Collapse Modes Research Articles

Stochastic reconstruction of geological reservoir models based on a concurrent multi-stage U-Net generative adversarial network

For complex geological reservoir modeling, some numerical-simulation-based methods, such as traditional multiple-point statistics (MPS), cannot extract nonstationary patterns of training images (TIs) effectively. CPU-intensive calculations require that variables information only be stored in RAM instead of other storage mediums to avoid huge time consumption if many simulations are performed successively. Generative Adversarial Network (GAN) modeling methods are able to overcome these issues, but when the training datasets are limited, mode collapse easily occurs. Meanwhile, fixed receptive fields might cause the problem that the corresponding reconstruction cannot take local features and global features of the TIs into account simultaneously. Therefore, we propose the Concurrent Multi-Stage U-Net GAN (Con-MSUGAN) based on a single TI. The pyramid structure is used for spatial pattern multi-scale representation in Con-MSUGAN, which can capture the corresponding feature information of TIs under different scales. Meanwhile, concurrent training patterns can make network parameters in adjacent stages correlated with each other to accelerate network convergence speed, thereby realizing stochastic multi-scale reconstructions. Stationary channel and nonstationary delta facies TIs were used to verify the performance of Con-MSUGAN. Results show that simulation of different datasets using this method are more similar in terms of spatial variance, connectivity degree and facies type frequency distribution to the original TIs. Meanwhile, muti-dimensional-scaling (MDS) plots show small discrepancies between simulation results and the corresponding TI. It indicates that Con-MSUGAN can overcome the problem that complex spatial patterns are hard to reproduce. In addition, saved network parameters can be reused. A variety of equiprobability simulation results can be obtained rapidly, thereby realizing stochastic reconstruction of geological reservoir models.

Tessellating the Latent Space for Non-Adversarial Generative Auto-Encoders.

Non-adversarial generative models are relatively easy to train and have less mode collapse than adversarial models. However, they are not very accurate in approximating the target distribution in latent space because they don't have a discriminator. To this end, we develop a novel divide-and-conquer model called Tessellated Wasserstein Auto-Encoders (TWAE) which has less statistical error in approximating the target distribution. TWAE tessellates the support of the target distribution into a given number of regions using the centroidal Voronoi tessellation (CVT) technique and designs data batches according to the tessellation instead of random shuffling for accurate computation of discrepancy. Theoretically, we demonstrate that the error in estimating the discrepancy decreases as the number of samples n and the regions m of the tessellation increase at rates of [Formula: see text] and [Formula: see text], respectively. TWAE is very flexible to different non-adversarial metrics and can significantly enhance their generative performance in terms of Fréchet inception distance (FID) compared to existing ones. Furthermore, numerical results demonstrate that TWAE is competitive to the adversarial model and shows powerful generative ability.

Wasserstein GAN-based architecture to generate collaborative filtering synthetic datasets

Currently, generative applications are reshaping different fields, such as art, computer vision, speech processing, and natural language. The computer science personalization area is increasingly relevant since large companies such as Spotify, Netflix, TripAdvisor, Amazon, and Google use recommender systems. Then, it is rational to expect that generative learning will increasingly be used to improve current recommender systems. In this paper, a method is proposed to generate synthetic recommender system datasets that can be used to test the recommendation performance and accuracy of a company on different simulated scenarios, such as large increases in their dataset sizes, number of users, or number of items. Specifically, an improvement in the state-of-the-art method is proposed by applying the Wasserstein concept to the generative adversarial network for recommender systems (GANRS) seminal method to generate synthetic datasets. The results show that our proposed method reduces the mode collapse, increases the sizes of the synthetic datasets, improves their ratings distributions, and maintains the potential to choose the desired number of users, number of items, and starting size of the dataset. Both the baseline GANRS and the proposed Wasserstein-based WGANRS deep learning architectures generate fake profiles from dense, short, and continuous embeddings in the latent space instead of the sparse, large, and discrete raw samples that previous GAN models used as a source. To enable reproducibility, the Python and Keras codes are provided in open repositories along with the synthetic datasets generated to test the proposed architecture (https://github.com/jesusbobadilla/ganrs.git).Graphical abstract

A survey on training challenges in generative adversarial networks for biomedical image analysis

In biomedical image analysis, the applicability of deep learning methods is directly impacted by the quantity of image data available. This is due to deep learning models requiring large image datasets to provide high-level performance. Generative Adversarial Networks (GANs) have been widely utilized to address data limitations through the generation of synthetic biomedical images. GANs consist of two models. The generator, a model that learns how to produce synthetic images based on the feedback it receives. The discriminator, a model that classifies an image as synthetic or real and provides feedback to the generator. Throughout the training process, a GAN can experience several technical challenges that impede the generation of suitable synthetic imagery. First, the mode collapse problem whereby the generator either produces an identical image or produces a uniform image from distinct input features. Second, the non-convergence problem whereby the gradient descent optimizer fails to reach a Nash equilibrium. Thirdly, the vanishing gradient problem whereby unstable training behavior occurs due to the discriminator achieving optimal classification performance resulting in no meaningful feedback being provided to the generator. These problems result in the production of synthetic imagery that is blurry, unrealistic, and less diverse. To date, there has been no survey article outlining the impact of these technical challenges in the context of the biomedical imagery domain. This work presents a review and taxonomy based on solutions to the training problems of GANs in the biomedical imaging domain. This survey highlights important challenges and outlines future research directions about the training of GANs in the domain of biomedical imagery.

A Variational Autoencoder Cascade Generative Adversarial Network for Scalable 3D Object Generation and Reconstruction

Generative Adversarial Networks (GANs) for 3D volume generation and reconstruction, such as shape generation, visualization, automated design, real-time simulation, and research applications, are receiving increased amounts of attention in various fields. However, challenges such as limited training data, high computational costs, and mode collapse issues persist. We propose combining a Variational Autoencoder (VAE) and a GAN to uncover enhanced 3D structures and introduce a stable and scalable progressive growth approach for generating and reconstructing intricate voxel-based 3D shapes. The cascade-structured network involves a generator and discriminator, starting with small voxel sizes and incrementally adding layers, while subsequently supervising the discriminator with ground-truth labels in each newly added layer to model a broader voxel space. Our method enhances the convergence speed and improves the quality of the generated 3D models through stable growth, thereby facilitating an accurate representation of intricate voxel-level details. Through comparative experiments with existing methods, we demonstrate the effectiveness of our approach in evaluating voxel quality, variations, and diversity. The generated models exhibit improved accuracy in 3D evaluation metrics and visual quality, making them valuable across various fields, including virtual reality, the metaverse, and gaming.

Variation-aware semantic image synthesis

Semantic image synthesis (SIS) aims to produce photorealistic images aligning to given conditional semantic layout and has witnessed a significant improvement in recent years. Although the diversity in image-level has been discussed heavily, class-level mode collapse widely exists in current algorithms. Therefore, we declare a new requirement for SIS to achieve more photorealistic images, variation-aware, which consists of inter- and intra-class variation. The inter-class variation is the diversity between different semantic classes while the intra-class variation stresses the diversity inside one class. Through analysis, we find that current algorithms elusively embrace the inter-class variation but the intra-class variation is still not enough. Further, we introduce two simple methods to achieve variation-aware semantic image synthesis (VASIS) with a higher intra-class variation, semantic noise and position code. We combine our method with several state-of-the-art algorithms and the experimental result shows that our models generate more natural images and achieve slightly better FIDs and/or mIoUs than the counterparts. Our codes and models will be publicly available.

DiffTAD: Denoising diffusion probabilistic models for vehicle trajectory anomaly detection

Vehicle trajectory anomaly detection plays an essential role in the fields of traffic video surveillance, autonomous driving navigation, and taxi fraud detection. Deep generative models have been shown to be promising solutions for anomaly detection, avoiding the costs involved in manual labeling. However, existing popular generative models such as Generative Adversarial Networks (GANs) and Variational AutoEncoders (VAEs) are often plagued by training instability, mode collapse, and poor sample quality. To resolve the dilemma, we present DiffTAD, a novel vehicle trajectory anomaly detection framework based on the emerging diffusion models. DiffTAD formalizes anomaly detection as a noisy-to-normal process that progressively adds noise to the vehicle trajectory until the path is corrupted to pure Gaussian noise. The core idea of our framework is to devise deep neural networks to learn the reverse of the diffusion process and to detect anomalies by comparing the difference between a query trajectory and its reconstruction. DiffTAD is a parameterized Markov chain trained with variational inference and allows the mean square error to optimize the reweighted variational lower bound. In addition, DiffTAD integrates decoupled Transformer-based temporal and spatial encoders to model the temporal dependencies and spatial interactions among vehicles in the diffusion models. Experiments on the real-world trajectory dataset TRAFFIC demonstrate that our DiffTAD achieves significant improvements over existing state-of-the-art methods, with the maximum enhancements reaching 25.87% and 35.59% in terms of AUC and F1. While on the synthetic datasets CROSS, SynTra, and MAAD, the maximum improvements in AUC/F1 are 27.47%/38.56%, 25.38%/31.42%, and 58.22%/50.04%, respectively.

High imbalance fault diagnosis of aviation hydraulic pump based on data augmentation via local wavelet similarity fusion

Data augmentation methods, such as SMOTE and deep generative networks, have the potential to solve class-imbalance issue in aviation hydraulic pump fault diagnosis. Nevertheless, the scarcity of faulty samples often leads augmented samples generated using such methods suffer from blurred category boundaries, limited diversity, and mode collapse. To avoid these issues, a new data augmentation method namely Local Wavelet Similarity Fusion (LWSF) is proposed, which augments faulty samples by preserving and distorting the wavelet packet coefficients (WPCs) of the original faulty samples. An important property is that LWSF can generate high-quality synthesis samples without complex model training. First, faulty samples are decomposed into a series of frequency bands using wavelet packet decomposition. Second, the amplitudes of WPCs at a few randomly selected frequency bands are distorted, which is accomplished through a similarity-weighted fusion of the target WPCs with the closest related WPCs matched from reference samples, with the goal of keeping distortion within an appropriate range. Third, these WPCs are used to reconstruct the samples, which serve as the augmented samples. Besides, to enhance diversity, normal samples are also employed to assist in the distortion. Finally, the effectiveness of LWSF is validated by a series of experimental comparisons. In an imbalanced fault diagnosis task, the accuracy and F1-score obtained by LWSF are 97.16 % and 97.35 %, respectively, which are 3.59 % and 7.30 % higher than the best one among the five compared methods.

Magnetoactive asymmetric mechanical metamaterial for tunable elastic cloaking

We propose a magnetic field-induced asymmetric mechanical metamaterial for tunable transformation-based elastic cloaking. The metamaterial is designed by integrating hard-Magnetic Active Elastomers (hMAEs) and combining zero-stress collapse modes to produce a unique asymmetric behavior controlled by an external magnetic field. The relationship bridging the microstructure and the desired cloaking performance is presented. This magneto-metamaterial design is applied to achieve tunable static elastic cloaking. The theoretical predictions, together with the numerical tests under various static loads, demonstrate encouraging cloaking performance. The study also highlights the impact of magneto-mechanical coupling and offers the first remotely-controllable hMAE-based cloaking solution, with promising potential in various applications including stress shielding and stealth technologies.

Design and Fabrication of 1-D CMUT Arrays for Dual-Mode Dual-Frequency Acoustic Angiography Applications.

When microbubble contrast agents are excited at low frequencies (less than 5 MHz), they resonate and produce higher-order harmonics due to their nonlinear behavior. We propose a novel scheme with a capacitive micromachined ultrasonic transducer (CMUT) array to receive high-frequency microbubble harmonics in collapse mode and to transmit a low-frequency high-pressure pulse by releasing the CMUT plate from collapse and pull it back to collapse again in the same transmit-receive cycle. By patterning and etching the substrate to create glass spacers in the device cavity we can reliably operate the CMUT in collapse mode and receive high-frequency signals. Previously, we demonstrated a single-element CMUT with spacers operating in the described fashion. In this article, we present the design and fabrication of a dual-mode, dual-frequency 1-D CMUT array with 256 elements. We present two different insulating glass spacer designs in rectangular cells for the collapse mode. For the device with torus-shaped spacers, the 3 dB receive bandwidth is from 8 to 17 MHz, and the transmitted maximum peak-to-peak pressure from 32 elements at 4 mm focal depth was 2.12 MPa with a 1.21 MPa peak negative pressure, which corresponds to a mechanical index (MI) of 0.58 at 4.3 MHz. For the device with line-shaped spacers, the 3-dB receive bandwidth at 150 V dc bias extends from 10.9 to 19.2 MHz. By increasing the bias voltage to 180 V, the 3 dB bandwidth shifts, and extends from 11.7 to 20.4 MHz. The transmitting maximum peak-to-peak pressure with 32 elements at 4 mm was 2.06 MPa with a peak negative pressure of 1.19 MPa, which corresponds to an MI of 0.62 at 3.7 MHz.

Distribution Fitting for Combating Mode Collapse in Generative Adversarial Networks.

Mode collapse is a significant unsolved issue of generative adversarial networks (GANs). In this work, we examine the causes of mode collapse from a novel perspective. Due to the nonuniform sampling in the training process, some subdistributions may be missed when sampling data. As a result, even when the generated distribution differs from the real one, the GAN objective can still achieve the minimum. To address the issue, we propose a global distribution fitting (GDF) method with a penalty term to confine the generated data distribution. When the generated distribution differs from the real one, GDF will make the objective harder to reach the minimal value, while the original global minimum is not changed. To deal with the circumstance when the overall real data is unreachable, we also propose a local distribution fitting (LDF) method. Experiments on several benchmarks demonstrate the effectiveness and competitive performance of GDF and LDF.

Experimental investigations on the collapse of spiral-stiffened cylindrical shells under external hydrostatic pressure

This paper presents a comprehensive experimental investigation into the collapse behavior of spiral stiffened cylindrical shells under external hydrostatic pressure. Spiral stiffened cylindrical shells with two different pitches and cylindrical shells were designed and fabricated. The fabricated specimens were divided into three groups. Each group has two identical specimens. Test specimens were welded from stainless steel. The shell was formed by cold bending, and the spiral ribs were welded to the surface of the shell to form a spiral stiffened cylindrical shell. Measurements of initial geometric imperfections and shell thickness were performed on the fabricated specimens to verify the accuracy of fabricating. Subsequently, the specimens were placed in a pressure chamber and pressurized until collapsed to obtain the ultimate collapse pressure and collapse modes. In addition, a uniaxial tensile test was conducted on the shell material to obtain its material properties. Experimental results indicated that the fabricated specimens had excellent repeatability and reasonable precision. The presence of a spiral rib effectively boosted the ultimate collapse pressure of the cylindrical shell with a more pronounced strengthening effect observed as the rib density increased.

Artistic Essence of Generative Adversarial Networks: Analyzing Training Data’s Impact on Performance

Generative adversarial networks (GANs) are powerful deep learning models for synthesizing realistic data. However, their performance critically depends on curating optimal training data. This research conducts a comprehensive study analyzing the impact of sample size, class balance, and heterogeneity in training datasets on GAN image and text generation quality. Through extensive experiments on CIFAR-10, it has been demonstrated that insufficient samples, imbalanced classes, and lack of diversity cause degraded sample quality, coherence, and mode collapse. The analysis conducted in this research work provides unique insights into data-GAN interplay. Models trained on balanced subsets with adequate samples per class produce superior Inception Scores and BLEU, avoiding limited variety in outputs. The techniques presented enable developing more generalizable and creative GANs. This work proves to be the first of its kind to rigorously evaluate the role of data characteristics like size, balance and heterogeneity in stabilizing GAN training and improving output fdelity across modalities. The data-centric findings would be valuable for researchers to curate optimal datasets that can unlock GANs’ full potential for diverse, realistic generation with wide applications in graphics, vision, language and beyond.

Regression generative adversarial network based on bounded losses for prediction of free calcium oxide in cement clinker

The data imbalance problem caused by multi-time scales phenomenon affects the prediction accuracy, validity and robustness of free calcium oxide (fCaO) content in cement clinker calcination process. Focusing on this problem, we propose an regression generative adversarial network model to predict fCaO content, which contains a generator, discriminator and predictor. Generator and discriminator are designed as bounded loss generative adversarial network to solve mode collapse problem and improve the stability. They learn actual data features in adversarial learning style and produce fake data to enlarge the scale and feature space of actual data to train predictor and finally achieve the prediction of fCaO content, which overcomes the problem of data imbalance. For performance assessment, we visually evaluate the validity of generated data from the perspective of univariate distribution and multivariate joint distribution and invent sequence change tendency consistency index (TCI) to evaluate the robustness of fCaO content prediction. Experiments implemented by cement production data demonstrate that the proposed model has advantages in accuracy, availability and robustness in fCaO content prediction, especially TCI is higher 22.69 percentage points than that of without data augmentation.

Detection and Classification of Respiratory Syndromes in Original and modified DCGAN Augmented Neonatal Infrared Datasets

In the current pandemic scenarios, a non-invasive method for determining a neonate's respiratory rate and categorizing them using a deep learning technique is highly pertinent. Acquiring an infrared neonatal dataset for detecting and classifying respiratory syndromes is challenging. The limited number of infrared videos and images representing different types of syndromes is a tremendous challenge to the accuracy of the deep learning model. This paper uses the Deep Convolutional Generative Adversarial Networks(DCGAN) with gradient penalty for the data augmentation. The Discriminator in a standard DCGAN architecture is a convolutional neural network (CNN) that receives an image as input and outputs a single scalar value that indicates the likelihood that the input image is real or fake. Adding a gradient penalty adds a regularisation term to the loss function. This modification helps to stabilize training by preventing mode collapse and generating higher-quality images. The augmented dataset helped to make the original imbalanced dataset more balanced and increased the size of the original dataset. When the accuracies of the deep learning models trained on the original and balanced augmented neonatal datasets were compared in this work, the model based on the balanced augmented dataset performed better.

Feedback linearization control for uncertain nonlinear systems via generative adversarial networks

This article presents a novel approach to leverage generative adversarial networks(GANs) techniques to learn a feedback linearization controller(FLC) for a class of uncertain nonlinear systems. By estimating uncertainty through the adversarial process, where ground truth samples are exclusively obtained from a predefined integral model, the feedback linearization controller, learned through a minimax two-player optimization framework, enhances the reference tracking performance of the input-output uncertain nonlinear system. Furthermore, we provide theoretical guarantee of convergence and stability, demonstrating the safe recovery of robust FLC. We also address the common challenge of mode collapse in GANs training through the strict convexity of our synthesized generator structure and an enhanced adversarial loss. Comprehensive simulations and practical experiments are conducted to underscore the superiority and efficacy of our proposed approach.

Experimental, numerical and analytical analysis of a single-span steel portal frame exposed to fire

The thermo-mechanical behaviour of single-span steel portal frames in fire conditions illustrates a sequential transition from an ambient temperature state through to eventual collapse, encompassing axial forces, moments and deformations. To investigate the thermo-mechanical behaviour, a fire test on an 8 m × 6 m steel portal frame assembly is conducted. Detailed thermal and structural responses of the frame are measured and presented. The experimental results reveal that the frame collapses post 31-min heating with a peak temperature approximating 800 °C. A typical inward collapse mode is observed. 2D and 3D finite element models are established and validated against experimental results, respectively. The numerical model is divided into three heating regions to mimic the behaviour exhibited during the fire scenario. Furthermore, a thermo-mechanical theory is proposed to analyse the fire behaviour of a single-span steel portal frame. Based on the fire response of a restrained steel beam, the thermo-mechanical behaviour can be segmented into four stages: the ambient temperature stage, thermal expansion stage, elasto-plastic stage and catenary stage. A comprehensive series of calculation formulas determine the displacement, internal forces and moments at different stages. The analytical computations exhibit a close consistence with the numerical results, within the deviation of 10% for the majority. This analytical method can be effortlessly employed for both the fire-resistant design and assessment of the current state of steel portal frames in fire conditions.

Li-Ion Cell in Operando Monitoring and Prognostication Using CMUT Devices

Capacitive micromachined ultrasonic transducer (CMUT) devices are ideal solutions for in operando battery monitoring applications especially for consumer electronics. Their minute size (2mm x 1mm x 0.5mm) allows direct placement on the battery without affecting the form factor of the device, the volume production techniques used for their manufacture allow for a very low-cost device, and they can be easily integrated with front-end integrated circuits either in a single device or in a discrete multichip form. Together with industry partners we have developed CMUT devices that we use to continuously monitor the behavior of a battery while it is cycling.For this work we used CMUTs (close-up shown in Fig. 1(a)) developed to operate with a low bias voltage (less than 30 V) which reduces the complexity of the driving hardware. The impulse response and its frequency content are shown in Fig. 1(b). In collapse mode, the peak frequency is 4.0 MHz and the one-way –3 dB BW is 76%, which allows flexibility in setting the drive conditions to interrogate the battery. Our testing also shows that the CMUTs show good device-to-device uniformity and they have a very stable performance in variable temperature environments (e.g. 0℃ - 60℃ ).The cell is interrogated in pulse-echo mode. The CMUT is excited by a wide-band unipolar pulse and the returned analog signal is filtered and amplified. The signal is then digitized and processed to extract features that allow us to quantify the state of the battery. Figure 1(c) depicts such signal features change under baseline cell cycling conditions (three cycles at a charge rate of C/4). These changes are the result of mechanical material property changes in the cell e.g., electrode stiffness and density [1] which directly affect the acoustic wave propagation inside the cell. Figure 1(d) depicts signal feature change due cell overcharge (red line). Ultrasound is especially sensitive to gas generation which occurs as the cell voltage extends beyond normal operating range [2]. Signal features provide early warning notice that cell is outside of normal operating conditions.Figure 1. a) Close-up image of CMUT device. b) CMUT impulse response and corresponding frequency content. c) Signal features (RMS and Time of Flight (ToF)) and thickness change during C/4 charge and discharge. d) Signal features during cell constant current overcharge. Cell overcharged after 4.4V (red line). Chang, W. et al. Measuring effective stiffness of Li-ion batteries via acoustic signal processing. Mater. Chem. A 8, 16624–16635 (2020).Appleberry, M. C. et al. Avoiding thermal runaway in lithium-ion batteries using ultrasound detection of early failure mechanisms. Journal of Power Sources 535, 231423 (2022). Figure 1

Multi-Disciplinary and Multi-Objective Optimization Method Based on Machine Learning

The optimization of aircraft is a typical multidisciplinary and multi-objective problem. To solve this problem, the difficulty lies not only in the high cost of discipline performance evaluation but also in the complex coupling relationship between different disciplines. To improve the optimization efficiency, a new optimization method is proposed, including two new algorithms: conditional generative adversarial nets with vector similarity (VS-CGAN) and distributed single-step deep reinforcement learning with transfer learning (TL-DSDRL). For low-cost disciplines, VS-CGAN learns the relationship between variables and objectives through presampling to compress the variable domains. The cosine function is used to evaluate the similarity between the random noise and generated variables to avoid mode collapse. For high-cost disciplines, TL-DSDRL improves optimization efficiency through pretraining. The newly designed reward function and multi-agent cooperation mechanism enhance the multi-objective search ability of reinforcement learning.

Auto-Encoding Generative Adversarial Networks towards Mode Collapse Reduction and Feature Representation Enhancement.

Generative Adversarial Nets (GANs) are a kind of transformative deep learning framework that has been frequently applied to a large variety of applications related to the processing of images, video, speech, and text. However, GANs still suffer from drawbacks such as mode collapse and training instability. To address these challenges, this paper proposes an Auto-Encoding GAN, which is composed of a set of generators, a discriminator, an encoder, and a decoder. The set of generators is responsible for learning diverse modes, and the discriminator is used to distinguish between real samples and generated ones. The encoder maps generated and real samples to the embedding space to encode distinguishable features, and the decoder determines from which generator the generated samples come and from which mode the real samples come. They are jointly optimized in training to enhance the feature representation. Moreover, a clustering algorithm is employed to perceive the distribution of real and generated samples, and an algorithm for cluster center matching is accordingly constructed to maintain the consistency of the distribution, thus preventing multiple generators from covering a certain mode. Extensive experiments are conducted on two classes of datasets, and the results visually and quantitatively demonstrate the preferable capability of the proposed model for reducing mode collapse and enhancing feature representation.