Time Series Feature Extraction Research Articles

Feasibility Analysis for Predicting Indian Ocean Bigeye Tuna (Thunnus obesus) Fishing Grounds Based on Temporal Characteristics of FY-3 Microwave Radiation Imager Data

Efficient and accurate fishery forecasting is of great significance in ensuring the efficiency of fishery operations. This paper proposes a fishery forecasting method using a brightness temperature (TB) time series spatial feature extraction and fusion model. Using Indian Ocean bigeye tuna fishery data from 2009 to 2021 as a reference, this paper discusses the feasibility of fishery forecasting using FY-3 Microwave Radiation Imager (MWRI) Level 1 TB data. For this paper, we designed a deep learning network model for radiometer TB time series feature extraction (TimeTB-FishNet) based on the Convolutional Neural Network (CNN), Gated Recurrent Unit (GRU), and Attention mechanism. After expanding the dimensions of TB features, the model uses them together with spatiotemporal feature factors (year, month, longitude, and latitude) as features. By adding the GRU and Attention to the CNN, the CNN-GRU-Attention model architecture is established and can extract deep time series spatial features from the data to achieve the best results. In the model validation experiments, the TimeTB-FishNet model performed optimally, with a coefficient of determination (R2) of 0.6643. In the generalization experiments, the R2 also reached 0.6261, with a root mean square error (RMSE) of 46.6031 kg/1000 hook. When the sea surface height (SSH) was introduced, the R2 further reached 0.6463, with a lower RMSE of 45.1318 kg/1000 hook. The experimental results show that the proposed method and model are feasible and effective. The proposed model can directly use enhanced radiometer TB data without relying on lagging ocean environmental product data, performing deep temporal and spatial feature extraction for fishery forecasting. This method can provide a reference for the fishing of bigeye tuna in the Indian Ocean.

ASK-HAR: Attention-Based Multi-Core Selective Kernel Convolution Network for Human Activity Recognition

Human Activity Recognition (HAR) is an increasingly popular field of study aimed at automatically identifying and categorizing human movements and activities using various tracking devices and measurements, such as sensors and cameras. Smartphones have emerged as a prevalent sensor modality for HAR, offering abundant data on an individual’s movements through GPS, accelerometers, and gyroscopes. In conventional convolutional neural networks (CNNs), the artificial neurons within each feature layer typically possess identical receptive fields (RFs). This paper introduces a novel model, ASK-HAR (Attention-based Multi-Core Selective Kernel Convolution Network for HAR), which enhances HAR by performing kernel selection between multiple branches with different RFs using attention mechanisms. The selective kernel mechanism is leveraged to optimize HAR performance. Additionally, the CBAM attention module is employed for time series feature extraction and activity recognition within the overall framework. Extensive experiments conducted on benchmark HAR datasets, including UCI-HAR, USC-HAD, WISDM, PAMAP2, and DSADS, demonstrate that the ASK-HAR model consistently achieves high accuracy across all datasets.

Clustering and Interpretation of time-series trajectories of chronic pain using evidential c-means

The most well-known unsupervised classification algorithms allow for the identification of hard or probabilistic partitions. However, when working with complex datasets such as those found in the healthcare domain, it is important to note that these partitions may have limitations, particularly when considering the uncertainty associated with medical data. Indeed, these algorithms do not account for modeling outliers, imprecise, or uncertain observations. In this study, we focused on analyzing time series of barometers attributes related to pain intensity in individuals with chronic pain. Our goal was to identify different typologies of care trajectories, highlighting distinct trajectories of chronic pain through clustering techniques. We propose a soft clustering approach based on feature extraction and selection from sequential data, aiming to improve interpretability and enhance the performance of the clustering procedure. Time series feature extraction and selection effectively handles complex and noisy data, improving interpretability and clustering performance. Evidential soft clustering models data uncertainty and imprecision, providing a better representation of nuances and variations in complex, raw data. The first step involves extracting features from the available time series and selecting the most important attributes. For this purpose, we use a time series feature extractor (Tsfresh) and unsupervised feature selection methods, including unsupervised Random Forest, Laplacian Score, and unsupervised Spectral Feature Selection. The second step involves using the evidential c-means (ECM) clustering algorithm on the extracted attributes. The ECM method, based on belief functions, allows for generating a credal partition that has the ability to model various forms of uncertainty. This partition can then be transformed into a hard partition to study individuals based on this uncertainty criterion. The results reveal the existence of two clusters of chronic pain related to discomfort and well-being, exhibiting excellent separability and compactness. Additionally, an uncertain cluster emerges grouping patients with intermediate characteristics. The interpretability of the determined partitions through descriptive analysis, statistical tests, and repeated measures multinomial regression on clinical and demographic data allowed us to determine the profile of patients in the identified trajectories.

Complex Emotion Dynamics Contribute to the Prediction of Depression: A Machine Learning and Time Series Feature Extraction Approach.

Emotion dynamics have demonstrated mixed ability to predict depressive symptoms and outperform traditional metrics like the mean and standard deviation of emotion reports. Here, we expand the types of emotion dynamic features used in prior work and apply a machine learning algorithm to predict depression symptoms. We obtained seven ecological momentary assessment (EMA) studies from previous work on depression and emotion dynamics (N = 890). These studies measured self-reported sadness, positive affect, and negative affect 5 to 10 times per day for 7 to 21days (schedule varied across studies). These data were fed through a feature extraction routine to generate hundreds of emotion dynamic features. A gradient boosting machine (GBM) using all available emotion dynamics features was the best of all models assessed. This model's out-of-sample prediction (R 2 pred) for depression severity ranged from .20 to .44 depending on EMA interpolation method and samples included in the analysis. It also explained significantly more variance than a benchmark model of individuals' mean emotion ratings over the assessment period, R 2 pred = .089. Comprehensive feature mining of emotion dynamics obtained during EMA may be necessary to identify processes that predict depression symptoms beyond mean emotion ratings.

Study on lung CT image segmentation algorithm based on threshold-gradient combination and improved convex hull method

Lung images often have the characteristics of strong noise, uneven grayscale distribution, and complex pathological structures, which makes lung image segmentation a challenging task. To solve this problems, this paper proposes an initial lung mask extraction algorithm that combines threshold and gradient. The gradient used in the algorithm is obtained by the time series feature extraction method based on differential memory (TFDM), which is obtained by the grayscale threshold and image grayscale features. At the same time, we also proposed a lung contour repair algorithm based on the improved convex hull method to solve the contour loss caused by solid nodules and other lesions. Experimental results show that on the COVID-19 CT segmentation dataset, the advanced lung segmentation algorithm proposed in this article achieves better segmentation results and greatly improves the consistency and accuracy of lung segmentation. Our method can obtain more lung information, resulting in ideal segmentation effects with improved accuracy and robustness.

Time-Series Feature Extraction by Return Map Analysis and Its Application to Bearing-Fault Detection

Developing new features for time-series characterization is a current challenge in data science and machine learning. In this paper, we propose a new metric based on a simple and efficient algorithm, namely, the return map. The return map analysis is well established in the field of non-linear dynamics, in particular, for fitting the parameters of a chaotic system from a waveform, or to attack a chaotic communication channel. We show that our metrics work well for both non-linear dynamics and time-series feature extraction problems in the field of machine learning. In an experiment aiming to classify vibration signals of normal and damaged bearings, we compare our method with two other methods that reported to have excellent accuracy, based on entropy and statistical feature distribution, respectively. We show that our method achieves higher accuracy with almost the lowest time costs, which was confirmed in experiments with two different datasets containing three main classes of bearings: normal, with inner race faults, and with outer race faults, having different damage origins and recorded in various conditions. In particular, for the dataset supplied by Case Western Reserve University, our method reached an accuracy of 100% at signals of 5000 sample points length, with a total time of 0.4 s required for feature estimation, while the entropy-based method reached an accuracy of 95% with a time of 100 s, and a statistical feature distribution method reached an accuracy of 93% with a total time of 1.9 s. Results show that the developed method is better suited to real-time bearing condition monitoring applications than most of the methods reported to date.

Spatial weighting EMD-LSTM based approach for short-term PM2.5 prediction research

Given the significant health and environmental risks posed by atmospheric PM2.5 pollution, accurately predicting its concentration changes is especially important. Current models fall short in researching time-series feature extraction from pollutants and spatial correlations among monitoring stations. In this study, a spatiotemporal prediction model is introduced to address these issues. The model combines spatial weighting, empirical mode decomposition (EMD), and a long short-term memory (LSTM) network. First, weights are allocated to sites using Pearson correlation analysis and distance weighting methods. Next, the pollutant time series is decomposed using the EMD method. The highly correlated intrinsic mode function (IMF) component is selected for signal reconstruction, enhancing denoising. Finally, the model uses an LSTM network to capture nonlinear and dynamic time series traits, which significantly improves the PM2.5 prediction accuracy. The model utilizes data collected from 10 monitoring stations across Hefei city during 2018-2019, employing the previous 24 h of observations to forecast PM2.5 concentrations for the subsequent hour. By comparing with RNN, HPO-RNN, GRU, LSTM, and CBAM-CNN-Bi LSTM, the results show that our model surpasses five benchmark models in terms of prediction accuracy. Relative to the best-performing CBAM-CNN-Bi LSTM model, our model reduces RMSE and MAE by 73.91% and 72.99%, respectively, and improves R2 by 8.15%. In summary, the proposed spatial weighting EMD-LSTM model offers an efficient new approach for predicting atmospheric PM2.5 pollution. It integrates spatial and time series analysis, significantly enhancing the prediction accuracy.

Timeseries Feature Extraction for Dataset Creation in Prognostic Health Management

Timeseries Feature Extraction for Dataset Creation in Prognostic Health Management

MDCNet: Long-term time series forecasting with mode decomposition and 2D convolution

Long-term time series forecasting is widely used in various real-world applications, such as weather, traffic, energy, healthcare, etc. Recently, time series decomposition techniques have been adopted in many mainstream forecasting models, such as the prevalent Transformer-based models, to help capture sophisticated temporal patterns and achieve success in several benchmark tasks. However, conventional decomposition algorithms are often based on simple operations or limited to specific fields, and therefore are not effective and applicable enough, especially for complex time series. In this paper, we propose Mode Decomposition and 2D Convolutional Network (MDCNet), a structure-simple yet effective forecasting architecture based on a more effective decomposition method and a multi-frequency time series feature extraction network with multi-scale 2D convolution. Specifically, we first introduce a Variational Mode Decomposition Block to discover intricate time patterns, which decompose time series into trend components and stationary modal components at different main frequencies. Then, we design a Trend Prediction Block and an Intrinsic Mode Functions Prediction Block to capture global correlation and hidden dependencies within different main frequencies, respectively. Furthermore, a Frequency Enhancement Module is designed to further reduce the impact of noise in long-term time series. Experiments on eight benchmark datasets show that MDCNet significantly reduces the error of the previous state-of-the-art method by 15.1% and 11.5% for multivariate and univariate time series, respectively.

Multivariate Time Series Feature Extraction and Clustering Framework for Multi-Function Radar Work Mode Recognition

Multi-Function Radars (MFRs) are sophisticated sensors with great agility and flexibility in adapting their transmitted waveform and control parameters. The recognition of MFR work modes based on the intercepted pulse sequences plays an important role in interpreting the functional purpose and threats of a non-cooperative MFRs. However, due to the increased flexibility of MFRs, radar work modes with emerging new modulations and control parameters always appear, and the supervised classification method suffers performance degradation or even failure. Unsupervised learning and clustering of MFR pulse sequences becomes urgent and important. This paper establishes a unified multivariate MFR time series feature extraction and clustering framework for MFR work mode recognition. At first, various features are collected to form the feature set. The feature set includes features extracted through deep learning based on recurrent auto-encoders, multidimensional time series toolkit features, and manually crafted features for radar inter-pulse modulations. Subsequently, several feature selection algorithms, combined with different clustering and classification methods, are used for the selection of an “optimal” feature subset. Finally, the effectiveness and superiority of the proposed framework and selected features are validated through simulated and measured datasets. In the simulated dataset containing 20 classes of work modes, under the most severe non-ideal conditions, we achieve a clustering purity of 73.46% and an NMI of 84.28%. In the measured dataset with seven classes of work modes, we achieve a clustering purity of 86.96% and an NMI of 90.10%.

Comparison of neural network architectures for feature extraction from binary black hole merger waveforms

We evaluate several neural-network architectures, both convolutional and recurrent, for gravitational-wave time-series feature extraction by performing point parameter estimation on noisy waveforms from binary-black-hole mergers. We build datasets of 100 000 elements for each of four different waveform models (or approximants) in order to test how approximant choice affects feature extraction. Our choices include SEOBNRv4P and IMRPhenomPv3, which contain only the dominant quadrupole emission mode, alongside IMRPhenomPv3HM and NRHybSur3dq8, which also account for high-order modes. Each dataset element is injected into detector noise corresponding to the third observing run of the LIGO-Virgo-KAGRA (LVK) collaboration. We identify the temporal convolutional network architecture as the overall best performer in terms of training and validation losses and absence of overfitting to data. Comparison of results between datasets shows that the choice of waveform approximant for the creation of a dataset conditions the feature extraction ability of a trained network. Hence, care should be taken when building a dataset for the training of neural networks, as certain approximants may result in better network convergence of evaluation metrics. However, this performance does not necessarily translate to data which is more faithful to numerical relativity simulations. We also apply this network on actual signals from LVK runs, finding that its feature-extracting performance can be effective on real data.

Deep learning model for state of health estimation of lithium batteries based on relaxation voltage

Lithium-ion batteries have been widely used in many fields due to their high energy density, long cycle life and low self-discharge rate. However, most existing SOH estimation methods require a lengthy amount of time to determine features, the acquisition of multi-dimensional raw data is challenging, and the application scenarios are limited, making SOH estimation algorithms less practical in developing battery management systems (BMS). To face the challenge, this paper develops a new battery SOH estimation method and investigates the transferability of machine learning models under multiple operating conditions using transfer learning. First, features are extracted from the 15 s relaxation voltage using tsfresh, a Python package for time series feature extraction. Subsequently, four statistical features were obtained using a feature selection method based on the Pearson correlation coefficient. Then, a Bidirectional Long Short-Term Memory (Bi-LSTM) model is established and hyperparameters are adjusted using Bayesian optimization. The root mean square error (RMSE) and the mean absolute percentage error (MAPE) of battery SOH estimation on battery data set are 1.210 % and 1.225 %, respectively. Finally, based on transfer learning, the Bi-LSTM model is applied to battery datasets under multiple operating conditions, reducing the model's dependence on original training data while ensuring model accuracy. Considering the easy accessibility of short-term relaxation battery data, the battery SOH estimation method proposed in this paper is expected to facilitate the estimation of battery SOH under various operating conditions.

Uncovering a stability signature of brain dynamics associated with meditation experience using massive time-series feature extraction

Previous research has examined resting electroencephalographic (EEG) data to explore brain activity related to meditation. However, previous research has mostly examined power in different frequency bands. The practical objective of this study was to comprehensively test whether other types of time-series analysis methods are better suited to characterize brain activity related to meditation. To achieve this, we compared >7000 time-series features of the EEG signal to comprehensively characterize brain activity differences in meditators, using many measures that are novel in meditation research. Eyes-closed resting-state EEG data from 49 meditators and 46 non-meditators was decomposed into the top eight principal components (PCs). We extracted 7381 time-series features from each PC and each participant and used them to train classification algorithms to identify meditators. Highly differentiating individual features from successful classifiers were analysed in detail. Only the third PC (which had a central-parietal maximum) showed above-chance classification accuracy (67 %, pFDR = 0.007), for which 405 features significantly distinguished meditators (all pFDR < 0.05). Top-performing features indicated that meditators exhibited more consistent statistical properties across shorter subsegments of their EEG time-series (higher stationarity) and displayed an altered distributional shape of values about the mean. By contrast, classifiers trained with traditional band-power measures did not distinguish the groups (pFDR > 0.05). Our novel analysis approach suggests the key signatures of meditators’ brain activity are higher temporal stability and a distribution of time-series values suggestive of longer, larger, or more frequent non-outlying voltage deviations from the mean within the third PC of their EEG data. The higher temporal stability observed in this EEG component might underpin the higher attentional stability associated with meditation. The novel time-series properties identified here have considerable potential for future exploration in meditation research and the analysis of neural dynamics more broadly.

Asteroids co-orbital motion classification based on Machine Learning

ABSTRACT In this work, we explore how to classify asteroids in co-orbital motion with a given planet using Machine Learning. We consider four different kinds of motion in mean motion resonance with the planet, nominally Tadpole at L4 and L5, Horseshoe and Quasi-Satellite, building three data sets defined as Real (taking the ephemerides of real asteroids from the JPL Horizons system), Ideal and Perturbed (both simulated, obtained by propagating initial conditions considering two different dynamical systems) for training and testing the Machine Learning algorithms in different conditions. The time series of the variable θ (angle related to the resonance) are studied with a data analysis pipeline defined ad hoc for the problem and composed by: data creation and annotation, time series features extraction thanks to the tsfresh package (potentially followed by selection and standardization) and the application of Machine Learning algorithms for Dimensionality Reduction and Classification. Such approach, based on features extracted from the time series, allows to work with a smaller number of data with respect to Deep Learning algorithms, also allowing to define a ranking of the importance of the features. Physical interpretability of the features is another key point of this approach. In addition, we introduce the SHapley Additive exPlanations for Explainability technique. Different training and test sets are used, in order to understand the power and the limits of our approach. The results show how the algorithms are able to identify and classify correctly the time series, with a high degree of performance.

Dimensionality reduction and features visual representation based on conditional probabilities applied to activity classification

A large part of the information emitted by contemporary technological devices comes in the form of time series. The massive commercialization of these kinds of devices has made the study of time series feature extraction techniques acquire a vital relevance in last years. Two main things are essential when applying feature extraction techniques to time series: to reduce the dimensionality so it occupies the least amount of storage memory possible, and to make features that contain the relevant information regarding the nature of the data set and the goals to be achieved. For this purpose, we propose in this work a brand new technique called the State Changes Representation for Time Series (SCRTS), which relies on the relevant data associated with the conditional probabilities of the time series (also known in the literature as Markov model’s features), and the distribution of its values. This method is length-independent, which means that we can apply it to time series of different dimensions obtaining the same number of features for each one. Also, it provides a visual representation of the input data, so it is possible to interpret what makes a certain time series different from the other. After explaining how it works, we apply it to 3 different wearable accelerometer data sets. This algorithm reduces the original dimension of the time series considerably (in the best case from 5499 values to 31), having a good performance in the classification results (in the best chance with an accuracy of 98%).

A multi-agent reinforcement learning framework for optimizing financial trading strategies based on TimesNet

An increasing number of studies have shown the effectiveness of using deep reinforcement learning to learn profitable trading strategies from financial market data. However, a single-agent model is not sufficient to handle complex financial scenarios. To address this problem, a novel approach called Multi-Agent Double Deep Q-Network (Later called MADDQN) is proposed in this study, which reasonably balances the pursuit of maximum revenue and the avoidance of risk under the multi-agent reinforcement learning framework by innovatively employing two different agents represented respectively by two time-series feature extraction networks, TimesNet, and the Multi-Scale Convolutional Neural Network. Furthermore, to achieve a more generalized model suitable for different underlying assets, a mixed dataset containing three major U.S. stock indexes is collected. And the proposed model has been pre-trained in this dataset and subsequently refined for the specified asset. The results from experiments on five different stock indices show that the proposed MADDQN has an average cumulative return of 23.08%, outperforming the other baseline methods. Besides, the multi-agent model demonstrates its advantage in balancing the risk and revenue, in comparison with the single-agent models. Additionally, The generalization experiments confirm that the proposed MADDQN method after pre-training in the proposed mixed dataset could be stably transferred to the other underlying assets with a refinement. These findings indicate that the proposed framework not only achieves good performance in complex financial market environments but also is able to generalize robustly across different scenarios in various markets.

Three-hourly PM2.5 and O3 concentrations prediction based on time series decomposition and LSTM model with attention mechanism

In recent years, air quality has attracted wide attention from all over the world, among which the high concentration of particulate matter with an aerodynamic less than 2.5 μm (PM2.5) and ozone (O3) has a great adverse impact on human health and daily life. Previous studies on two pollutant predictions are overly concerned with model improvement but hardly focus on influence variable screening, and pollutant's time series features extraction and identification. To better improve the prediction accuracy and enhance the application of the model in practice, in the present study, a novel model RF-CEEMDAN-Attention-LSTM was proposed, which has three processes: (1) random forest (RF) screened out highly correlated influence variables; (2) complete ensemble empirical mode decomposition with adaptive noise (CEEMADN) method was adopted to decompose the PM2.5 and O3 concentration time series into multiple sub-time sequences; (3) the double hidden layer LSTM model with attention and dropout mechanism captured nonlinear relationships and dynamic changes of time series features. Three-hourly PM2.5 and O3 concentration in Chengdu were used to validate the effectiveness of the developed RF-CEEMDAN-Attention-LSTM model by comparing five other parallel models. The final results showed that the model not only had a better fitting effect on both PM2.5 and O3 than other comparable models in the entire timeline, but the model also had the highest R2 (PM2.5:0.916, O3:0.525) values for these two air pollutants at high concentration values.

Prediction method of peak carbon emission in power grid based on energy consumption elasticity coefficient

With the implementation of China’s “two-carbon” target strategy, the status of renewable energy in the power system is gradually improving. In order to improve the carbon emission control level of power grid, it is necessary to predict the peak carbon emission of power grid. A prediction method of grid carbon emission peak based on energy elasticity coefficient is proposed. A time series model of grid carbon emission samples was constructed, and a combination of combinatorial sorting and machine learning was used to reconstruct the time series of grid carbon emission peaks. Based on the results of time series reconstruction, feature extraction and classification training of grid carbon emission peak are carried out, and the measurement error correction of grid carbon emission under the constraint of energy consumption elasticity is realized. Combined with the association rule mining method, the fusion processing of grid carbon emission peak samples under the elastic constraint of energy consumption is realized. Through characteristic detection and fusion processing of energy consumption elasticity coefficient, the reconstituted carbon emission samples of power grid constrained by energy consumption elasticity were reconstructed, and the peak carbon emission of power grid was predicted. The simulation results show that this method has high accuracy and convergence in predicting the peak carbon emission of power grid, and improves the error correction ability in the prediction process.

Classification of recurrent major depressive disorder using a new time series feature extraction method through multisite rs-fMRI data

BackgroundMajor depressive disorder (MDD) has a high rate of recurrence. Identifying patients with recurrent MDD is advantageous in adopting prevention strategies to reduce the disabling effects of depression. MethodWe propose a novel feature extraction method that includes dynamic temporal information, and inputs the extracted features into a graph convolutional network (GCN) to achieve classification of recurrent MDD. We extract the average time series using an atlas from resting-state functional magnetic resonance imaging (fMRI) data. Pearson correlation was calculated between brain region sequences at each time point, representing the functional connectivity at each time point. The connectivity is used as the adjacency matrix and the brain region sequences as node features for a GCN model to classify recurrent MDD. Gradient-weighted Class Activation Mapping (Grad-CAM) was used to analyze the contribution of different brain regions to the model. Brain regions making greater contribution to classification were considered to be the regions with altered brain function in recurrent MDD. ResultWe achieved a classification accuracy of 75.8 % for recurrent MDD on the multi-site dataset, the Rest-meta-MDD. The brain regions closely related to recurrent MDD have been identified. LimitationThe pre-processing stage may affect the final classification performance and harmonizing site differences may improve the classification performance. ConclusionThe experimental results demonstrate that the proposed method can effectively classify recurrent MDD and extract dynamic changes of brain activity patterns in recurrent depression.

Wind Turbine Blade Early Fault Detection With Faulty Label Unknown and Labeling Bias

In practical industrial applications, the need for a large number of accurately labeled training samples is a significant challenge for fault detection tasks. However, labeling all training samples is expensive and prone to labeling errors, especially for early fault detection of wind turbine blades. This paper proposes a labeling bias (LB) hypothesis. Assuming the labeler only needs to label parts of normal samples that are easy to judge, we design a probability ratio least-squares importance fitting (PRL-SIF) method based on variable homogeneity. Unlike other state-of-the-art positive unlabeled (PU) learning methods, PRL-SIF does not require knowledge of the class priors to achieve training. Furthermore, to better handle the multi-dimensional time-series data of wind turbines, we provide a data preprocessing method based on functional analysis to achieve time series feature extraction and dimensionality reduction. The effectiveness and robustness of the proposed method are verified on 23 real-world wind turbine datasets. Experimental results show that the proposed method can achieve nearly 90% accuracy while only labeling 20% of normal samples.