algorithm (GA) is adopted to search the optimal parameter combination of variational mode decomposition (VMD); a convolutional neural network (CNN) is established to discover the relationship between influencing factors and carbon prices; a bidirectional long and short-term memory network (BiLSTM) is applied to extract time series information; and an attention mechanism is used to strengthen the influence of important information on carbon prices. Compared to 11 other models, the GA-VMD-CNN-BiLSTM-Attention model has a higher accuracy and stronger model reliability. In addition to deterministic point prediction, this study uses non-parametric kernel density estimation with the Gaussian kernel function (KDE-Gaussian) for interval forecasting. The forecasting can quantify the uncertainty of carbon prices and serve as a more practical reference for decision-makers. By revealing the particularly challenging issue that underlies carbon price forecasting, our analysis also sheds light on current low-carbon policies in China.

A novel study for depression detecting using audio signals based on graph neural network

Depression is a prevalent mental health disorder. The absence of specific biomarkers makes clinical diagnosis highly subjective. This makes it difficult to make a definitive diagnosis for the patient. Recently, deep learning methods have shown promise for depression detection. However, current methods tend to focus solely on the connections within or between audio signals, leading to limitations in the model’s ability to recognize depression-related cues in audio signals and affecting its classification performance. To address these limitations, we propose a graph neural network approach for depression recognition that incorporates potential connections within and between audio signals. Specifically, we first use a gated recurrent unit (GRU) to extract time-series information between frame-level features of audio signals. We then construct two graph neural network modules sequentially to explore the potential connections within and between audio signals. The first graph network module constructs a graph using the frame-level features of each audio sample as nodes. The output is obtained as a graph-embedded feature vector representation after the graph convolution layers. Subsequently, the output graph embedding feature vector representation of the first graph network model is used as the nodes of the graph to construct the second graph network. The internal relationship between audio signals is encoded by the property of node neighborhood information propagation. In addition, we use a pre-trained emotion recognition network to extract emotional features that are highly correlated with depression. By further strengthening the connection weights among nodes in the second graph network through a self-attention mechanism, relevant cues are provided for the model to complete depression detection from audio signals. We conducted extensive experiments on three depression datasets, including DAIC-WOZ, MODMA, and D-Vlog. The proposed model achieves better results on several performance evaluation metrics such as accuracy, F1-score, precision, and recall compared to all the compared algorithms, validating its effectiveness.

Rolling bearing fault diagnosis method based on multi-information fusion characteristics under complex working conditions

To address the problem of low accuracy of traditional neural network models for rolling bearing fault diagnosis under complex working conditions, a novel method based on the multi-information fusion axis attention mechanism (MFA) and improved multi-scale convolutional neural network (MFA-IMSCNN) is proposed. Firstly, a multi-information fusion axis attention mechanism is presented to accurately extract the feature channel information and the number axis location information at various scales. Secondly, a multi-scale dense cyclic feature extraction module (MSDC) is built to improve the feature extraction ability of the MFA-IMSCNN under small sample conditions. The module better captures the potential feature information of time series and ensures the integrity of the information. Finally, two different datasets are imported into the MFA-IMSCNN model to evaluate its effectiveness, and the experimental results verify that the MFA-IMSCNN obtains better classification performance and stronger robustness under various experimental conditions.

Deep temporal networks for EEG-based motor imagery recognition.

The electroencephalogram (EEG) based motor imagery (MI) signal classification, also known as motion recognition, is a highly popular area of research due to its applications in robotics, gaming, and medical fields. However, the problem is ill-posed as these signals are non-stationary and noisy. Recently, a lot of efforts have been made to improve MI signal classification using a combination of signal decomposition and machine learning techniques but they fail to perform adequately on large multi-class datasets. Previously, researchers have implemented long short-term memory (LSTM), which is capable of learning the time-series information, on the MI-EEG dataset for motion recognition. However, it can not model very long-term dependencies present in the motion recognition data. With the advent of transformer networks in natural language processing (NLP), the long-term dependency issue has been widely addressed. Motivated by the success of transformer algorithms, in this article, we propose a transformer-based deep learning neural network architecture that performs motion recognition on the raw BCI competition III IVa and IV 2a datasets. The validation results show that the proposed method achieves superior performance than the existing state-of-the-art methods. The proposed method produces classification accuracy of 99.7% and 84% on the binary class and the multi-class datasets, respectively. Further, the performance of the proposed transformer-based model is also compared with LSTM.

Remaining electrical life prediction of AC contactor based on CAE-BiGRU-Attention

To tackle the challenges of low prediction accuracy caused by single-feature modeling, and the hidden state of the neural network easily loses some information of the long time series, a method for predicting the remaining electrical life of AC contactor using a convolutional autoencoder-bidirectional gated recurrent unit-attention (CAE-BiGRU-Attention) was proposed in this work. Firstly, the feature parameters were extracted from the AC contactor full-life test, and an optimal feature subset was selected using neighborhood component analysis and Spearman rank correlation coefficient to characterize the degradation state of electrical life effectively. Then, the deep information of the optimal feature subset was extracted using CAE. Finally, the remaining electrical life of the AC contactor was treated as a long time series problem and predicted in time series by BiGRU-Attention accurately. The case analysis demonstrates that the model has better prediction accuracy than recurrent neural network (RNN), long short-term memory (LSTM), GRU, BiGRU and CAE-BiGRU models, with an average effective accuracy of 97.12%. This effectively demonstrates the model’s feasibility to accurately predict temporal sequences in the remaining electrical life prediction of electrical equipment.

Esophageal cancer detection framework based on time series information from smear images

The gold standard for esophageal cancer diagnosis and treatment is the Thinprep Cytologic Test (TCT) of suspected sections. TCT refers to analyze the features of the lesion areas using tissue regions stained with hematoxylin and eosin. However, pathologists are currently subjected to high workloads and limited experiences. Moreover, automatic esophageal cancer detection methods typically screen the smears by focusing on suspicious lesions identified by cytotechnicians, cell misclassification often results in inaccurate and weak robust diagnoses, smear utilization rate is also reduced. Computer-assisted diagnosis also faces challenges such as information loss, non-quantitative analysis, and nonstandard diagnosing processes. To address these issues, this article proposes a deep learning based detection framework called the Esophageal Time Series Thinprep Cytologic Test (ETS-TCT) to provide a quantitative and robust screening of Whole Slide Imaging (WSI) for esophageal cancer. Our system is divided into three modules: Coarse–fine detection model, quantitative analysis model, and Fusion Long Short Term Memory-Support Vector Machine (FLSTM-SVM) model. These modules are organized to map the pathological information from cell-level to smear-level and offer interpretable predictions to pathologists. For each smear, we design the coarse–fine segmentation model for adaptively locating and classifying target cells, give mathematical descriptions for cell Deoxyribonucleic Acid (DNA) values using the quantitative analysis model, establish an early risk screening model for esophageal cancer using the optimized FLSTM-SVM network. To validate the methodology, we use two clinical datasets containing 580 samples, achieving the accuracy, sensitivity, specificity, and area under ROC (AUC) of 92.0%, 93.4%, 94.4%, and 0.913, respectively. In short, the framework provides robust and accurate diagnosis of esophageal cancer WSIs using the time series features. This screening can assist the pathologists in strengthening the diagnosis by robustly presenting the quantitative intermediate results as well as visual interpretations. Moreover, we clearly present and expatiate the concepts of time series features, highlighting the relationship between these features to physiological states of patients. Our results indicate the possibility of using the framework as a second system to the computer-assisted esophageal cancer screening.

Incorporating stock prices and text for stock movement prediction based on information fusion

Forecasting stock market via historical financial data is an important issue for market participants because even if the prediction accuracy is only slightly improved, better trading decisions can be made. Historical financial data has evolved from the initial single text or stock price to the fusion of multisource information. However, how to adopt a method that adaptively fuses numerical data and text so that the prediction model can learn time series information in parallel remains a challenging problem. In this paper, we propose a collaborative attention Transformer fusion model for stock movement prediction (CoATSMP), including parallel extraction of text and prices features, parameter-level fusion and a joint feature processing module, that can successfully deeply fuse text and stock prices in view of the soft fusion method. The experiments show that (1) the proposed approach outperforms the baselines, (2) the soft fusion method proposed in this paper has better modeling performance under the CoATSMP framework, which brings greater improvement in the prediction performance, (3) models containing prices and text are better than those using only one data source, and (4) quantitative analysis of experimental results indicates that text plays a relatively more critical role in the CoATSMP framework. Real simulation trading shows that the trading strategy based on CoATSMP can significantly improve profits; thus, the model has practical application value.

Dynamic community detection based on graph convolutional networks and contrastive learning

With the continuous development of technology and networks, real-life interactions are gradually being abstracted into social networks for study. Social circles are a fundamental structural feature that is prevalent on social networks. Thus, exploring social circle structure plays an important role in revealing the characteristics of complex social networks and provides guidance for understanding social behavior in real life. For example, it can aid in precision marketing, personalized recommendation, and knowledge dissemination within social circles. One of the important means of identifying social network circles lies on community detection algorithms. However, real-world social networks are often dynamic and can be studied and analyzed by building dynamic networks, while existing dynamic network community detection methods tend to ignore the global structure information and time-series information of nodes. To address this problem, this paper proposes a dynamic network community detection algorithm based on graph convolutional neural networks and contrastive learning, which fully captures the adjacent characteristics between nodes based on the correlation information and leverages the feature smoothing strategy to efficiently extract node representations of dynamic networks under unsupervised scenario. Specifically, the proposed algorithm first utilizes node correlation based aggregation strategy to compute the feature matrix for single time-step of the dynamic network. Then, mutual information maximization is implemented based on cross-entropy between learned local and global representations. To reduce the computational overhead in the optimization process, an additional LSTM module is further equipped for updating the parameters of graph convolutional networks in each time-step. Additionally, a contrastive learning based network smoothing strategy is designed to minimize the feature differences between neighboring nodes. Comparative experiments demonstrate that the proposed algorithm achieves excellent performance on both synthetic and real networks.

OODT: LOS Signal Identification for Acoustic Indoor Localization From Stream Perspective

Acoustic-based technologies have attracted more attention in indoor ranging-based localization and tracking applications. To achieve high-accuracy ranging measurements, the precise identification of line-of-sight(LOS) acoustic signals is essential. This paper proposes a novel Off-Online Dynamic Training method(OODT) to identify LOS acoustic signals from stream perspective based on a few training data in dynamic indoor environment. The dynamic online training method is proposed to identify the unlabeled acoustic feature stream by the Parent-Child model, utilizing the dynamic prior probability from time series information and a selection strategy based on the prediction risk. Then, the trustworthy pseudo-labeled streaming samples are accumulated into the Child-Models by online learning in real-time. To reduce the impact of discarding the untrustworthy LOS signals, the Off-Online retraining method is proposed to incorporate the spatial information into the Parent-Models, which uses the category distribution of all pseudo-labeled streaming samples as the prior probability for iterative retraining. Subsequently, a new round of the dynamic online training is conducted to update the pseudo-labels of feature stream. Experiments demonstrate that the dynamic online training method has a higher identification precision of LOS acoustic signals from stream perspective, reaching 98% and more than 93.06% in above-ground and underground experimental scenarios. Moreover, the adaption for the concept drift is also verified with identification precision of 97.98% and 98.23%, respectively. The Off-Online retraining method further optimize the performance of the Parent-Child models in a single scenario. In conclusion, the proposed OODT method can autonomously identify dynamic acoustic signals with strong robustness and scenario drift adaptability.

Research on Automatic Classification of Coal Mine Microseismic Events Based on Data Enhancement and FCN-LSTM Network

Efficient and accurate classification of the microseismic data obtained in coal mine production is of great significance for the guidance of coal mine production safety, disaster prevention and early warning. In the early stage, the classification of microseismic events relies on human experiences, which is not only inefficient but also often causes some misclassifications. In recent years, the neural network-based classification method has become more favored by people because of its advantages in modeling procedures. A microseismic signal is a kind of time-series signal and the application of the classification method is widely optimistic. The number and the balance of the training data samples have an important impact on the accuracy of the classification result. However, the quality of the training data set obtained from the production cannot be guaranteed. A long short-term memory (LSTM) network can analyze the time-series input data, where the image classification at the pixel level can be achieved by the fully convolutional network (FCN). The two structures in the network can not only use the advantages of the FCN for extracting signal details but also use the characteristics of LSTM for conveying and expressing the long time-series information effectively. In this paper, a time-series data enhancement combination process is proposed for the actual poor microseismic data. A hybrid FCN-LSTM network structure was built, the optimal network parameters were obtained by experiments, and finally a reasonable microseismic data classifier was obtained.

Geospatial dataset for hydrologic analyses in India (GHI): a quality-controlled dataset on river gauges, catchment boundaries and hydrometeorological time series

Abstract. Streamflow gauging stations not only track the pulse of rivers but also act as common reference points for hydrologic and other environmental analyses. As such, streamflow data and metadata on gauging stations – Geographic Information System (GIS) data on station locations, their upstream catchment boundaries and river flow networks – are critical for analyses. However, for India's river basins, the availability of such data is limited; when available, data are not in an analysis-ready format and can have substantial errors. Studies often use available information from India's water agencies as is, without checking its validity. This study addresses the above limitations by building a new dataset using existing metadata (from the Central Water Commission, CWC, and the Water Resources Information System, WRIS) and checking it against publicly available information from global data sources (e.g., World Wildlife Fund, Multi-Error-Removed Improved-Terrain Hydro and Copernicus) and online maps (e.g., Google Maps). The quality control process categorizes existing metadata based on their consistency with these sources; also, existing metadata are supplemented with additional information where needed. The new dataset developed here is called the “Geospatial dataset for Hydrologic analyses in India” (GHI) and uses Hydrological data and maps based on Shuttle Elevation Derivatives at multiple Scales (HydroSHEDS) data as the underlying template. GHI has both geospatial and time series information. In this initial version of GHI, the spatial domain includes only the river basins of Peninsular India where daily streamflow data are publicly available. Following the quality control process, the CWC's 645 stations in Peninsular India were categorized into three groups: Group 1 (reliable metadata and adequate daily streamflow data; 213 stations), Group 2 (reliable metadata and inadequate or no daily streamflow data; 259 stations) and Group 3 (missing or unreliable metadata; 173 stations). For each of the 472 stations falling into groups 1 and 2, catchment-specific annual and monthly time series spanning 71 water years (1950–2020) of the following were compiled: observed precipitation from the Indian Meteorological Department (IMD); observed streamflow from WRIS; estimated precipitation, evapotranspiration (ET) and streamflow from ERA5-Land; and ET from the Global Land Evaporation Amsterdam Model (GLEAM). A preliminary analysis of catchment-scale time series of data indicates that, while the compiled data appear reasonable over most of the study domain, spurious runoff–precipitation ratios were observed in the hilly coastal regions of Western India. This adds to yet another data-related obstacle faced by the hydrologic community. In order to quantify historical changes and reconcile them with anticipated future changes, the community needs robust and reliable hydrographic and hydrometeorological datasets as well as unrestricted access to such datasets. The goal of this study is to highlight the limitations of existing datasets and pave the way for a community-led effort towards building the needed datasets. GHI serves as a placeholder until such datasets become available. Potential improvements to GHI are discussed. GHI is publicly available at https://doi.org/10.5281/zenodo.7563599 (Goteti, 2023).

Optimizing Spaced Repetition Schedule by Capturing the Dynamics of Memory

Spaced repetition, namely, learners review items in a given schedule, has been proven powerful for memorization and practice of skills. Most current spaced repetition methods focus on either predicting student recall or designing an optimal review schedule, thus omitting the integrity of the spaced repetition system. In this work, we propose a novel spaced repetition schedule framework by capturing the dynamics of memory, which alternates memory prediction and schedule optimization to improve the efficiency of learners' reviews. Firstly, the framework collects logs from students' reviews and builds memory models with Markov property to capture the dynamics of memory. Then, the spaced repetition optimization is transformed into a stochastic shortest problem and solved via the value iteration method. We also construct a new benchmark dataset for spaced repetition, which is the first to contain time-series information during learners' memorization. Experimental results on the collected data from the real world and the simulated environment demonstrate that the proposed approach reduces 64% error and 17% cost in predicting recall rates and optimizing schedules compared to several baselines. We have publicly released the dataset containing 220 million rows and codes used in this paper at: <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/maimemo/SSP-MMC-Plus</uri> .

Multivariate Time Series Prediction Based on Temporal Change Information Learning Method.

In the multivariate time series prediction tasks, the impact information of all nonpredictive time series on the predictive target series is difficult to be extracted at different time stages. Through the emphasis on optimal-related sequences in the target series, the deep learning model with the attention mechanism achieves a good predictive performance. However, temporal change information in the objective function and optimization algorithm is completely ignored in these models. To this end, a temporal change information learning (CIL) method is proposed in this article. First, mean absolute error (MAE) and mean squared error (MSE) losses are contained in the objective function to evaluate different amplitude errors. Meanwhile, the second-order difference technology is used in the correlation terms of the objective function to adaptively capture the impact of the abrupt and slow change information in each series on the target series. Second, the long short-term memory (LSTM) network with the transformation mechanism is used in the method so that temporal dependence information can be fully extracted (i.e., avoiding the supersaturation region). Third, to effectively obtain the optimal model parameters, the current and historical moment estimation information is adaptively memorized without the introduction of additional hyperparameters, and therefore, the acquisition ability of temporal change information in the error gradient flow is greatly enhanced by the proposed optimization algorithm. Finally, three datasets with different scales are used to verify the advantages of the CIL method in computational overhead and prediction effect.

A method for the spatiotemporal correlation prediction of the quality of multiple operational processes based on S-GGRU

Predicting quality indicators is essential in process production, process production often involves multiple processes. However, modeling the prediction model for quality indicators faces challenges due to frequent fluctuations in raw material composition and the control system's automatic adjustment of process parameters. Additionally, due to the mutual influence between processes, the problem of predicting quality indicators in process production is a systematic prediction problem. In order to deal with the challenges, firstly, the physical control relationships are transformed into graph data, which is input into GNN to calculate the spatial characteristics of process parameters on quality indicators under physical information constraints. GRU captures the temporal features of quality indicator data. Based on above study, a combination network GGRU integrating quality indicators' spatial and temporal characteristics was proposed. Finally, each GGRU is then connected based on the material flow between each process, and path information between processes is converted to the calculation flow between GGRU models to form the S-GGRU prediction model. Experimental results show that the proposed S-GGRU model, a multi-process prediction model based on GRU and GNN, outperforms the recurrent neural network model and the graphical neural network model in solving the quality prediction problem for multi-process and multi-metrics in-process production. Therefore, both the time-series information of quality metrics and the influencing factors among variables should be considered to achieve better predictions, and the S-GGRU performance better in predicting quality indicators for a 5-process process line.

Convolutional gated recurrent unit-driven multidimensional dynamic graph neural network for subject-independent emotion recognition

Electroencephalogram (EEG) could directly reflect human brain activities. Recently, EEG-based emotion recognition technology has attracted widespread attention. However, although most of the existing methods can extract temporal and spatial features of EEG signals, the extraction process of the time-series information and spatial features is separate from each other. Some spatial features may be lost when extracting time-series information, and vice versa. Additionally, the non-Euclidean representations between the EEG electrodes cannot be ignored. In response to these issues, we construct a convolutional gated recurrent unit-driven multidimensional dynamic graph neural network (CGRU-MDGN), which mines its local spatial characteristics of EEG signals while learning the temporal information and captures non-Euclidean spatial features between EEG channels to classify emotions more accurately. Specifically, we substitute the fully connected layers of the GRU with convolution layers, which can take into account both the local spatial features and time-series information of EEG signals. Next, multiple non-Euclidean spatial representations are generated using multidimensional dynamic graph convolution. Furthermore, the characteristics of different dimensions or views are merged through the adaptive-weighted summation technique to realize the maximized feature extraction as much as possible. Comprehensive experiments reveal that the constructed CGRU-MDGN could achieve competitive recognition results on subject-independent emotion recognition tasks.

Statistically validated coeherence and intensity in temporal networks of information flows

We propose a method for characterizing the local structure of weighted multivariate time series networks. We draw intensity and coherence of network motifs, i.e. statistically recurrent subgraphs, to characterize the system behavior via higher-order structures derived upon effective transfer entropy networks. The latter consists of a model-free methodology enabling to correct for small sample biases affecting Shannon transfer entropy, other than conducting inference on the estimated directional time series information flows. We demonstrate the usefulness of our proposed method with an application to a set of global commodity prices. Our main result shows that, despite simple triadic structures are the most intense, coherent and statistically recurrent over time, their intensity suddenly decreases after the Global Financial Crisis, in favor of most complex triadic structures, while all types of subgraphs tend to become more coherent thereafter.

A deep learning framework for crop mapping with reconstructed Sentinel-2 time series images

Timely and accurate access to regional scale crop plant area and spatial distribution is essential for regional agricultural production and food security, especially in the context of global population growth and climate change. Deep learning has become prevalent in crop mapping under complex conditions due to its powerful feature extraction and nonlinear ability. This study proposes a time-series image classification framework using Attention-based Bidirectional Gated Recurrent Unit (A-BiGRU) to map rice, maize, and soybean in Fujin, China, from reconstructed Sentinel-2 time-series images. Firstly, the reconstructed Sentinel-2 time-series images with 10 spectral dimensions and 22 temporal dimensions were obtained by linear interpolation and Savitzky-Golay (SG) filter. Then, a neural network, the A-BiGRU was developed to identify different crops by taking advantage of their unique growth patterns. The attention mechanism enables temporal neural networks to focus on the critical growth periods of crops. Additionally, the structure of GRU is simpler than that of long short-term memory (LSTM) and simple recurrent neural network (SRNN), which reduces the number of parameters and alleviates overfitting. Compared to GRU, BiGRU can fully uses the time-series information of the entire crop growth cycle. To assess the effectiveness of the proposed method, we compared two deep learning methods (LSTM and SRNN) and three widely used non-deep learning classifiers (Spectral Angle Mapping (SAM), Support Vector Machine (SVM)) and eXtreme Gradient Boosting (XGBoost). The results demonstrate that A-BiGRU achieved the highest accuracy, with an overall accuracy of 0.9804, a macro F1 score of 0.9788 and a kappa score of 0.9714. We also selected four typical regions and compared the classification results with optical images, which showed that the proposed method has a good recognition effect. Therefore, the A-BiGRU method is capable of achieving high-precision crop mapping.