Long Short-term Memory Network Research Articles

Explaining the Mechanism of Multiscale Groundwater Drought Events: A New Perspective From Interpretable Deep Learning Model

AbstractThis study presents a new approach to understand the causes of groundwater drought events with interpretable deep learning (DL) models. As prerequisites, accurate long short‐term memory (LSTM) models for simulating groundwater are built for 16 regions representing three types of spatial scales in the southeastern United States, and standardized groundwater index is applied to identify 233 groundwater drought events. Two interpretation methods, expected gradients (EG) and additive decomposition (AD), are adopted to decipher the DL‐captured patterns and inner workings of LSTM networks. The EG results show that: (a) temperature‐related features were the primary drivers of large‐scale groundwater droughts, with their importance increasing from 56.1% to 63.1% as the drought events approached from 6 months to 15 days. Conversely, precipitation‐related features were found to be the dominant factors in the formation of groundwater drought in small‐scale catchments, with the overall importance ranging from 59.8% to 53.3%; (b) Seasonal variations in the importance of temperature‐related factors are inversely related between large and small spatial scales, being more significant in summer for larger regions and in winter for catchments; and (c) temperature‐related factors exhibited an overall “trigger effect” on causing groundwater drought events in the studying areas. The AD method unveiled how the LSTM network behaved differently in retaining and discarding information when emulating different groundwater droughts. In summary, this study provides a new perspective for the causes of groundwater drought events and highlights the potential and prospect of interpretable DL in enhancing our understanding of hydrological processes.

Polarimetric inverse scattering using LSTM-aided association-learning sparse reconstruction framework

Conventional polarimetric inverse scattering is under the assumption of joint sparsity of different polarization channels. However, in practical situation, there are strong correlations among these multi-channel measurements, but the inherent dependencies are unknown; the deterministic joint sparsity assumption may cause insufficient utilization of association information among multi-polarization channels, and consequently lead to limited improvement of canonical scatterers extraction accuracy. Aimed at this problem, we propose a long short-term memory (LSTM)-aided association-learning sparse reconstruction framework for polarimetric inverse scattering, in which the dependencies among multi-channel measurements are captured via a data driven pattern. In the proposed scheme, instead of directly solving the full inverse problem, polarimetric inverse scattering with multi-attribute canonical scatterers extraction is implemented via hierarchical pattern, and consequently the required computational burden is reduced. Particularly, the LSTM network is utilized to perform atomic screening for estimating first type of attributes of canonical scatterers, and then the optimal adaptive filtering (OAF) process is carried out for estimating second and third types of attributes, followed by residual update process. Extensive experimental results demonstrate that the proposed approach is able to outperform existing algorithms in terms of reconstruction accuracy and computational speed.

Multi-mode vehicle pose estimation under different GNSS conditions

The integrated navigation system combining the global navigation satellite system (GNSS) and inertial navigation system (INS) is a crucial method for pose estimation in the field of autonomous driving technologies. Nevertheless, the accuracy of pose estimation is severely compromised when GNSS signals are obstructed or disrupted. To address this issue, this study introduces a multi-mode pose estimation framework designed to ensure accurate pose estimation even under unstable GNSS conditions. By integrating vehicle kinematics model that considers steering characteristics (VKMSC) and the convolutional neural network-long short-term memory (CNN-LSTM) neural network (NN) model into various estimation modes, the framework enhances the robustness of the integrated navigation system against signal interference. The system dynamically selects the optimal estimation strategy based on the degree of GNSS signal disruption. The proposed method has been validated through real-vehicle experiments, which demonstrate its efficacy in providing precise pose estimation across a spectrum of interference scenarios. Under the multipath and non-line-of-sight (MP/NLOS) mode, compared to the integrated navigation system and the fusion of traditional vehicle kinematic models, the proposed method improved positional estimation accuracy by 61.8 % and 19.7 %, respectively. In GNSS outage mode, the proposed method increased the estimation accuracy by 36.5 % and 12.0 %, respectively, compared to the INS navigation system assisted by the VKMSC and CNN-LSTM network model. The proposed method effectively reduces pose estimation errors in the integrated navigation system during interference and suppresses data fluctuations, thereby enhancing the system's precision and robustness.

Hybrid machine learning model combining of CNN-LSTM-RF for time series forecasting of Solar Power Generation

Forecasting solar power generation (SPG) is vital for the development and planning of power systems, offering significant benefits in terms of technical performance and financial efficiency. It enhances system reliability, safety, stability and it reduces the operational costs. This paper's primary goal is to develop models that can precisely forecast solar power generation by analyzing real first-hand dataset of solar power. The value of these forecasting models lies in their ability to anticipate future solar power generation, thus optimizing resource use and minimizing expenses. To achieve this, the study utilizes various classical machine learning, deep learning, and hybrid machine learning techniques, including Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), Recurrent Neural Networks (RNN), Random Forest (RF), Support Vector Regression (SVR), Bi-directional LSTM (Bi-LSTM), and Convolutional Neural Network (CNN). Among these, the hybrid model combining CNN-LSTM-RF demonstrated superior accuracy with R-squared of 92 %, a Root Mean Square Error (RMSE) of 0.07 kW, and a Mean Absolute Error (MAE) of 0.05 kW. This indicates that the hybrid machine learning model combining of CNN-LSTM-RF is effective in forecasting solar power generation.

Spatial and temporal forecasting of groundwater anomalies in complex aquifer undergoing climate and land use change

Monitoring groundwater (GW) level variations, or anomalies in multiple wells, over long periods of time is essential to understanding changes in regional groundwater resource availability. However, it is challenging to predict these GW anomalies over the long term in agricultural areas due to complicated boundary conditions, heterogeneous hydrogeological characteristics, and groundwater extraction, as well as nonlinear interactions among these factors. To overcome this challenge, we developed an advanced modeling framework based on a recurrent neural network of long short-term memory (LSTM) as an alternative to complex and computationally expensive physical models. GW anomalies were forecast two months in advance (t + 2) based on the evaluation of drivers that influence GW dynamics in densely irrigated agricultural regions. An application and evaluation of this new approach was conducted in the Wisconsin Central Sands (WCS) region in the U.S., one of the most productive agricultural regions. The modeling framework was developed for the period 1958–2020 by utilizing easily accessible dynamic and static variables to represent hydrometeorological and geological characteristics. GW anomaly observations were acquired from 26 piezometers (wells) installed in the sandy aquifer in WCS over 10–60 years. The subset of GW observations from ∼ 10–60 years, not used in model training, can forecast GW anomalies two months out with a coefficient of determination R2 ∼ 0.8. Additionally, MAE was less than 0.34 m/month across the study region for both temporal and spatial modeling frameworks. Groundwater anomalies showed high spatiotemporal variability, and their responses are influenced differently by boundary conditions, catchment geology, climate, and topography across locations. Sites with higher autocorrelation with previous two-months GW anomalies reduced bias by increasing R2. Land use change and irrigation pumping have interactive effects on GW anomalies forecasting. The novelty of this study is identifying the regional drivers of GW fluxes. This case-specific information and location-related simplification, modification, and assumption of LSTM is a unique contribution to the existing literature. Our framework can be used as an alternative method of simulating water availability and groundwater level changes in areas where subsurface properties are unknown or difficult to determine.

Transfer learning-based updating method of transient stability assessment model

An updating method for data-driven based transient stability assessment model of power systems is proposed in this paper. To cope with potential faults in the real-time operation of power systems, the parameters of models can be modified online through the proposed method. Firstly, according to the long-short term memory (LSTM) network, transient stability assessment models are initially trained offline by expected faults. Then, the feature distribution differences between expected faults and potential faults are evaluated by the time-sequence maximum mean discrepancy (TMMD). Compared with the traditional maximum mean discrepancy (MMD), the proposed TMMD approach takes the temporal properties of transient stability into account, which can reflect the relationship between fault samples and time series better. Finally, the parameters of models are adjusted by transfer learning to reduce the feature distribution differences, by which the updated models can be more suitable for the different but related potential faults. Therefore, through the proposed model updating method, the extensibility of evaluation models is greatly enhanced, which is more in line with the practical conditions of power systems. In this paper, the effectiveness of the proposed method has been demonstrated by the evaluation results in the IEEE 39-bus system and a realistic power system.

Enhanced Air Quality Prediction Using a Coupled DVMD Informer-CNN-LSTM Model Optimized with Dung Beetle Algorithm.

Accurate prediction of air quality is crucial for assessing the state of the atmospheric environment, especially considering the nonlinearity, volatility, and abrupt changes in air quality data. This paper introduces an air quality index (AQI) prediction model based on the Dung Beetle Algorithm (DBO) aimed at overcoming limitations in traditional prediction models, such as inadequate access to data features, challenges in parameter setting, and accuracy constraints. The proposed model optimizes the parameters of Variational Mode Decomposition (VMD) and integrates the Informer adaptive sequential prediction model with the Convolutional Neural Network-Long Short Term Memory (CNN-LSTM). Initially, the correlation coefficient method is utilized to identify key impact features from multivariate weather and meteorological data. Subsequently, penalty factors and the number of variational modes in the VMD are optimized using DBO. The optimized parameters are utilized to develop a variationally constrained model to decompose the air quality sequence. The data are categorized based on approximate entropy, and high-frequency data are fed into the Informer model, while low-frequency data are fed into the CNN-LSTM model. The predicted values of the subsystems are then combined and reconstructed to obtain the AQI prediction results. Evaluation using actual monitoring data from Beijing demonstrates that the proposed coupling prediction model of the air quality index in this paper is superior to other parameter optimization models. The Mean Absolute Error (MAE) decreases by 13.59%, the Root-Mean-Square Error (RMSE) decreases by 7.04%, and the R-square (R2) increases by 1.39%. This model surpasses 11 other models in terms of lower error rates and enhances prediction accuracy. Compared with the mainstream swarm intelligence optimization algorithm, DBO, as an optimization algorithm, demonstrates higher computational efficiency and is closer to the actual value. The proposed coupling model provides a new method for air quality index prediction.

Video-Based Sign Language Recognition via ResNet and LSTM Network.

Sign language recognition technology can help people with hearing impairments to communicate with non-hearing-impaired people. At present, with the rapid development of society, deep learning also provides certain technical support for sign language recognition work. In sign language recognition tasks, traditional convolutional neural networks used to extract spatio-temporal features from sign language videos suffer from insufficient feature extraction, resulting in low recognition rates. Nevertheless, a large number of video-based sign language datasets require a significant amount of computing resources for training while ensuring the generalization of the network, which poses a challenge for recognition. In this paper, we present a video-based sign language recognition method based on Residual Network (ResNet) and Long Short-Term Memory (LSTM). As the number of network layers increases, the ResNet network can effectively solve the granularity explosion problem and obtain better time series features. We use the ResNet convolutional network as the backbone model. LSTM utilizes the concept of gates to control unit states and update the output feature values of sequences. ResNet extracts the sign language features. Then, the learned feature space is used as the input of the LSTM network to obtain long sequence features. It can effectively extract the spatio-temporal features in sign language videos and improve the recognition rate of sign language actions. An extensive experimental evaluation demonstrates the effectiveness and superior performance of the proposed method, with an accuracy of 85.26%, F1-score of 84.98%, and precision of 87.77% on Argentine Sign Language (LSA64).

A hybrid deep approach to recognizing student activity and monitoring health physique based on accelerometer data from smartphones

Smartphone sensors have gained considerable traction in Human Activity Recognition (HAR), drawing attention for their diverse applications. Accelerometer data monitoring holds promise in understanding students’ physical activities, fostering healthier lifestyles. This technology tracks exercise routines, sedentary behavior, and overall fitness levels, potentially encouraging better habits, preempting health issues, and bolstering students’ well-being. Traditionally, HAR involved analyzing signals linked to physical activities using handcrafted features. However, recent years have witnessed the integration of deep learning into HAR tasks, leveraging digital physiological signals from smartwatches and learning features automatically from raw sensory data. The Long Short-Term Memory (LSTM) network stands out as a potent algorithm for analyzing physiological signals, promising improved accuracy and scalability in automated signal analysis. In this article, we propose a feature analysis framework for recognizing student activity and monitoring health based on smartphone accelerometer data through an edge computing platform. Our objective is to boost HAR performance by accounting for the dynamic nature of human behavior. Nonetheless, the current LSTM network’s presetting of hidden units and initial learning rate relies on prior knowledge, potentially leading to suboptimal states. To counter this, we employ Bidirectional LSTM (BiLSTM), enhancing sequence processing models. Furthermore, Bayesian optimization aids in fine-tuning the BiLSTM model architecture. Through fivefold cross-validation on training and testing datasets, our model showcases a classification accuracy of 97.5% on the tested dataset. Moreover, edge computing offers real-time processing, reduced latency, enhanced privacy, bandwidth efficiency, offline capabilities, energy efficiency, personalization, and scalability. Extensive experimental results validate that our proposed approach surpasses state-of-the-art methodologies in recognizing human activities and monitoring health based on smartphone accelerometer data.

Non-stationary vibration fatigue life prediction of automotive components based on long short-term memory network

Automotive components are prone to fatigue failure as a result of the long-term effects of vibration loads. Due to the significant non-stationarity of irregular excitations from various road surfaces, the classical frequency-domain method struggles to accurately estimate the fatigue life of automotive components. Based on long short-term memory (LSTM) networks, an efficient time-domain method for non-stationary vibration fatigue life prediction is proposed. Firstly, the data augmentation method for simulating long-time non-stationary loads is studied. Short-time histories are transformed into time–frequency spectrograms, and then the time–frequency spectrums are warped and masked to reconstruct the long-time non-stationary loads. Furthermore, employing only short-time loads and responses as training samples, the LSTM network is trained to construct a surrogate model for calculating structural stress time histories. Finally, the responses of varying-length long-time loads are calculated, and fatigue life is predicted by the combination of rainflow counting and Miner rule. Additionally, the representative response durations required for the fatigue analysis are estimated. Numerical simulation of control arms shows that the fatigue life prediction results using the LSTM surrogate model are within 1.9% difference compared to transient dynamics analysis results based on finite element method, and the calculation efficiency is improved by orders of magnitude.

Inverse design of nanohole all-dielectric metasurface based on deep convolutional neural network

Compared with the traditional micro-nano device design method, the data-driven neural network model method can greatly improve the inverse design efficiency of micro-nano structures with tiny design errors. Micro-nano structures based on the Fano resonance effect are widely used and have various functions. In recent years, researchers have attached great importance to its inverse design. This paper addresses the application requirements of the Fano resonance effect and develops a Gramian angular field-Convolutional Neural Network-Long Short Term Memory (GAF–CNN–LSTM) hybrid model based on the Gramian angular field transition. The model takes the designed nanohole all-dielectric metasurface structure as an example. This structure can produce different types of Fano resonance effects under different parameter combinations, implement different types of Fano resonances relative to multiple structures,the proposed structure can be realized by dynamic tuning of a single structure to obtain multi-Fano resonances, thus better application to different sensors.Through the ablation experiments and the contrast experiments, the results indicate that the proposed hybrid model can achieve high-precision inverse prediction of multifunctional Fano resonance effects, with a model accuracy of up to 95%. This study provides a new perspective on the inverse design of array-type multifunctional Fano resonance effect micro-nano devices.

A model with high-precision on proton exchange membrane fuel cell performance degradation prediction based on temporal convolutional network-long short-term memory

Prediction with high-precision of the performance degradation is crucial for the development of proton exchange membrane fuel cells (PEMFCs) with long lifetime. In this work, a new hybrid neural network model on fuel cell performance degradation prediction (TCN-LSTM) is developed by incorporating the advantages of temporal convolutional network (TCN) and long short-term memory (LSTM). The innovation of TCN-LSTM is that LSTM captures the long-term correlation of data from the high-level features of the original data extracted by TCN. On the available PEMFC datasets (IEEE 2014 PHM Data Challenge), the TCN-LSTM reduces the root mean square error (RMSE) by up to 94% compared with other state-of-the-art algorithms. In addition, the TCN-LSTM is also assessed on 2673 h self-tested fuel cell voltage degradation dataset recorded under dynamic cyclic load conditions with complicated decay mechanism. The RMSE is only 0.00205 even under the 50% training length, which further verifies the excellent prediction performance of TCN-LSTM. It can be concluded that the TCN-LSTM will be a promising fuel cell performance degradation prediction model.

Time-Dependent Deep Learning Prediction of Multiple Sclerosis Disability.

The majority of deep learning models in medical image analysis concentrate on single snapshot timepoint circumstances, such as the identification of current pathology on a given image or volume. This is often in contrast to the diagnostic methodology in radiology where presumed pathologic findings are correlated to prior studies and subsequent changes over time. For multiple sclerosis (MS), the current body of literature describes various forms of lesion segmentation with few studies analyzing disability progression over time. For the purpose of longitudinal time-dependent analysis, we propose a combinatorial analysis of a video vision transformer (ViViT) benchmarked against traditional recurrent neural network of Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) architectures and a hybrid Vision Transformer-LSTM (ViT-LSTM) to predict long-term disability based upon the Extended Disability Severity Score (EDSS). The patient cohort was procured from a two-site institution with 703 patients' multisequence, contrast-enhanced MRIs of the cervical spine between the years 2002 and 2023. Following a competitive performance analysis, a VGG-16-based CNN-LSTM was compared to ViViT with an ablation analysis to determine time-dependency of the models. The VGG16-LSTM predicted trinary classification of EDSS score in 6years with 0.74 AUC versus the ViViT with 0.84 AUC (p-value < 0.001 per 5 × 2 cross-validation F-test) on an 80:20 hold-out testing split. However, the VGG16-LSTM outperformed ViViT when patients with only 2years of MRIs (n = 94) (0.75 AUC versus 0.72 AUC, respectively). Exact EDSS classification was investigated for both models using both classification and regression strategies but showed collectively worse performance. Our experimental results demonstrate the ability of time-dependent deep learning models to predict disability in MS using trinary stratification of disability, mimicking clinical practice. Further work includes external validation and subsequent observational clinical trials.

Machine learning-based prediction approach for ranging resolution enhancement of FMCW LiDAR system with LSTM networks

Ranging resolution is one of the most important factors of frequency-modulated continuous wave (FMCW) light detection and ranging (LiDAR) systems, which is determined by the effective sweep bandwidth and observation time of the ranging signal used for Fourier transform. Conventional methods for resolution enhancement either require expensive devices or complicated operation with human intervention, which is high-cost and inefficient. For the first time, a machine learning-based prediction approach is proposed for resolution improvement. A long short-term memory (LSTM) network with a special architecture, which is driven by a theoretical model of FMCW ranging signals, is established for signal prediction. A resampling-based preprocessing block is then designed to denoise the original ranging data and enhance its feature. Finally, multi-step prediction of the input time-series signal is performed using the well-trained LSTM network to increase effective data length so as to enhance the resolution. The effectiveness of the proposed method is verified through several experiments. The results show that the proposed method not only has the advantages of low-cost and high-efficiency, but also enables the FMCW LiDAR system to possess larger equivalent bandwidth, better ranging resolution, and higher signal-to-noise ratio (SNR), enlightening the development of machine learning in high-precision ranging systems.

Diagnostic and prognostic for prescriptive maintenance and control of PEMFC systems in an industrial framework

This paper designed and defined the framework of prescriptive maintenance and its four axes, for hydrogen-energy systems, considering the constraints of actual systems, to improve durability, availability, reliability, performance, safety and operating costs of these systems. Different data-driven approaches for diagnostic and different approaches for prognostic of Remaining Useful Life (RUL) for Proton Exchange Membrane Fuel Cell (PEMFC) systems are discussed. The use of measures always available on actual systems allows to validate the approach in an industrial framework. Several diagnostic algorithms (classification and clustering) are compared to find the best compromise between computation time and accuracy in an online diagnostic framework. For the prognostic part, the data set comes from experimental tests performed under steady load for a PEMFC stack. A degradation trend of PEMFC stack voltage is extracted from these data and two well-known prediction algorithms namely Bidirectional Echo State Network (BiESN) and Bidirectional Long-Short Term Memory network (BiLSTM) are tested and compared to evaluate the robustness of the PEMFC voltage prediction for RUL estimation. The diagnostic and prognostic approaches proposed in this paper are first steps towards future work related to prescriptive maintenance for hydrogen-energy systems (PEMFC, Proton Exchange Membrane Water Electrolyzer and hybridization of the two).

Dynamic prediction model of landslide displacement based on (SSA-VMD)-(CNN-BiLSTM-attention): a case study

Accurately predicting landslide displacement is essential for reducing and managing associated risks. To address the challenges of both under-decomposition and over-decomposition in landslide displacement analysis, as well as the low predictive accuracy of individual models, this paper proposes a novel prediction model based on time series theory. This model integrates a Convolutional Neural Network (CNN) with a Bidirectional Long-Short Term Memory network (BiLSTM) and an attention mechanism to form a comprehensive CNN-BiLSTM-Attention model. It harnesses the feature extraction capabilities of CNN, the bidirectional data mining strength of BiLSTM, and the focus-enhancing properties of the attention mechanism to enhance landslide displacement predictions. Furthermore, this paper proposes utilizing the Variational Mode Decomposition (VMD) method to decompose both landslide displacement and its influencing factors. The VMD algorithm’s parameters are optimized through the Sparrow Search Algorithm (SSA), which effectively minimizes the influence of subjective bias while maintaining the integrity of the decomposition. A fusion of the Maximal Information Coefficient (MIC) and Grey Relational Analysis (GRA) is then employed to identify the critical influencing factors. The selected sequence of factors that conforms to the criteria is used as the input variable for displacement prediction via the CNN-BiLSTM-Attention model. The cumulative displacement prediction is derived by aggregating the results from each sequence. The study reveals that the SSA-VMD-CNN-BiLSTM-Attention model introduced herein achieves superior predictive accuracy for both periodic and random term displacements than individual models. This advancement provides a dependable benchmark for forecasting displacement in similar landslide scenarios.

Prediction of the river water environment carrying capacity using LSTM networks

ABSTRACT River basins receive wastewater from socio-economic activities. In such a context, a comprehensive assessment and forecast of load-bearing capacity needs to be developed. This capacity depends on many complex factors, such as hydrology, hydraulics, and environment, leading to applying a modelling approach. This study aims to propose a hybrid modelling approach to evaluate and predict the river water environmental capacity (RWEC) in a basin. Big data technology with Python is applied to process modelling results and RWEC forecasts. The long short-term memory (LSTM) model predicts RWEC on each reach of the river channel network. The {RWECP, i, P = Nitrate, BOD, Phosphate, i = hourly} data set and related factors form a time series of data used for the LSTM forecasting model. Forecast results are evaluated through the root mean square error (RMSE) and mean absolute error (MAE). The results show that the average level over all 24 reaches for 7 forecast days: RMSENitrate = 22.16 (kg/day), RMSEBOD = 38.92 (kg/day), and RMSEPhosphate = 0.79 (kg/day). This is an acceptable result for a complex system and 7-day forecast. The results of the study help significantly with pollution control.

Composing jazz music pieces using LSTM neural networks approach

The domain of artificial intelligence has increasingly extended into the creative arts, aiming to emulate and augment human creativity with automated processes. This is particularly evident in the field of music, where AI's ability to learn and produce intricate compositions has attracted significant attention. This study explores the challenge of generating jazz music using artificial intelligence, specifically focusing on the application of Long Short-Term Memory (LSTM) neural networks for jazz composition. By optimizing the model to address the genre's complexity, the research demonstrates the LSTM's capacity to capture and reproduce jazz's essential harmonic progressions and rhythmic nuances. Quantitative analyses show high accuracy and a deep understanding of musical structures, whereas qualitative feedback confirms the model's efficacy in producing compositions that embody jazz's spontaneity. Despite its achievements, the model's tendency to generate repetitive sequences suggests areas for improvement. This paper advances the field of AI in music, illustrating the potential of LSTM networks to mimic complex musical genres and emphasizing the necessity of ongoing model refinement to foster creativity. It highlights the evolving role of machine learning in music generation, proposing a foundation for future work aimed at diminishing the gap between AI capabilities and artistic expression.

Long short-term memory (LSTM) neural networks for predicting dynamic responses and application in piezoelectric energy harvesting

Long Short-Time Memory (LSTM) deep neural networks are capable of learning order dependence in sequence problems and capturing long-term, non-linear temporal dependencies between the input and out of a system. With the long-term vision to model dynamical systems to which analytical or numerical methods are impossible or difficult to apply, this paper presents a study of modeling system dynamics and predicting responses using the LSTM networks, which have demonstrated excellent capability in predicting single-mode responses in a prior study. However, the LSTM network exhibits difficulties in modeling and predicting multi-mode responses accurately. To resolve the multi-mode issue, this paper presents an approach that obtains an equivalent network consisting of a set of sub-networks learned on isolated modes, and demonstrates its effectiveness on a simulated 2-degree-of-freedom mass-spring-damper system of nonlinear Duffing springs. The second part of the paper is focused on the application of the proposed approach in piezoelectric energy harvesting. Experiments are conducted on a harvester subjected to random base-motion excitation and exhibiting nonlinearity in its multi-mode response. Both the direct and mode-separation LSTM modeling approaches are applied to predict the output voltage given a random base-motion excitation. The mode-separation approach outperforms the direct approach significantly, and yields an excellent match between the actual and predicted responses. Specifically, for a test electrical voltage response of RMS value 0.2241 V, the difference between the actual test and predicted responses by using the mode-separation approach has an RMS value of 0.0504 V, compared to 0.1645 V obtained by using the direct LSTM approach. It is also much lower than the RMS value of 0.1835 V obtained by using the attention-based LSTM network, another comparison method. Leveraging a deep learning strategy, the validated approach opens up opportunities for accurately modeling energy harvesting systems of high complexities and/or strong nonlinearities.

AI-Enabled Portable E-Nose Regression Predicting Harmful Molecules in a Gas Mixture.

The biomimetic electronic nose (e-nose) technology is a novel technology used for the identification and monitoring of complex gas molecules, and it is gaining significance in this field. However, due to the complexity and multiplicity of gas mixtures, the accuracy of electronic noses in predicting gas concentrations using traditional regression algorithms is not ideal. This paper presents a solution to the difficulty by introducing a fusion network model that utilizes a transformer-based multikernel feature fusion (TMKFF) module combined with a 1DCNN_LSTM network to enhance the accuracy of regression prediction for gas mixture concentrations using a portable electronic nose. The experimental findings demonstrate that the regression prediction performance of the fusion network is significantly superior to that of single models such as convolutional neural network (CNN) and long short-term memory (LSTM). The present study demonstrates the efficacy of our fusion network model in accurately predicting the concentrations of multiple target gases, such as SO2, NO2, and CO, in a gas mixture. Specifically, our algorithm exhibits substantial benefits in enhancing the prediction performance of low-concentration SO2 gas, which is a noteworthy achievement. The determination coefficient (R2) values of 93, 98, and 99% correspondingly demonstrate that the model is very capable of explaining the variation in the concentration of the target gases. The root-mean-square errors (RMSE) are 0.0760, 0.0711, and 3.3825, respectively, while the mean absolute errors (MAE) are 0.0507, 0.0549, and 2.5874, respectively. These results indicate that the model has relatively small prediction errors. The method we have developed holds significant potential for practical applications in detecting atmospheric pollution detection and other molecular detection areas in complex environments.