Convolutional Autoencoder Network Research Articles

Local phase-constrained convolutional autoencoder network for identifying multivariate geochemical anomalies

Autoencoder is a powerful tool for identifying multivariate geochemical anomalies. However, existing autoencoder-based geochemical anomaly detection methods primarily rely on a global reconstruction error (e.g., mean square error) to define the lower limit of geochemical anomalies, neglecting the common, local structure information of geochemical data. This limitation inevitably results in the decreased accuracy of geochemical anomaly identification. This study proposed a local Phase-Constrained Convolutional AutoEncoder network (PC-CAE) for the identification of multivariate geochemical anomalies. Initially, we employed a local Fourier transform to extract phase information from both the original and the reconstructed data. Subsequently, a convolutional autoencoder network was utilized to learn the latent representation of geochemical background, using the local phase difference between the original and reconstructed data to preserve the local data structure related to geology setting. Additionally, an adaptive weighting strategy was employed to mitigate the overfitting issue. The training samples with high reconstruction errors were finally identified as anomalies. We tested the validity of PC-CAE using the stream sediment geochemical dataset collected in the Jiaodong gold province, Eastern China. The results demonstrated that PC-CAE outperforms existing convolutional autoencoder network and spectrum–area multifractal model in identifying multivariate geochemical anomalies associated with Au mineralization.

CAEM-GBDT: a cancer subtype identifying method using multi-omics data and convolutional autoencoder network

The identification of cancer subtypes plays a very important role in the field of medicine. Accurate identification of cancer subtypes is helpful for both cancer treatment and prognosis Currently, most methods for cancer subtype identification are based on single-omics data, such as gene expression data. However, multi-omics data can show various characteristics about cancer, which also can improve the accuracy of cancer subtype identification. Therefore, how to extract features from multi-omics data for cancer subtype identification is the main challenge currently faced by researchers. In this paper, we propose a cancer subtype identification method named CAEM-GBDT, which takes gene expression data, miRNA expression data, and DNA methylation data as input, and adopts convolutional autoencoder network to identify cancer subtypes. Through a convolutional encoder layer, the method performs feature extraction on the input data. Within the convolutional encoder layer, a convolutional self-attention module is embedded to recognize higher-level representations of the multi-omics data. The extracted high-level representations from the convolutional encoder are then concatenated with the input to the decoder. The GBDT (Gradient Boosting Decision Tree) is utilized for cancer subtype identification. In the experiments, we compare CAEM-GBDT with existing cancer subtype identifying methods. Experimental results demonstrate that the proposed CAEM-GBDT outperforms other methods. The source code is available from GitHub at https://github.com/gxh-1/CAEM-GBDT.git.

A Coarse and Fine-Grained Deep Multi View Subspace Clustering Method for Unsupervised Fault Diagnosis of Rolling Bearings

Abstract To address the issue of insufficient characterization of fault features in inherent vibration data that affects the performance of unsupervised learning-based fault diagnosis, a coarse and fine-grained deep multi view subspace clustering method (CFG-DMVSC) for unsupervised fault diagnosis of rolling bearings is proposed. The proposed method designs a convolutional autoencoder network based on the Gramian angular field transformation for multi-signal analysis domains. A multi-view coarse-grained self-expressive method based on information entropy is designed to handle differences in information across different views. Furthermore, a fine-grained common and independent information separation loss function based on mutual information is proposed to ensure compactness among multiple views. Both the Case Western Reserve University rolling bearing dataset and privately built bearing fault test bench data demonstrate that, compared to existing methods, the proposed method can perform coarse and fine-grained division in multi-view subspaces, achieving better clustering diagnosis performance on the extracted common information among views.

Classification of compressed full-waveform airborne lidar data

ABSTRACT Airborne laser scanners produce 3D data that can be used for a range of applications, such as urban planning, facility monitoring, flood mapping, and forest management. Additional information on the surveyed area can be obtained from the backscattered waveforms recorded by modern light detection and ranging (lidar) sensors. However, the high-dimensional representation of full-waveform data has hindered progress in its use due to difficulties in processing and storage. This paper develops a quantized convolutional autoencoder network to compress lidar waveform data into a condensed feature representation, resulting in a compression rate of up to 20:1. This, together with height information, is fed into a U-net convolutional neural network that achieves an accuracy of 93.7% on six classes.

Revised clusters of annotated unknown sounds in the Belgian part of the North sea

Acoustic signals, especially those of biological source, remain unexplored in the Belgian part of the North Sea (BPNS). The BPNS, although dominated by anthrophony (sounds from human activities), is expected to be acoustically diverse given the presence of biodiverse sandbanks, gravel beds and artificial hard structures. Under the framework of the LifeWatch Broadband Acoustic Network, sound data have been collected since the spring of 2020. These recordings, encompassing both biophony, geophony and anthrophony, have been listened to and annotated for unknown, acoustically salient sounds. To obtain the acoustic features of these annotations, we used two existing automatic feature extractions: the Animal Vocalization Encoder based on Self-Supervision (AVES) and a convolutional autoencoder network (CAE) retrained on the data from this study. An unsupervised density-based clustering algorithm (HDBSCAN) was applied to predict clusters. We coded a grid search function to reduce the dimensionality of the feature sets and to adjust the hyperparameters of HDBSCAN. We searched the hyperparameter space for the most optimized combination of parameter values based on two selected clustering evaluation measures: the homogeneity and the density-based clustering validation (DBCV) scores. Although both feature sets produced meaningful clusters, AVES feature sets resulted in more solid, homogeneous clusters with relatively lower intra-cluster distances, appearing to be more advantageous for the purpose and dataset of this study. The 26 final clusters we obtained were revised by a bioacoustics expert. We were able to name and describe 10 unique sounds, but only clusters named as ‘Jackhammer’ and ‘Tick’ can be interpreted as biological with certainty. Although unsupervised clustering is conventional in ecological research, we highlight its practical use in revising clusters of annotated unknown sounds. The revised clusters we detailed in this study already define a few groups of distinct and recurring sounds that could serve as a preliminary component of a valid annotated training dataset potentially feeding supervised machine learning and classifier models.

Single-Pixel Imaging Based on Deep Learning Enhanced Singular Value Decomposition.

We propose and demonstrate a single-pixel imaging method based on deep learning network enhanced singular value decomposition. The theoretical framework and the experimental implementation are elaborated and compared with the conventional methods based on Hadamard patterns or deep convolutional autoencoder network. Simulation and experimental results show that the proposed approach is capable of reconstructing images with better quality especially under a low sampling ratio down to 3.12%, or with fewer measurements or shorter acquisition time if the image quality is given. We further demonstrate that it has better anti-noise performance by introducing noises in the SPI systems, and we show that it has better generalizability by applying the systems to targets outside the training dataset. We expect that the developed method will find potential applications based on single-pixel imaging beyond the visible regime.

Temperature scaling unmixing framework based on convolutional autoencoder

Hyperspectral unmixing is a key technology in the development of remote sensing applications. However, since both endmembers and abundances are unknown, unmixing is a non-convex problem with a large solution space. To solve this, existing methods usually impose the same strength of sparsity constraint. However, this often does not hold in practice. Because the abundances of purer regions are generally sparse, while the abundances distribution of more mixed regions should be smoother. Temperature scaling is a technique of introducing a temperature parameter T into softmax activation function to adjust the sparsity of the output. Inspired by this, we propose a temperature scaling unmixing (TSU) framework based on convolutional autoencoder (CAE). In this framework, sparse constraints of different intensities are applied to diverse regions by considering spatial similarity of ground objects distribution while preserving the ability of CAE to extract spatial features. What is more, equal-frequency binning is adopted to guide the division of regions by similarity matrix to realize the automatic temperature parameter setting. In addition, a CAE network is designed under the TSU framework in this paper, called TSUCAE. The TSUCAE method exhibits superior accuracy compared to state-of-the-art approaches, as demonstrated through extensive comparative experiments. Furthermore, the TSU framework can be transferred to other CAE-based unmixing methods directly while keeping the network structure of these methods unchanged. Sufficient ablation experiments also prove that the transfer of framework can improve the performance of unmixing. The code is publicly available at https://github.com/UPCGIT/TSUCAE.

Deep learning-based prediction of saturation magnetic flux density in Fe-based metallic glasses via transfer learning

Transfer learning (TL) is a widely recognized machine learning model that leverages domain knowledge from one or more predictive tasks to enhance performance in a related task. This study demonstrated the improved prediction performance of saturation magnetic flux density (Bs) of bulk metallic glasses (BMGs) with deep learning through the utilization of TL, which enabled the extension of compositional knowledge to overcome the limitations posed by scarce data. Specifically, TL was demonstrated by training a convolutional autoencoder network on full BMG datasets and transferring the acquired compositional knowledge to scarce iron (Fe)-based metallic glass datasets. Through this transfer, the broad compositional knowledge derived from the full BMG dataset is embedded in data representations, allowing its domain knowledge to be shared with smaller datasets. Here, a convolutional neural network was used as a predictor, with the prediction performance of Bs evaluated by the determination coefficient (R2 score) and root-mean-squared error (RMSE). The predictor, having received knowledge about the broad composition through TL, exhibited enhanced predictive performance of Bs on unseen BMG datasets while maintaining its prediction capability for Fe-based metallic glass alloys with R2 = 0.939 and RMSE = 0.084 T. This performance surpassed the performance of directly transferred knowledge on scarce data. These findings underscore the benefits of TL in predictive tasks with limited datasets and highlight the potential for applying this method to the development of new materials, where well-trained models can be repurposed for unlabeled material datasets with desired properties.

Generative approaches for solving tangram puzzles

The Tangram is a dissection puzzle composed of seven polygonal pieces that can form different patterns. Solving the Tangram is an irregular shape packing problem known to be NP-hard. This paper investigates the application of four deep-learning architectures—Convolutional Autoencoder, Variational Autoencoder, U-Net, and Generative Adversarial Network—specifically designed for solving Tangram puzzles. We explore the potential of these architectures in learning the complex spatial relationships inherent in Tangram configurations. Our experiments show that the Generative Adversarial Network competes well with other architectures and converges considerably faster. We further prove that traditional evaluation metrics based on pixel accuracy often fail in assessing the visual quality of the generated Tangram solutions. We introduce a loss function based on a Weighted Mean Absolute Error that prioritizes pixels representing inter-piece sections over those covered by individual pieces. Extending this loss function, we propose a novel evaluation metric as a more fitting measure for assessing Tangram solutions compared to traditional metrics. This investigation advances our understanding of the capabilities of artificial intelligence in complex geometrical problem domains.

Vector Quantized Convolutional Autoencoder Network for LDCT Image Reconstruction with Hybrid Loss

Vector Quantized Convolutional Autoencoder Network for LDCT Image Reconstruction with Hybrid Loss

Fault state identification of rolling bearings based on deep transfer convolutional autoencoder and a new health indicator

For the problem of bearing fault state identification, in recent years, the main research is how to distinguish the normal state and the fault state. However, it is still difficult to accurately identify the more detailed fault severity in the cross-working environment. Therefore, this paper proposes a new health indicator (HI) and deep transferable convolutional autoencoder network (DTCAE) to solve the problem of rolling bearing early fault monitoring and fault state recognition. Firstly, the envelope spectrum of the vibration signal is calculated, and then the ratio of the energy at the maximum amplitude of the envelope spectrum to other energies is obtained, and the final HI is obtained by weighting the standard deviation with it, which has good generalization and engineering practicability. HI is used for early bearing fault monitoring and classifying bearing fault status in source domain. Finally, aiming at the problem that the existing transfer learning methods do not consider the weakening of the target domain data features after domain adaptation, DTCAE is proposed to identify the bearing fault state. Specifically, in addition to the distribution difference loss and classification loss commonly seen in transfer learning models, a decoder is also added after feature extraction to calculate the difference between the reconstructed data of the target domain and the original data, and the feature distribution difference between the reconstructed source domain and the target domain is also calculated. This allows the source-domain bearing and the target-domain bearing features to be matched in such a way that both not only have the same feature distribution, but also retain the degradation trend characteristics of the target-domain bearing. Experiments in three challenging cases have achieved good results and are superior to some existing methods.

Feature-enhanced X-ray imaging using fused neural network strategy with designable metasurface

Abstract Scintillation-based X-ray imaging can provide convenient visual observation of absorption contrast by standard digital cameras, which is critical in a variety of science and engineering disciplines. More efficient scintillators and electronic postprocessing derived from neural networks are usually used to improve the quality of obtained images from the perspective of optical imaging and machine vision, respectively. Here, we propose to overcome the intrinsic separation of optical transmission process and electronic calculation process, integrating the imaging and postprocessing into one fused optical–electronic convolutional autoencoder network by affixing a designable optical convolutional metasurface to the scintillator. In this way, the convolutional autoencoder was directly connected to down-conversion process, and the optical information loss and training cost can be decreased simultaneously. We demonstrate that feature-specific enhancement of incoherent images is realized, which can apply to multi-class samples without additional data precollection. Hard X-ray experimental validations reveal the enhancement of textural features and regional features achieved by adjusting the optical metasurface, indicating a signal-to-noise ratio improvement of up to 11.2 dB. We anticipate that our framework will advance the fundamental understanding of X-ray imaging and prove to be useful for number recognition and bioimaging applications.

Prediction of Contact Fatigue Performance Degradation Trends Based on Multi-Domain Features and Temporal Convolutional Networks.

Contact fatigue is one of the most common failure forms of typical basic components such as bearings and gears. Accurate prediction of contact fatigue performance degradation trends in components is conducive to the scientific formulation of maintenance strategies and health management of equipment, which is of great significance for industrial production. In this paper, to realize the performance degradation trend prediction accurately, a prediction method based on multi-domain features and temporal convolutional networks (TCNs) was proposed. Firstly, a multi-domain and high-dimensional feature set of vibration signals was constructed, and performance degradation indexes with good sensitivity and strong trends were initially screened using comprehensive evaluation indexes. Secondly, the kernel principal component analysis (KPCA) method was used to eliminate redundant information among multi-domain features and construct health indexes (HIs) based on a convolutional autoencoder (CAE) network. Then, the performance degradation trend prediction model based on TCN was constructed, and the degradation trend prediction for the monitored object was realized using direct multi-step prediction. On this basis, the effectiveness of the proposed method was verified using a bearing common-use data set, and it was successfully applied to performance degradation trend prediction for rolling contact fatigue specimens. The results show that using KPCA can reduce the feature set from 14 dimensions to 4 dimensions and retain 98.33% of the information in the original preferred feature set. The method of constructing the HI based on CAE is effective, and change processes versus time of the constructed HI can truly reflect the degradation process of rolling contact fatigue specimen performance; this method has obvious advantages over the two commonly used methods for constructing HIs including auto-encoding (AE) networks and gaussian mixture models (GMMs). The model based on TCN can accurately predict the performance degradation of rolling contact fatigue specimens. Compared with prediction models based on long short-term memory (LSTM) networks and gating recurrent units (GRUs), the model based on TCN has better performance and higher prediction accuracy. The RMS error and average absolute error for a prediction step of 3 are 0.0146 and 0.0105, respectively. Overall, the proposed method has universal significance and can be applied to predict the performance degradation trend of other mechanical equipment/parts.

A novel transfer gear life prediction method by the cross-condition health indicator and nested hierarchical binary-valued network

As one of the crucial techniques in prognostics and health management, the long-term remaining useful life (RUL) prediction of gears is essential for maintenance decisions in engineering applications. However, most previous RUL prediction approaches may have unacceptable prediction performance due to insufficient prior life-cycle data and variable operating conditions. In this paper, a novel transfer life prediction method is proposed for gear RUL estimation under different working conditions. First, a health indicator (HI) transfer construction framework, QFMDCAET, is developed for generating cross-condition HIs under various working conditions based on a quadratic function-based multi-scale deep convolutional auto-encoder network and multi-kernel maximum mean discrepancy. With the acquired gear HIs, the nested hierarchical binary-valued network (NHBN) is designed to estimate the gear RUL. In NHBN, a Gumbel function is applied to improve the memory ability of NHBN. Moreover, a nested hierarchical mechanism is also used to fully utilize the sequence information of the cell state in NHBN. The experimental results illustrate the effectiveness of the proposed NHBN in gear RUL prediction, especially for long-term RUL prediction. Lastly, the comparison results indicate that the proposed NHBN-based RUL prediction method is superior to other conventional and advanced RUL prediction methods.

Time-feature attention-based convolutional auto-encoder for flight feature extraction

Quick Access Recorders (QARs) provide an important data source for Flight Operation Quality Assurance (FOQA) and flight safety. It is generally characterized by large volume, high-dimensionality and high frequency, and these features result in extreme complexities and uncertainties in its usage and comprehension. In this study, we proposed a Time-Feature Attention (TFA)-based Convolutional Auto-Encoder (TFA-CAE) network model to extract essential flight features from QAR data. As a case study, we used the QAR data landing at the Kunming Changshui International Airport and Lhasa Gonggar International Airport as the experimental data. The results show that (1) the TFA-CAE model performs the best in extracting representative flight features in comparison to some traditional or similar approaches, such as Principal Component Analysis (PCA), Convolutional Auto-Encoder (CAE), Self-Attention-based CAE (SA-CAE), Gate Recurrent Unit based Auto-Encoder (GRU-AE) and TFA-GRU-AE models; (2) flight patterns corresponding to different runways can be recognized; and (3) anomalous flights can effectively deviate from many observations. Overall, the TFA-CAE model provides a well-established technique for further usage of QAR data, such as flight risk detection or FOQA.

Radar Active Jamming Recognition under Open World Setting

To address the issue that conventional methods cannot recognize unknown patterns of radar jamming, this study adopts the idea of zero-shot learning (ZSL) and proposes an open world recognition method, RCAE-OWR, based on residual convolutional autoencoders, which can implement the classification of known and unknown patterns. In the supervised training phase, a residual convolutional autoencoder network structure is first constructed to extract the semantic information from a training set consisting solely of known jamming patterns. By incorporating center loss and reconstruction loss into the softmax loss function, a joint loss function is constructed to minimize the intra-class distance and maximize the inter-class distance in the jamming features. Moving to the unsupervised classification phase, a test set containing both known and unknown patterns is fed into the trained encoder, and a distance-based recognition method is utilized to classify the jamming signals. The results demonstrate that the proposed model not only achieves sufficient learning and representation of known jamming patterns but also effectively identifies and classifies unknown jamming signals. When the jamming-to-noise ratio (JNR) exceeds 10 dB, the recognition rate for seven known jamming patterns and two unknown jamming patterns is more than 92%.

Deep audio embeddings for vocalisation clustering.

The study of non-human animals' communication systems generally relies on the transcription of vocal sequences using a finite set of discrete units. This set is referred to as a vocal repertoire, which is specific to a species or a sub-group of a species. When conducted by human experts, the formal description of vocal repertoires can be laborious and/or biased. This motivates computerised assistance for this procedure, for which machine learning algorithms represent a good opportunity. Unsupervised clustering algorithms are suited for grouping close points together, provided a relevant representation. This paper therefore studies a new method for encoding vocalisations, allowing for automatic clustering to alleviate vocal repertoire characterisation. Borrowing from deep representation learning, we use a convolutional auto-encoder network to learn an abstract representation of vocalisations. We report on the quality of the learnt representation, as well as of state of the art methods, by quantifying their agreement with expert labelled vocalisation types from 8 datasets of other studies across 6 species (birds and marine mammals). With this benchmark, we demonstrate that using auto-encoders improves the relevance of vocalisation representation which serves repertoire characterisation using a very limited number of settings. We also publish a Python package for the bioacoustic community to train their own vocalisation auto-encoders or use a pretrained encoder to browse vocal repertoires and ease unit wise annotation.

End-to-end sound field reproduction based on deep learning.

Sound field reproduction, which attempts to create a virtual acoustic environment, is a fundamental technology in the achievement of virtual reality. In sound field reproduction, the driving signals of the loudspeakers are calculated by considering the signals collected by the microphones and working environment of the reproduction system. In this paper, an end-to-end reproduction method based on deep learning is proposed. The inputs and outputs of this system are the sound-pressure signals recorded by microphones and the driving signals of loudspeakers, respectively. A convolutional autoencoder network with skip connections in the frequency domain is used. Furthermore, sparse layers are applied to capture the sparse features of the sound field. Simulation results show that the reproduction errors of the proposed method are lower than those generated by the conventional pressure matching and least absolute shrinkage and selection operator methods, especially at high frequencies. Experiments were performed under conditions of single and multiple primary sources. The results in both cases demonstrate that the proposed method achieves better high-frequency performance than the conventional methods.

Model order reduction for parameterized electromagnetic problems using matrix decomposition and deep neural networks

A non-intrusive model order reduction (MOR) method for solving parameterized electromagnetic scattering problems is proposed in this paper. A database collecting snapshots of high-fidelity solutions is built by solving the parameterized time-domain Maxwell equations for some values of the material parameters using a fullwave solver based on a high order discontinuous Galerkin time-domain (DGTD) method. To perform a prior dimensionality reduction, a set of reduced basis (RB) functions are extracted from the database via a two-step proper orthogonal decomposition (POD) method. Intrinsic coordinates of the high-fidelity solutions are further compressed through a convolutional autoencoder (CAE) network. Singular value decomposition (SVD) is then used to extract the principal components of the low dimensional coding matrices generated by CAE, and a cubic spline interpolation-based (CSI) approach is employed for approximating the dominating time- and parameter-modes of these matrices. The generation of the reduced basis and the training of the CAE and CSI are accomplished in the offline stage, thus the RB solution for given time/parameter values can be quickly recovered via outputs of the interpolation model and decoder network. In particular, the offline and online stages of the proposed RB method are completely decoupled, which ensures the validity of the method. The performance of the proposed CAE-CSI ROM is illustrated with numerical experiments for scattering of a plane wave by a 2-D dielectric disk and a multi-layer heterogeneous medium.

Airfoil-based convolutional autoencoder and long short-term memory neural network for predicting coherent structures evolution around an airfoil

Airfoil stall induced by unsteady flow separation is a major inhibitor that constrains the aircraft design. Hence, it is critical to have a better understanding of the flow dynamics around the airfoil near stall conditions. Recently, machine learning (ML) has been widely used to predict the spatiotemporal evolutions of coherent structures over the airfoil. Nonetheless, due to the existence of strong nonlinearity introduced by turbulence, most of the existing approaches focus on prediction of the flow fields in the laminar flow regime. In this work, we seek to examine the applicability of deep neural network (DNN) to prediction of turbulent flows around a NACA (National Advisory Committee for Aeronautics) 4412 airfoil. Large eddy simulations (LES) are performed to prepare the training dataset. Both proper orthogonal decomposition (POD) and DNN-based reduced-order models (ROM) are employed for reconstructing the flow fields. A convolutional autoencoder (CAE) is applied to map the high-dimensional flow dynamics into the low-dimensional nonlinear manifold. The geometric features are considered in the input matrix of the CAE network by setting physical quantities inside the airfoil to zero to reflect the physics. Then, the multi-head attention-strengthened long short-term memory (MH-LSTM) is utilized to predict the evolutions of latent vectors. Finally, the high-dimensional flow dynamics can be reconstructed by utilizing the decoder part of CAE. We believe that the MH-LSTM methodology provides a promising path towards predicting the unsteady flow field over the airfoil by capitalizing on the capability of multi-head attention allowing for attending to distinct parts of input sequence.