Long-Short Term Research Articles

ENFformer: Long-short term representation of electric network frequency for digital audio tampering detection

Detecting digital audio tampering is essential for ensuring judicial fairness and societal security. Traditional methods, primarily based on Electric Network Frequency (ENF), have been limited by their reliance on singular, static features, which overlook critical temporal dynamics inherent in ENF data. This results in suboptimal detection accuracy. Addressing these limitations, this paper introduces a novel transformer model, ENFformer, designed for the detection of digital audio tampering by leveraging both short and long-term temporal features of ENF data. Initially, the ENFformer model extracts traditional zero-order and first-order phase features using discrete Fourier transforms (DFT0 and DFT1), along with frequency features obtained through the Hilbert transform. These features are processed using a frame-based algorithm to develop their temporal counterparts. To enhance feature extraction, the model employs a two-layer one-dimensional Convolutional Long Short-Term Memory (ConvLSTM) network to assimilate short-term temporal features, followed by a Bidirectional Long Short-Term Memory (BiLSTM) network for long-term feature integration. A branch attention mechanism then synergizes these long-term features, which are further refined by a transformer module for accurate tampered audio identification. Our empirical evaluations on the Carioca and ENF-EDIT1 databases demonstrate that ENFformer achieves detection accuracies of 97.33% and 93.50% respectively, surpassing existing state-of-the-art methods. These results confirm the effectiveness of our approach, which significantly advances the field of digital audio tampering detection by incorporating a comprehensive analysis of temporal information in ENF features. The source code of this study is publicly available at https://github.com/CCNUZFW/ENFformer.

Forecasting the Future Value of a Colombian Investment Fund with LSTM Recurrent Neural Networks (LSTM)

Forecasting the Future Value of a Colombian Investment Fund with LSTM Recurrent Neural Networks (LSTM)

LONG SHORT-TERM RELEVANCE LEARNING

To incorporate sparsity knowledge as well as measurement uncertainties in the traditional long short-term memory (LSTM) neural networks, an efficient relevance vector machine algorithm is introduced to the network architecture. The proposed scheme automatically determines relevant neural connections and adapts accordingly, in contrast to the classical LSTM solution. Due to its flexibility, the new LSTM scheme is less prone to overfitting and hence can approximate time-dependent solutions by use of a smaller data set. On a structural nonlinear finite element application, we show that the self-regulating framework does not require prior knowledge of a suitable network architecture and size, while ensuring satisfying accuracy at reasonable computational cost.

Stator Turn Fault Diagnosis and Severity Assessment in Converter-Fed Induction Motor Using Flat Diagnosis Structure Based on Deep Learning Approach

This work aims to explore the hybrid deep learning techniques for early interturn fault diagnosis and fault severity assessment in variable frequency induction motor drive. However, the existing literature mostly focuses on identifying the fault patterns, whereas the assessment of fault deterioration before it reaches the extreme degradation state has been less investigated. In this paper, a hybrid structure combining two-dimensional convolution neural network (2DCNN) and long-short term memory/gated recurrent unit (LSTM/GRU) has been modelled as a relatively lightweight flat diagnosis structure (FDS) for early fault identification and its severity assessment simultaneously in converter-fed induction motor. Initially, the hierarchical layers of 2DCNN extract the effective features automatically for fault detection and its severity identification simultaneously. Then, these features are presented as suitable inputs to LSTM/GRU for classification. A comparative evaluation of the proposed hybrid method with the standalone structure, namely 2DCNN, LSTM and SVM proves the advantages of the proposed scheme, and its applicability for both fault diagnosis and fault severity assessment in industrial drive systems, under variable frequencies and load variations.

Monthly precipitation prediction based on the EMD–VMD–LSTM coupled model

Abstract Precipitation prediction is one of the important issues in meteorology and hydrology, and it is of great significance for water resources management, flood control, and disaster reduction. In this paper, a precipitation prediction model based on the empirical mode decomposition–variational mode decomposition–long short-term memory (EMD–VMD–LSTM) is proposed. This model is coupled with EMD, VMD, and LSTM to improve the accuracy and reliability of precipitation prediction by using the characteristics of EMD for noise removal, VMD for trend extraction, and LSTM for long-term memory. The monthly precipitation data from 2000 to 2019 in Luoyang City, Henan Province, China, are selected as the research object. This model is compared with the standalone LSTM model, EMD–LSTM coupled model, and VMD–LSTM coupled model. The research results show that the maximum relative error and minimum relative error of the precipitation prediction using the EMD–VMD–LSTM neural network coupled model are 9.64 and −7.52%, respectively, with a 100% prediction accuracy. This coupled model has better accuracy than the other three models in predicting precipitation in Luoyang City. In summary, the proposed EMD–VMD–LSTM precipitation prediction model combines the advantages of multiple methods and provides an effective way to predict precipitation.

Deep learning based on the application of voice emotion

A language is a valuable tool for human development and progress, and it is also an important medium for human beings to transmit information and express emotions. Language signals are ubiquitous, and it is an indispensable part of human life. This article will take the analysis of language as the starting point, combined with the relevant content of computer deep learning, and summarize various methods of language emotion recognition based on a convolutional neural network. In recent years, with the gradual intelligentization of computers, more in-depth discoveries and research have been made on language emotion research. In the deep neural network sector, most of the models used are CNN, LSTM, MO-LSTM models, and this paper aims to propose a new CLDNN (CONVOLUTIONAL, LONG SHORT-TERM) that integrates CNN, DNN, and LSTM into the same network. MEMORY, FULL CONNECTED DEEP NEURAL NETWORKS) model, compare with it, and summarize the advantages and disadvantages of CLDNN.

AN EMOTION RECOGNITION MODEL BASED ON LONG SHORT-TERM MEMORY NETWORKS AND EEG SIGNALS AND ITS APPLICATION IN PARAMETRIC DESIGN

One of the design objectives of a product is to create a positive emotional user experience. Through careful design, the product can evoke emotional resonance in users and stimulate their pleasure and satisfaction. Therefore, emotion recognition is crucial for parameterized product design. Considering that emotion recognition based on electroencephalogram (EEG) signals is more objective and accurate compared to methods such as text and surveys, this paper proposes an emotion analysis model based on long short-term memory (LSTM) and EEG and applies it to parameterized design. The main contributions of this paper are as follows. (1) Constructing a high-accuracy emotion recognition model. First, EEG data reflecting the characteristic patterns of brain activities in different emotional states are collected through EEG electrodes. Then, the EEG data are input into the LSTM network for training, enabling it to learn and capture the features associated with emotional states. During the training process, the model learns to extract crucial emotional features from the EEG data for emotion state recognition. This model can automatically learn emotional features, handle long-term dependencies and provide a more accurate and reliable solution for emotion recognition tasks. (2) Creating an EEG dataset specifically for evaluating emotions related to a product and using the trained emotion recognition model to classify this dataset, obtaining emotion classification results. The emotion classification results can be used to determine which parameter designs in product development need to be retained or discarded. These parameter designs can involve aspects such as user experience, functionality, aesthetics, usability and user-friendliness. Decisions can be made based on the emotion classification results to improve the quality and user satisfaction of the product.

DYNAMIC PREDICTION OF RESIDUAL LIFE WITH LONGITUDINAL COVARIATES USING LONG SHORT-TERM MEMORY NETWORKS.

Sepsis, a complex medical condition that involves severe infections with life-threatening organ dysfunction, is a leading cause of death worldwide. Treatment of sepsis is highly challenging. When making treatment decisions, clinicians and patients desire accurate predictions of mean residual life (MRL) that leverage all available patient information, including longitudinal biomarker data. Biomarkers are biological, clinical, and other variables reflecting disease progression that are often measured repeatedly on patients in the clinical setting. Dynamic prediction methods leverage accruing biomarker measurements to improve performance, providing updated predictions as new measurements become available. We introduce two methods for dynamic prediction of MRL using longitudinal biomarkers. in both methods, we begin by using long short-term memory networks (LSTMs) to construct encoded representations of the biomarker trajectories, referred to as "context vectors." In our first method, the LSTM-GLM, we dynamically predict MRL via a transformed MRL model that includes the context vectors as covariates. In our second method, the LSTM-NN, we dynamically predict MRL from the context vectors using a feed-forward neural network. We demonstrate the improved performance of both proposed methods relative to competing methods in simulation studies. We apply the proposed methods to dynamically predict the restricted mean residual life (RMRL) of septic patients in the intensive care unit using electronic medical record data. We demonstrate that the LSTM-GLM and the LSTM-NN are useful tools for producing individualized, real-time predictions of RMRL that can help inform the treatment decisions of septic patients.

A Survey of Unsupervised Learning Algorithms for Zero-Day Attacks in Intrusion Detection Systems.

Intrusion detection systems (IDS) are systems that are used to monitor networks for malicious events, abnormal activities, and policy violations. They are systems that are capable of detecting and classifying network attacks based on behaviors or signatures of previously known attacks based on markers. However, since network attacks are constantly evolving and it is almost impossible to infuse all possible combinations and signatures of the attacks, the effectiveness of Machine Learning based IDS is often challenged and called into play as a result of novel attacks generated, known as Zero-day attacks. This has facilitated the need to have intelligent-based IDS that could detect anomalies without relying on a detailed signature repository. In this paper, we present a literature-based survey of popular deep learning algorithms and evaluated their capabilities, strengths, limitations, and resource requirements for detecting anomalies and Zero-Day attacks. Based on our evaluation, we propose Long Short-Term (LSTM) networks and Autoencoder networks as the best algorithms for further analysis in intrusion detection.

Design and Implementation of Technical Analysis Based LSTM Model for Stock Price Prediction

The paper describes the design and implementation of a Long Short-Term Memory (LSTM) model for stock price prediction based on technical analysis. The model use technical indicators such as moving averages and Bollinger Bands to discover trends in the stock market and forecast future stock values. Historical stock data was used to extract technical indicators for the model. These indicators were then used as input features to train the LSTM model using a supervised learning strategy. Metrics such as mean absolute error, mean squared error, and root mean squared error were used to assess the model's performance. However, as investment became more accessible, the stock market became more difficult and volatile. This paper proposes a stock price prediction system that employs a (LSTM) oriented neural network to forecast the next-day closing price of APPLE shares. Regression and LONG SHORT-TERM MEMORY models are constructed using selected input variables, and their performance is evaluated using RMSE, MAPE, and R squared error metrics to analyze the stock's trend for buying and selling.

Prediction of COVID-19 Spreading using LSTM Recurrent Neural Networks in Southeast Asia

The world has come to a standstill due to the coronavirus. The number of confrmed cases and the number of deaths are increasing day by day. Now accurate and real-time predictions are very important. Currently, a lot of work is being done on the spread prediction of coronavirus. If we can predict the number of corona patients in a country in advance, then it will be benefcial for the government of that country to take action in advance. In this paper, we experiment with Bangladesh and India data using the deep learning method. We have used LSTM (LONG SHORT-TERM MEMORY) neural network which is a type of recurrent neural network. The LSTM method can work very well even with very little data. Since we do not currently have much data on covid19, LSTM neural networks may be a suitable model. We have got very good prediction results using this method even with very little data. The prediction of various parameters (number of confrmed cases per day, number of deaths per day, number of total confrmed cases, and number of total deaths). We also calculate MAPE, MAE, and RMSE values and compare them among different models. Our modeloutperforms others’ models. We think this research will be a benefcial tool for administrators and health offcials. GUB JOURNAL OF SCIENCE AND ENGINEERING, Vol 8, Dec 2021 P 1-6

Compound Positioning Method for Connected Electric Vehicles Based on Multi-Source Data Fusion

With the development of electrified transportation, electric vehicle positioning technology plays an important role in improving comprehensive urban management ability. However, the traditional positioning methods based on the global positioning system (GPS) or roadside single sensors make it hard to meet requirements of high-precision positioning. Considering the advantages of various sensors in the cooperative vehicle-infrastructure system (CVIS), this paper proposes a compound positioning method for connected electric vehicles (CEVs) based on multi-source data fusion technology, which can provide data support for the CVIS. Firstly, Dempster-Shafer (D-S) evidence theory is used to fuse the position probability in multi-sensor detection information, and screen vehicle existence information. Then, a hybrid neural network model based on a long short-term (LSTM) framework is constructed to fit the mapping relationship between measured and undetermined coordinates. Moreover, the fused data are proceeded as the input of the hybrid LSTM model, which can export the vehicular real-time compound positioning information. Finally, an intersection in Shijingshan District, Beijing is selected as the test field for trajectory information collection of CEVs. The experimental results have shown that the uncertainty of fusion data can be reduced to 0.38% of the original level, and the maximum error of real-time positioning accuracy is less than 0.0905 m based on the hybrid LSTM model, which can verify the effectiveness of the model.

Automated detection of scar-related ventricular tachycardia origins from implanted device electrograms: a combined computational-AI platform

Abstract Funding Acknowledgements Type of funding sources: Public Institution(s). Main funding source(s): EPSRC Background Existing strategies that identify ventricular tachycardia (VT) ablation targets either employ invasive and time-consuming electrophysiological (EP) mapping, or non-invasive modalities that utilise standard electrocardiogram (ECG) signals. Success of these pre-procedure ablation approaches in localising re-entrant VTs often relies on VT induction, which could be avoided by utilising recordings of clinical VT episodes stored as electrograms (EGMs) in implanted devices. Such a non-invasive approach that localises VT substrates from EGMs may aid ablation planning, enhancing safety and speed. Purpose Our goal is to automate scar-related VT localisation by utilising EGM recordings of VT episodes from implanted devices. To achieve this, deep-learning algorithms will be trained on computational data to return VT sites of origin from implanted device EGMs. Ultimately, we intend to utilise this computational-artificial intelligence (AI) framework to detect ablation targets of clinical VT episodes and guide pre-procedure ablation planning non-invasively. Methods A comprehensive library of ECGs and EGMs from simulated paced beats (~15000) and scar-related VTs (500) was generated across five detailed torso models within a fast EP computational environment, combining reaction-eikonal and lead field methods. ECG (or EGM) traces from simulated paced beats were used to initially pre-train two convolutional neural network (CNN) long short-term (LSTM) attention-based architectures. Subsequently, signals of the in-silico, re-entrant VTs were used to re-train the networks to output the sites of origin of these episodes in a standardised ventricular coordinate space. Finally, the retrained CNN architectures were tested on re-entrant VTs of unseen models, and median localisation errors (LEs) were estimated against known VT origins from simulations. Results The performance of the networks to localise scar-related VT episodes was asserted for each torso model. When a torso model was only seen during initial training on simulated paced beats, implanted device EGMs and ECGs successfully localised VT sources with LEs 10.04 – 16.36 mm and 10.05 – 12.79 mm, respectively. When a torso model was not seen during pacing or VT training, recreating potential clinical application settings where ECGs or EGMs of clinical VTs would be the only inputs to the networks, LEs ranged 12.42 - 22.79 mm and 12.41 - 19.68 mm for EGM and ECG-based testing, respectively. Conclusions Our study successfully detected VT ablation substrates with accuracy that could be beneficial in clinical ablation settings. The proposed computational-AI framework may be used to automate the localisation of scar-related VTs from clinical ECGs or EGM recordings from implanted devices, ultimately aiding ablation planning.

Multi-step-ahead solar irradiance modeling employing multi-frequency deep learning models and climatic data

In this paper two enhanced long-short-term memory (LSTM) models of sequenced-LSTM (SLSTM) and wavelet-LSTM (WLSTM), provided for multi-step-ahead simulation of solar irradiance of six stations, located in Iran and USA. In this respect, twenty-year recorded solar irradiance and climate data were employed. The proposed multi-frequency models serve all the capabilities of classic LSTM network and also handle its weakness in detecting and modeling multi-frequency information that often included in natural datasets. The suggested methodology improved the long-short auto-regressive term of climate-solar irradiance data by including very long frequencies of time series. The results revealed that the suggested multi-frequency LSTM methods could exceed the feed forward neural network and classic LSTM network in test phase up to 23% and 13%, respectively.

Adaptive ML-based technique for renewable energy system power forecasting in hybrid PV-Wind farms power conversion systems

Large scale integration of renewable energy system with classical electrical power generation system requires a precise balance to maintain and optimize the supply–demand limitations in power grids operations. For this purpose, accurate forecasting is needed from wind energy conversion systems (WECS) and solar power plants (SPPs). This daunting task has limits with long-short term and precise term forecasting due to the highly random nature of environmental conditions. This paper offers a hybrid variational decomposition model (HVDM) as a revolutionary composite deep learning-based evolutionary technique for accurate power production forecasting in microgrid farms. The objective is to obtain precise short-term forecasting in five steps of development. An improvised dynamic group-based cooperative search (IDGC) mechanism with a IDGC-Radial Basis Function Neural Network (IDGC-RBFNN) is proposed for enhanced accurate short-term power forecasting. For this purpose, meteorological data with time series is utilized. SCADA data provide the values to the system. The improvisation has been made to the metaheuristic algorithm and an enhanced training mechanism is designed for the short term wind forecasting (STWF) problem. The results are compared with two different Neural Network topologies and three heuristic algorithms: particle swarm intelligence (PSO), IDGC, and dynamic group cooperation optimization (DGCO). The 24 h ahead are studied in the experimental simulations. The analysis is made using seasonal behavior for year-round performance analysis. The prediction accuracy achieved by the proposed hybrid model shows greater results. The comparison is made statistically with existing works and literature showing highly effective accuracy at a lower computational burden. Three seasonal results are compared graphically and statistically.

Distribution System State Estimation Using Model-Optimized Neural Networks

Maintaining reliability during power system operation relies heavily on the operator’s knowledge of the system and its current state. With the increasing complexity of power systems, full system monitoring is needed. Due to the costs to install and maintain measurement devices, a cost-effective optimal placement is normally employed, and as such, state estimation is used to complete the picture. However, in order to provide accurate state estimates in the current power system climate, the models must be fully expanded to include probabilistic uncertainties and non-linear assets. Recognizing its analogous relationship with state estimation, machine learning and its ability to summarily model unseen and complex relationships between input data is used. Thus, a power system state estimator was developed using modified long short-term (LSTM) neural networks to provide quicker and more accurate state estimates over the conventional weighted least squares-based state estimator (WLS-SE). The networks are then subject to standard polynomial scheduled weight pruning to further optimize the size and memory consumption of the neural networks. The state estimators were tested on a hybrid AC/DC distribution system composed of the IEEE 34-bus AC test system and a 9-bus DC microgrid. The conventional WLS-SE has achieved a root mean square error (RMSE) of 0.0151 p.u. for voltage magnitude estimates, while the LSTM’s were able to achieve RMSE’s between 0.0019 p.u. and 0.0087 p.u., with the latter having 75% weight sparsity, estimates about ten times faster, and half of its full memory requirement occupied.

Proposing a Fuzzy Soft‐max‐based classifier in a hybrid deep learning architecture for human activity recognition

Human Activity Recognition (HAR) is the process of identifying and analysing activities performed by a person (or persons). This paper proposes an efficient HAR system based on wearable sensors that uses deep learning techniques. The proposed HAR takes the advantage of staking Convolutional Neural Network and Long Short-Term (LSTM), for extracting the high-level features of the sensors data and for learning the time-series behaviour of the abstracted data, respectively. This paper proposed a Fuzzy Soft-max classifier for the dense layer which classifies the output of LSTM Blocks to the associated activity classes. The authors’ decision for proposing this classifier was because sensor data related to the resembling human activities, such as walking and running or opening door and closing door, are often very similar to each other. For this reason, the authors expect that adding fuzzy inference power to the standard Soft-max classifier will increase its accuracy for distinguishing between similar activities. The authors were also interested in considering a post-processing module that considers activity classification over a longer period. Using the proposed Fuzzy Soft-max classifier and by the post-processing technique, the authors were able to reach the 97.03 and 85.1 rates of accuracy for the PAMAP2 and Opportunity dataset, respectively.

Performance Evaluation of Neural Network-Based Short-Term Solar Irradiation Forecasts

Due to the globally increasing share of renewable energy sources like wind and solar power, precise forecasts for weather data are becoming more and more important. To compute such forecasts numerous authors apply neural networks (NN), whereby models became ever more complex recently. Using solar irradiation as an example, we verify if this additional complexity is required in terms of forecasting precision. Different NN models, namely the long-short term (LSTM) neural network, a convolutional neural network (CNN), and combinations of both are benchmarked against each other. The naive forecast is included as a baseline. Various locations across Europe are tested to analyze the models’ performance under different climate conditions. Forecasts up to 24 h in advance are generated and compared using different goodness of fit (GoF) measures. Besides, errors are analyzed in the time domain. As expected, the error of all models increases with rising forecasting horizon. Over all test stations it shows that combining an LSTM network with a CNN yields the best performance. However, regarding the chosen GoF measures, differences to the alternative approaches are fairly small. The hybrid model’s advantage lies not in the improved GoF but in its versatility: contrary to an LSTM or a CNN, it produces good results under all tested weather conditions.

Spatio-Temporal Variations and Driving Forces of Harmful Algal Blooms in Chaohu Lake: A Multi-Source Remote Sensing Approach

Harmful algal blooms (hereafter HABs) pose significant threats to aquatic health and environmental safety. Although satellite remote sensing can monitor HABs at a large-scale, it is always a challenge to achieve both high spatial and high temporal resolution simultaneously with a single earth observation system (EOS) sensor, which is much needed for aquatic environment monitoring of inland lakes. This study proposes a multi-source remote sensing-based approach for HAB monitoring in Chaohu Lake, China, which integrates Terra/Aqua MODIS, Landsat 8 OLI, and Sentinel-2A/B MSI to attain high temporal and spatial resolution observations. According to the absorption characteristics and fluorescence peaks of HABs on remote sensing reflectance, the normalized difference vegetation index (NDVI) algorithm for MODIS, the floating algae index (FAI) and NDVI combined algorithm for Landsat 8, and the NDVI and chlorophyll reflection peak intensity index (ρchl) algorithm for Sentinel-2A/B MSI are used to extract HAB. The accuracies of the normalized difference vegetation index (NDVI), floating algae index (FAI), and chlorophyll reflection peak intensity index (ρchl) are 96.1%, 95.6%, and 93.8% with the RMSE values of 4.52, 2.43, 2.58 km2, respectively. The combination of NDVI and ρchl can effectively avoid misidentification of water and algae mixed pixels. Results revealed that the HAB in Chaohu Lake breaks out from May to November; peaks in June, July, and August; and more frequently occurs in the western region. Analysis of the HAB’s potential driving forces, including environmental and meteorological factors of temperature, rainfall, sunshine hours, and wind, indicated that higher temperatures and light rain favored this HAB. Wind is the primary factor in boosting the HAB’s growth, and the variation of a HAB’s surface in two days can reach up to 24.61%. Multi-source remote sensing provides higher observation frequency and more detailed spatial information on a HAB, particularly the HAB’s long-short term changes in their area.

Deep Learning-Based Approach for Detecting Trajectory Modifications of Cassini-Huygens Spacecraft

There were necessary trajectory modifications of Cassini spacecraft during its last 14 years movement cycle of the interplanetary research project. In the scale 1.3 hour of signal propagation time and 1.4-billion-kilometer size of Earth-Cassini channel, complex event detection in the orbit modifications requires special investigation and analysis of the collected big data. The technologies for space exploration warrant a high standard of nuanced and detailed research. The Cassini mission has accumulated quite huge volumes of science records. This generated a curiosity derives mainly from a need to use machine learning to analyze deep space missions. For energy saving considerations, the communication between the Earth and Cassini was executed in non-periodic mode. This paper provides a sophisticated in-depth learning approach for detecting Cassini spacecraft trajectory modifications in post-processing mode. The proposed model utilizes the ability of Long Short Term Memory (LSTM) neural networks for drawing out useful data and learning the time series inner data pattern, along with the forcefulness of LSTM layers for distinguishing dependencies among the long-short term. Our research study exploited the statistical rates, Matthews correlation coefficient, and F1 score to evaluate our models. We carried out multiple tests and evaluated the provided approach against several advanced models. The preparatory analysis showed that exploiting the LSTM layer provides a notable boost in rising the detection process performance. The proposed model achieved a number of 232 trajectory modification detections with 99.98% accuracy among the last 13.35 years of the Cassini spacecraft life.