Clean Features Research Articles

Data-Driven and Sustainable Transportation Route Optimization in Green Logistics Supply Chain

In the green logistics supply chain, transportation route optimization faces urgent problems such as high energy consumption and environmental pollution. This paper aims to achieve sustainable optimization of transportation routes through a data-driven approach. First, a large amount of transportation-related data is collected, including vehicle operation information, cargo characteristics, road conditions and meteorological factors. Big data analysis technology is used to clean and extract features from the data, and a transportation demand forecasting model is constructed. Then, a new optimization model is designed using the particle swarm optimization algorithm, with the goal of minimizing transportation costs and carbon dioxide emissions. In actual application cases, by optimizing the transportation routes of a logistics company, the results showed that the lowest transportation cost was 1,512 dollars and the lowest carbon dioxide emissions were 1.2 tons. Data-driven transportation route optimization not only improves logistics efficiency, but also promotes environmental sustainable development.

Towards generalized face forgery detection with domain-robust representation learning

Face forgery detection is crucial for the security of digital identities. However, existing methods often struggle to generalize effectively to unseen domains due to the domain shift between training and testing data. We propose a Domain-robust Representation Learning (DRRL) method for generalized face forgery detection. Specifically, we observe that domain shifts in face forgery detection tasks are often caused by forgery differences and content differences between domain data, while the limitations of training data lead the model to overfit to these feature expressions in the seen domain. Therefore, DRRL enhances the model's generalization to unseen domains by first adding representative data representations to mitigate overfitting to seen data and then removing the features of expressed domain information to learn a robust, discriminative representation of domain variation. Data augmentation is achieved by stylizing sample representations and exploring representative new styles to generate rich data variants, with the Content-style Augmentation (CSA) module and Forgery-style Augmentation (FSA) module implemented for content and forgery expression, respectively. Based on this, the Content Decorrelation (CTD) module and Sensitive Channels Drop (SCD) module are used to remove content features irrelevant to forgery and domain-sensitive forgery features, encouraging the model to focus on clean and robust forgery features, thereby achieving the goal of learning domain-robust representations. Extensive experiments on five large-scale datasets demonstrate that our method exhibits advanced and stable generalization performance in practical scenarios.

The Prediction and Evaluation of Surface Quality during the Milling of Blade-Root Grooves Based on a Long Short-Term Memory Network and Signal Fusion.

The surface quality of milled blade-root grooves in industrial turbine blades significantly influences their mechanical properties. The surface texture reveals the interaction between the tool and the workpiece during the machining process, which plays a key role in determining the surface quality. In addition, there is a significant correlation between acoustic vibration signals and surface texture features. However, current research on surface quality is still relatively limited, and most considers only a single signal. In this paper, 160 sets of industrial field data were collected by multiple sensors to study the surface quality of a blade-root groove. A surface texture feature prediction method based on acoustic vibration signal fusion is proposed to evaluate the surface quality. Fast Fourier transform (FFT) is used to process the signal, and the clean and smooth features are extracted by combining wavelet denoising and multivariate smoothing denoising. At the same time, based on the gray-level co-occurrence matrix, the surface texture image features of different angles of the blade-root groove are extracted to describe the texture features. The fused acoustic vibration signal features are input, and the texture features are output to establish a texture feature prediction model. After predicting the texture features, the surface quality is evaluated by setting a threshold value. The threshold is selected based on all sample data, and the final judgment accuracy is 90%.

Blood pressure prediction based on multi-sensor information fusion of electrocardiogram, photoplethysmography, and pressure pulse waveform

Cardiovascular diseases are now the leading cause of death that endangers people’s health. Thus, a precise and reliable blood pressure (BP) prediction method is essential. This paper proposes a noninvasive BP prediction method with the multi-feature fusion of electrocardiogram (ECG), photoplethysmography (PPG), and pressure pulse waveform (PPW). A multi-sensor information acquisition platform was developed to collect cardiovascular-related signals. Besides, the algorithms were designed to clean, preprocess, and extract features from sample data. Furthermore, the BP prediction model was constructed by using feature selection and feature fusion based on Random Forest Regression (RFR). Finally, the importance of features used for blood pressure prediction was analyzed, and the results of RFR-based blood pressure prediction were compared with those of other machine learning algorithms. The mean absolute errors of systolic and diastolic blood pressure prediction reached 0.90 mmHg and 2.47 mmHg, respectively. The results of the BP prediction model based on multi-sensor information fusion meet both the AAMI and BHS standards.

Natural-like generation of grain boundary models and the combined effects of microstructural elements and lithiation on the plastic behavior of TiO2: A computational study

Microstructural features of ceramics have a significant effect on mechanical properties and ion transport in different technologies ranging from optical materials to renewable energy generation and storage. Atomistic modeling is usually done on postulated grain boundaries which may not account for the distributions of different grain boundaries in real materials. We generate microstructural features of TiO2 with molecular dynamics (MD) by simulating a heat treatment; in this way, a natural-like distribution of grain sizes and boundaries is obtained. We used a model system with about 545,000 atoms and obtained structures with average grain sizes ranging about 6 to 12 nm and trending with the annealing temperature. While the model is nano- rather than micro-meter structured, it present key microstructural features. We observe ‘clean’ grain boundaries as well as boundaries with an amorphous interlayer. The clean boundaries feature low-index surfaces. When an amorphous interlayer is present, higher-index surfaces can be observed. The resulting structures show significant softening (reduction of maximum stress) vs. ideal crystalline titania, with isotropy of mechanical properties. Elastic constants, besides significant isotropy, show moderate stiffening due to the presence of grains. When introducing lithium, we observe segregation at grain boundaries, and lithiation of grainy structures leads to further softening.

A comparative study of LSTM-ED architectures in forecasting day-ahead solar photovoltaic energy using Weather Data

Solar photovoltaic (PV) energy, with its clean, local, and renewable features, is an effective complement to traditional energy sources today. However, the photovoltaic power system is highly weather-dependent and therefore has unstable and intermittent characteristics. Despite the negative impact of these features on solar sources, the increase in worldwide installed PV capacity has made solar energy prediction an important research topic. This study compares three encoder-decoder (ED) networks for day-ahead solar PV energy prediction: Long Short-Term Memory ED (LSTM-ED), Convolutional LSTM ED (Conv-LSTM-ED), and Convolutional Neural Network and LSTM ED (CNN-LSTM-ED). The models are tested using 1741-day-long datasets from 26 PV panels in Istanbul, Turkey, considering both power and energy output of the panels and meteorological features. The results show that the Conv-LSTM-ED with 50 iterations is the most successful model, achieving an average prediction score of up to 0.88 over R-square (R2). Evaluation of the iteration counts’ effect reveals that the Conv-LSTM-ED with 50 iterations also yields the lowest Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) values, confirming its success. In addition, the fitness and effectiveness of the models are evaluated, with the Conv-LSTM-ED achieving the lowest Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) values for each iteration. The findings of this work can help researchers build the best data-driven methods for forecasting PV solar energy based on PV features and meteorological features.

Feature decoupling and interaction network for defending against adversarial examples

Recently, it was found that deep neural networks (DNNs) are susceptible to adversarial input perturbations. Most defense strategies adopt the denoising method based on preprocessing, which mitigates the impacts of adversarial perturbations on DNNs by learning the distributions of nonadversarial datasets and projecting adversarial inputs into the learned nonadversarial manifolds. However, existing defense strategies commonly focus on reconstructing clean images while ignoring the role of adversarial perturbations, which results in the reconstructed images failing to achieve the visual quality and classification accuracy of the original clean images, and the induced adversarial robustness improvement is limited. This paper proposes a feature decoupling-interaction network (FDIN), which introduces the concepts of clean features and adversarial features to separate the two kinds of features from the input adversarial examples (AEs) in a feature decoupling-interaction manner. The clean features are used to reconstruct the input image so that it is infinitely close to the original clean image, and the adversarial features are used to reconstruct the adversarial perturbations. Adversarial perturbations are removed from the adversarial examples across multiple cross cycles to improve further the reconstructed image's visual quality and classification accuracy. The features of the original clean image are used as prior knowledge to guide the network to learn the clean features of the adversarial examples and improve the classification accuracy of the model on the clean examples. In addition, a classification loss function based on the Carlini & Wagner (CW) attack algorithm is used instead of the conventional cross-entropy loss function to improve the adversarial robustness of the FDIN. The experimental results show that the proposed method achieves better defense performance than the current state-of-the-art methods on both standard tests and various attack tests and even exceeds the test accuracy of the target classifier on the original test set.

GLAD: Global–Local Approach; Disentanglement Learning for Financial Market Prediction

Accurate prediction of financial market trends can have a great impact on maximizing profits and avoiding risks. Conventional methods, e.g., regression or SVR, or end-to-end training approaches, coined as deep learning algorithms, have restraints as a consequence of capturing noisy and unnecessary data. Financial market’s data are composed of stock’s price time series that are correlated, and each time series has both global and local dynamics. Inspired by recent advancements in disentanglement representation learning, in this paper, we present a promising model for predicting financial markets that learn disentangled representations of features and eliminate those features that cause interference. Our model uses the informer encoder to extract features, capturing global–local patterns by using the time and frequency domains, augmenting the clean features with time and frequency-based features, and using the decoder to predict. To be more specific, we adopt contrastive learning in the time and frequency domains to learn both global and local patterns. We argue that our methodology, disentangling and learning the influential factors, holds the potential for more accurate predictions and a better understanding of how time series move and behave. We conducted our experiments using the S&P 500, CSI 300, Hang Seng, and Nikkei 225 stock market datasets to predict their next-day closing prices. The results showed that our model outperformed existing methods in terms of prediction error (mean squared error and mean absolute error), financial risk measurement (volatility and max drawdown), and prediction net curves, which means that it may enhance traders’ profits.

Objformer: Boosting 3D object detection via instance-wise interaction

Deep learning on point clouds drives 3D object detection. Despite rapid progress, point-based methods still suffer from the problems such as incompletion and occlusion, which are caused by the material properties of objects and cluttered scenes. These difficult targets increase the difficulty of identification or even lead to misidentification, severely weakening the performance of point-based methods on 3D object detection. To alleviate the above problems, we propose the Objformer to boost point-based 3D object detection via instance-wise interaction. We design an instance feature encoder to encode clean instance features, which contain key geometric priors and holistic semantic information. Further, an instance interaction module is devised to aggregate the complementary features across instances with label-guided interaction, boosting the performance of the 3D object detection. Experiments show that Objformer outperforms previous point-based state-of-the-arts on two popular benchmarks, ScanNet V2 and SUN RGB-D. Especially, our single-modal Objformer even outperforms the competing advanced multi-modal fusion method on both SUN RGB-D and ScanNet V2.

A Decompressed Spectral-Spatial Multiscale Semantic Feature Network for Hyperspectral Image Classification

Convolutional neural networks (CNNs) have shown outstanding feature extraction capability and become a hot topic in the field of hyperspectral image (HSI) classification. However, most of the prior works usually focus on designing deeper or wider network architectures to extract spatial and spectral features, which give rise to difficulty for optimization and more parameters along with higher computation. Moreover, how to learn spatial and spectral information more effectively is still being researched. To tackle the aforementioned problems, a decompressed spectral-spatial multiscale semantic feature network (DSMSFNet) for HSI classification is proposed. This model is composed of a decompressed spectral-spatial feature extraction module (DSFEM) and a multiscale semantic feature extraction module (MSFEM). The former is devised to extract more discriminative and representative global decompressed spectral-spatial features in a lightweight extraction manner, while the latter is constructed to expand the range of available receptive fields and generate clean multiscale semantic features at a granular level to further enhance the classification performance. Compared with progressive classification approaches, abundant experimental results on three benchmark datasets prove the superiority of our developed DSMSFNet model.

Multi-Task Learning for UAV Aerial Object Detection in Foggy Weather Condition

Adverse weather conditions such as haze and snowfall can degrade the quality of captured images and affect performance of drone detection. Therefore, it is challenging to locate and identify targets in adverse weather scenarios. In this paper, a novel model called Object Detection in a Foggy Condition with YOLO (ODFC-YOLO) is proposed, which performs image dehazing and object detection jointly by multi-task learning approach. Our model consists of a detection subnet and a dehazing subnet, which can be trained end-to-end to optimize both tasks. Specifically, we propose a Cross-Stage Partial Fusion Decoder (CSP-Decoder) in the dehazing subnet to recover clean features of encoder from complex weather conditions, thereby reducing the feature discrepancy between hazy and clean images, thus enhancing the feature consistency between different tasks. Additionally, to increase the feature modeling and representation capabilities of our network, we also propose an efficient Global Context Enhanced Extraction (GCEE) module to extract beneficial information from blurred images by constructing global feature context long-range dependencies. Furthermore, we propose a Correlation-Aware Aggregated Loss (CAALoss) to average noise patterns and tune gradient magnitudes across different tasks, accordingly implicitly enhancing data diversity and alleviating representation bias. Finally, we verify the advantages of our proposed model on both synthetic and real-world foggy datasets, and our ODFC-YOLO achieves the highest mAP on all datasets while achieving 36 FPS real-time detection speed.

Scale-progressive Multi-patch Network for image dehazing

Image dehazing is a classical vision task, which aims to recover a clean image from a hazy one. Previous dehazing methods usually follow a coarse-to-fine architecture to mine clean features by introducing generic CNNs components. However, this manner usually results in undesirable model complexity and computational burden. In this work, we present a Scale-progressive Multi-patch Network (SPM-Net) to handle common problems in previous dehazing networks. Specifically, our approach utilizes a scale-progressive multi-patch mechanism to efficiently model uneven hazy distribution on local patches and progressively explore clean cues in multiple scales in a fine-to-coarse way. Besides of above, we found the feature misalignment problem in the patch-based methods, and a practical solution is proposed to handle this previously neglected problem. A comprehensive evaluation of both synthetic datasets and real-world datasets demonstrates that the proposed dehazing method surpasses the previous state-of-the-art approaches with a margin both quantitatively and qualitatively. Our proposed SPM-Net achieved a PSNR of 29.47 dB on the Haze4k dataset, significantly surpassing the previous state-of-the-art method DMT-Net (29.47 dB vs. 28.53 dB) while having much fewer parameters than DMT-Net (16.1 M vs. 54.9 M) and faster inferencing efficiency (0.072 s vs. 0.192 s).

Dropout Autoencoder Fingerprint Augmentation for Enhanced Wi-Fi FTM-RSS Indoor Localization

In this letter, we propose a dropout autoencoder fingerprint augmentation approach for enhanced Wi-Fi fine time measurement and received signal strength signals-based indoor localization. Due to complex indoor environment, fingerprinting techniques suffers from unrecorded measurements at some reference points, leading to incomplete fingerprint datasets. The dropout autoencoder was employed to reconstruct clean signal features for the unrecorded fingerprint measurement which can significantly affect the localization accuracy of fingerprinting systems. The localization is accomplished by utilizing deep neural networks (DNN)-based regression. We collected two datasets from experiments conducted in two indoor offices using commercial off-the-shelf devices. The performance of our proposed method was compared to existing methods, and on the respective datasets, our proposal method showed better performance with a localization accuracy of 0.3m and 0.6m for the 1-σ percentile errors and 0.66m and 1.5m for the 2-σ percentile errors.

Optimized deep neural network models for blood pressure classification using Fourier analysis-based time-frequency spectrogram of photoplethysmography signal.

Appropriate blood pressure (BP) management through continuous monitoring and rapid diagnosis helps to take preventive care against cardiovascular diseases (CVD). As hypertension is one of the leading causes of CVDs, keeping hypertension under control by a timely screening of subjects becomes lifesaving. This work proposes estimating BP from motion artifact-affected photoplethysmography signals (PPG) by applying signal processing techniques in realtime. This paper proposes a deep neural network-based methodology to accurately classify PPG signals using a Fourier theory-based time-frequency (TF) spectrogram. This work uses the Fourier decomposition method (FDM) to transform a PPG signal into a TF spectrogram. In the proposed work, the last three layers of the pre-trained deep neural network, namely, GoogleNet, DenseNet, and AlexNet, are modified and then used to classify the PPG signal into normotension, pre-hypertension, and hypertension. The proposed framework is trained and tested using the MIMIC-III and PPG-BP databases using five-fold training and testing. Out of the three deep neural networks, the proposed framework with the DenseNet-201 network performs best, with a test accuracy of 96.5%. The proposed work uses FDM to compute the TF spectrogram to accurately separate the motion artifacts and noise components from a noise-corrupted PPG signal. Capturing more frequency components that contain more information from PPG signals makes the deep neural networks extract better and more meaningful features. Thus, training a deep neural network model with clean PPG signal features improves the generalized capability of a BP classification model when tested in realtime.

A framework for reconstructing building parametric models with hierarchical relationships from point clouds

Parametric reconstruction based on point clouds of 3D scenes has attracted much attention due to its application prospects in many fields. This paper proposes a novel framework for reconstructing parametric models of buildings with hierarchical relationships from point clouds. Different from traditional approaches focusing on extracting geometric primitives, this work aims to combine the semantic hierarchy of building components with edge measurements based on vertex connections to output parametric models. We show a three-stage automated pipeline for reconstructing parametric models from point clouds of buildings. Regarding to the first stage, we design a deep network architecture to achieve the task of hierarchical segmentation, and propose a metric to evaluate the semantic consistency of different hierarchies. In the second stage, in order to predict the position of corners and filter the correct connections to construct the skeleton-graph of the model, we present a deep network architecture that converts point clouds into skeleton-graph model. The network takes the labeled 3D points output by the semantic hierarchical segmentation network as input, and then outputs the skeleton-graph of the point clouds, which is a set of edge segments connected by corner points. In the third stage, the semantic hierarchical segmentation information of the point clouds is embedded as attributes, and then the geometric parameters are measured according to the edges connected by vertices, and finally the semantic information and geometric features are mapped into the schema of CityGML. We pre-train and validate the reliability of our framework on a finely crafted synthetic dataset and finally we conduct transfer learning and fine-tune the framework on a real scene dataset. Experiments show that our method not only generates high-quality hierarchical parametric models but also recovers clean features and is robust to noise. This study provides practical guidance and technical references for developing more intelligent modeling algorithms that could support data-driven decision-making in smart cities.

Hydrogen Energy: Development Prospects, Current Obstacles and Policy Suggestions under China’s “Dual Carbon” Goals

Hydrogen energy has an advantage over conventional fossil energy because of its clean and low-carbon features, while its stability gives it an advantage over renewable energy sources such as hydropower, photovoltaic (PV) power, and wind power. Using hydrogen as a fuel can achieve zero carbon emissions or even net zero emissions at the end of energy conversion. Moreover, with hydrogen as the carrier, an energy system of multi-energy complementarity will be established by coupling multiple energy systems such as power grids, gas grids and heat networks, which can realize the comprehensive utilization of renewable energy sources in an efficient, stable, and flexible manner. Based on the “dual carbon” goals, this paper targets sectors with high carbon emissions represented by electric power, steel and transportation, analyzes the feasible path of using hydrogen energy to promote deep carbon reduction, and points out the obstacles faced by hydrogen energy development in various sectors. On account of the current development and expected prospects, this paper proposes that the current focus of hydrogen energy industry is to coordinate the development of industrial sectors and regions, build a pattern that can facilitate the industry’s systematic and concerted development, address key technical challenges (such as green hydrogen production or hydrogen storage and transportation via pipeline), improve industrial standards, and promote the industry’s development in a balanced, coordinated and orderly way.

Generating Time-Series Data Using Generative Adversarial Networks for Mobility Demand Prediction

The increasing penetration rate of electric kickboard vehicles has been popularized and promoted primarily because of its clean and efficient features. Electric kickboards are gradually growing in popularity in tourist and education-centric localities. In the upcoming arrival of electric kickboard vehicles, deploying a customer rental service is essential. Due to its free-floating nature, the shared electric kickboard is a common and practical means of transportation. Relocation plans for shared electric kickboards are required to increase the quality of service, and forecasting demand for their use in a specific region is crucial. Predicting demand accurately with small data is troublesome. Extensive data is necessary for training machine learning algorithms for effective prediction. Data generation is a method for expanding the amount of data that will be further accessible for training. In this work, we proposed a model that takes time-series customers’ electric kickboard demand data as input, pre-processes it, and generates synthetic data according to the original data distribution using generative adversarial networks (GAN). The electric kickboard mobility demand prediction error was reduced when we combined synthetic data with the original data. We proposed Tabular-GAN-Modified-WGAN-GP for generating synthetic data for better prediction results. We modified The Wasserstein GAN-gradient penalty (GP) with the RMSprop optimizer and then employed Spectral Normalization (SN) to improve training stability and faster convergence. Finally, we applied a regression-based blending ensemble technique that can help us to improve performance of demand prediction. We used various evaluation criteria and visual representations to compare our proposed model’s performance. Synthetic data generated by our suggested GAN model is also evaluated. The TGAN-Modified-WGAN-GP model mitigates the overfitting and mode collapse problem, and it also converges faster than previous GAN models for synthetic data creation. The presented model’s performance is compared to existing ensemble and baseline models. The experimental findings imply that combining synthetic and actual data can significantly reduce prediction error rates in the mean absolute percentage error (MAPE) of 4.476 and increase prediction accuracy.

Underwater Forward-Looking Sonar Images Target Detection via Speckle Reduction and Scene Prior

Forward-looking sonar (FLS) imagery system plays a significant role in oceanic object recognition and detection since it can overcome the limitation of lighting conditions and reflect the real situation of the underwater environment. However, object detection algorithms for FLS images remain challenging for two main reasons: 1) The noise caused by the coherent characteristic of the scattering phenomenon impairs the detector capture of target information; 2) The scene prior based on the uneven target scale distribution is generally neglected, which leads to the detector generating redundant anchors and slows down detection efficiency. Confronting such challenges, this paper characterizes the noise and the uneven target scale distribution in FLS images as multiplicative speckle noise and scene prior, respectively. Therefore, we propose a novel underwater FLS image detection network, namely UFIDNet, to further improve detection performance by considering speckle noise reduction and scene prior in FLS images. More specifically, a speckle reduction auxiliary branch (SRAB) is designed to introduce additional despeckled supervision information to encourage the feature extractor to produce clean features and share them with the detection pipeline during the training phase. In particular, the noise distribution of FLS images is excavated for synthetic dataset construction and Despeckle Network (DSN) design to obtain despeckled supervision images. In addition, a feature selection strategy (FSS) embedded in detection branch is designed to screen out feature levels that do not match the target size, thus significantly reducing the generation of redundant anchors and improving detection speed. Experimental results show that our UFIDNet achieves 70.5% and 47.3% average precision (AP), 81.3% and 54.6% average recall (AR <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">max=10</sub> ), 27.0 and 26.1 FPS on two real FLS datasets respectively, outperforming many state-of-the-art general detectors and sonar image detectors.

Exploiting exotic LHC datasets for long-lived new particle searches

Motivated by the expectation that new physics may manifest itself in the form of very heavy new particles, most of the operation time of the Large Hadron Collider (LHC) is devoted to proton-proton (pp) collisions at the highest achievable energies and collision rates. The large collision rates imply tight trigger requirements that include high thresholds on the final-state particles’ transverse momenta pT and an intrinsic background in the form of particle pileup produced by different collisions occurring during the same bunch crossing. This strategy is potentially sub-optimal for several well-motivated new physics models where new particles are not particularly heavy and can escape the online selection criteria of the multi-purpose LHC experiments due to their light mass and small coupling.A solution may be offered by complementary datasets that are routinely collected by the LHC experiments. These include heavy ion collisions, low-pileup runs for precision physics, and the so-called “parking” and “scouting” datasets. While some of them are motivated by other physics goals, they all have the usage of mild pT thresholds at the trigger-level in common. In this study, we assess the relative merits of these datasets for a representative model whose particular clean signature features long-lived resonances yielding displaced dimuon vertices. We compare the reach across those datasets for a simple analysis, simulating LHC data in Run 2 and Run 3 conditions with the Delphes simulation. We show that the scouting and parking datasets, which afford low-pT trigger thresholds by only using partial detector information and delaying the event reconstruction, respectively, have a reach comparable to the standard pp dataset with conventional thresholds. We also show that heavy ion and low-pileup datasets are far less competitive for this signature.

An Adaptive Specific Emitter Identification System for Dynamic Noise Domain

With the sharply increasing number of wireless devices, specific emitter identification (SEI) technology based on radio frequency fingerprint (RFF) is employed to enhance devices’ security for authentication. However, the problems of poor performance in a highly dynamic interference environment, high dependence on quality of data sets, and less consideration of calculation resources are urgent to be resolved in a practical SEI applying domain. Our adaptive SEI system focuses on these problems, proposing a preprocessing algorithm improving synchrosqueezed wavelet transforms by energy regularization (ISWTE) to simplify the process of signal preprocessing while improving the robustness of RFF, an unsupervised neural network noise feature extracting GAN (NEGAN) to reduce the dependence of the data sets quality while obtaining precise clean RFF features from signals with noises and an optimized structure of waveform and classifier to achieve a better performance with less complexity and little cost. To test the system performance, we have generated a real wireless device data sets with ten devices, including a training set and a base test set both with SNR 10 dB, and a dynamic test set with SNR ranges −20 to 10 dB. The proposed adaptive SEI system by our work obtains a test accuracy of 0.96 at SNR 10 dB, 0.85 at SNR 0 dB, and 0.25 at SNR −20 dB only through SNR 10-dB training set. Most of the existing methods through neither of high SNR or enhanced mixed SNRs training set failed to complete devices identification under SNR −10 dB, and test accuracy of them deceased dramatically at SNR 0 dB.