Baseline Convolutional Neural Network Research Articles

An Empirical Study on the Application of Data Augmentation Techniques to Enhance the Performance of CNN Models in Spam Detection

Spam detection is a critical task in cybersecurity, aiming to filter out unsolicited and potentially harmful communications. This study investigates the impact of various data augmentation techniques on enhancing the performance of Convolutional Neural Network (CNN) models for spam detection. Utilizing the Enron Email Dataset, we implemented several augmentation methods, including synonym replacement, random insertion, random swap, random deletion, back translation, and noise addition. Our results indicate significant performance improvements with these techniques. The baseline CNN model achieved an accuracy of 87.5%, precision of 85.2%, recall of 83.7%, and F1-score of 84.4%. The application of back translation, the most effective technique, increased accuracy to 90.3% and F1-score to 88.0%. These findings demonstrate the potential of data augmentation in improving spam detection systems, providing a robust foundation for future research. The study also highlights the importance of combining augmentation techniques and adapting them to different languages and real-world scenarios for even greater performance gains.

Vision transformer to differentiate between benign and malignant slices in 18F-FDG PET/CT

Fluorine-18-fluorodeoxyglucose (18F-FDG) positron emission tomography (PET)/computed tomography (CT) is widely used for the detection, diagnosis, and clinical decision-making in oncological diseases. However, in daily medical practice, it is often difficult to make clinical decisions because of physiological FDG uptake or cancers with poor FDG uptake. False negative clinical diagnoses of malignant lesions are critical issues that require attention. In this study, Vision Transformer (ViT) was used to automatically classify 18F-FDG PET/CT slices as benign or malignant. This retrospective study included 18F-FDG PET/CT data of 207 (143 malignant and 64 benign) patients from a medical institute to train and test our models. The ViT model achieved an area under the receiver operating characteristic curve (AUC) of 0.90 [95% CI 0.89, 0.91], which was superior to the baseline Convolutional Neural Network (CNN) models (EfficientNet, 0.87 [95% CI 0.86, 0.88], P < 0.001; DenseNet, 0.87 [95% CI 0.86, 0.88], P < 0.001). Even when FDG uptake was low, ViT produced an AUC of 0.81 [95% CI 0.77, 0.85], which was higher than that of the CNN (DenseNet, 0.65 [95% CI 0.59, 0.70], P < 0.001). We demonstrated the clinical value of ViT by showing its sensitive analysis of easy-to-miss cases of oncological diseases.

Convolution neural network hyperparameter optimization using modified particle swarm optimization

Based on the literature review, a convolutional neural network (CNN) is one of the deep learning techniques most often used for classification problems, especially image classification. Various approaches have been proposed to improve accuracy performance. In CNN architecture, parameter determination is very influential on accuracy performance. Particle swarm optimization (PSO) is a type of metaheuristic algorithm widely used for hyperparameter optimization. PSO convergence is faster than genetic algorithm (GA) and attracts many researchers for further studies such as genetic algorithms and ant colony. In PSO, determining the value of the weight parameter is very influential on accuracy. Therefore, this paper proposes CNN hyperparameter optimization using modified PSO with linearly decreasing randomized weight. The experiments use the modified National Institute of Standards and Technology (MNIST) dataset. The accuracy of the proposed method is superior, and the execution time is slower to random search. In epoch 1, epoch 3, and epoch 5, the proposed method is superior to baseline CNN, linearly decreasing weight PSO (LDWPSO), and RL-based optimization algorithm (ROA). Meanwhile, the accuracy performance of the proposed method is superior to previous studies, namely LeNet-1, LeNet-2, LeNet-3, PCANet-2, RANDNet-2, CAE1, CAE-2, and bee colony. Otherwise, lost to PSO-CNN, distributed PSO (DPSO), recurrent CNN, and CNN-PSO. However, the four methods have a complex architecture and wasteful execution time.

The Investigation of Multiple Optimization Methods on Convolutional Neural Network

In this study, the optimization of a Convolutional Neural Network (CNN) was conducted using the Fruits 360 dataset, with a specific emphasis on the impacts of Spatial Transformer Network (STN) and Stochastic Gradient Descent (SGD) optimization methods. Firstly, a baseline CNN model is built, which achieves 97.84% accuracy with a loss of 0.0999 after 50 epochs. Then, the impact of integrating STN and SGD into CNN models separately is investigated. The addition of STN slightly increased the accuracy to 97.92%, reduced the loss to 0.0994, and decreased the validation accuracy. This result suggests that while STN enhances the model's generalization ability, it may slightly reduce the maximum accuracy achievable on the validation set. After SGD optimization, the verification accuracy is increased to 98.19%, the loss is reduced to 0.0537, and the verification accuracy is increased to 98.40%. These results highlight the effectiveness of SGD in fine-tuning model parameters, resulting in more accurate models and improved generalization capabilities. A comparative analysis of these methods highlights their respective advantages. The effectiveness of the STN is rooted in its capacity to improve model generalization and mitigate overfitting, which is particularly beneficial in situations that demand robustness against varied data sets. In contrast, SGD stands out for its ability to significantly improve model accuracy and reduce loss, making it a balanced choice for comprehensive model optimization. Future research directions include exploring these optimization techniques on various datasets and investigating the potential of combining STN and SGD to achieve higher performance in CNN models.

Real-Time Medical Image Classification with ML Framework and Dedicated CNN–LSTM Architecture

In the domain of modern deep learning and classification techniques, the convolutional neural network (CNN) stands out as a highly successful and preferred method for image classification in artificial intelligence. Especially in the medical field, CNN has proven to be an ideal approach for analyzing medical data and accurately identifying diseases. Over the recent years, CNN has demonstrated significant potential and success in various computer vision tasks, with medical image classification being one of the prominent applications. In our study, we introduce a novel custom CNN model called MedvCNN, designed for classifying different types of classes. We conduct experiments with various image sizes to explore their versatility. In addition, long short-term memory (LSTM), a type of recurrent neural network (RNN), is incorporated into our approach. LSTM is specifically tailored to handle sequential data, making it ideal for time series analysis. However, its capabilities extend beyond time series data and are effectively applied to various sequential data types, including sequential vectors derived from image data. One of the key advantages of utilizing LSTM for image classification is its ability to effectively memorize and capture important features in the image data. This feature is particularly advantageous in medical image processing, where precise and accurate identification of key attributes is crucial for successful diagnosis and analysis. Furthermore, our experiments reveal that the hybrid custom LSTM model, MedvLSTM, a RNN algorithm, surpasses other methods in the domain of medical image classification. Our study places significant emphasis on attaining robust classification performance for medical image data through a sophisticated, parameter free approach, complemented by an ablation study, and comprehensive statistical analysis. This comprehensive analysis and evaluation allow us to gain a deeper understanding of the model’s effectiveness and its potential impact in the field of medical image analysis. We compare these two approaches to a baseline CNN architecture, aiming to streamline the classification process, reduce time consumption, and improve cost efficiency. Additionally, we present a real-time web-based AutoML framework along with a practical demonstration. Ultimately, our research provides a thorough investigation of the current state-of-the-art in medical image analysis accuracy, focusing on the utilization of neural networks and LSTM.

Deep Learning Approaches for Classification of Emotion Recognition based on Facial Expressions

Automated emotion recognition (AEE) plays a crucial role in numerous industries that depend on understanding human emotional responses, such as advertising, technology, and human-robot interaction, particularly within the Information Technology (IT) field. However, current systems can often come up short in comprehensively understanding an individual's emotions, as prior research has mainly focused on assessing facial expressions and categorizing them into seven primary emotions, including neutrality. In this study, we present several Deep Convolutional Neural Network (CNN) models designed specifically for the task of facial emotion recognition, utilizing the FER2013 and RAF datasets. The baseline CNN model is established through a trial-and-error method, and its results are compared with more complex deep learning techniques, including ResNet18, VGGNet16, VGGNet19, and Efficient Net-B0 models. Among these models, the VGGNet19 model achieved the best results with a test accuracy of 71.02% on the FER2013 dataset. In comparison, the ResNet18 model outperformed all other models with an 86.02% test accuracy on the RAF-DB dataset. These results underscore the potential for advancing automated emotion recognition through complex deep-learning techniques.

Graph neural network for predicting the effective properties of polycrystalline materials: A comprehensive analysis

We develop a polycrystal graph neural network (PGNN) model for predicting the effective properties of polycrystalline materials, using the Li7La3Zr2O12 ceramic as an example. A large-scale dataset with >5000 different three-dimensional polycrystalline microstructures of finite-width grain boundary is generated by Voronoi tessellation and processing of the electron backscatter diffraction images. The effective ion conductivities and elastic stiffness coefficients of these microstructures are calculated by high-throughput physics-based simulations. The optimized PGNN model achieves a low error of <1.4% in predicting all three diagonal components of the effective Li-ion conductivity matrix, outperforming a linear regression model and two baseline convolutional neural network models. Sequential forward selection method is used to quantify the relative importance of selecting individual grain (boundary) features to improving the property prediction accuracy, through which both the critical and unwanted node (edge) feature can be determined. The extrapolation performance of the trained PGNN model is also investigated. The transfer learning performance is evaluated by using the PGNN model pretrained for predicting conductivities to predict the elastic properties of the same set of microstructures.:

Efficient Generation of Pretraining Samples for Developing a Deep Learning Brain Injury Model via Transfer Learning.

The large amount of training samples required to develop a deep learning brain injury model demands enormous computational resources. Here, we study how a transformer neural network (TNN) of high accuracy can be used to efficiently generate pretraining samples for a convolutional neural network (CNN) brain injury model to reduce computational cost. The samples use synthetic impacts emulating real-world events or augmented impacts generated from limited measured impacts. First, we verify that the TNN remains highly accurate for the two impact types (N = 100 each; [Formula: see text] of 0.948-0.967 with root mean squared error, RMSE, ~ 0.01, for voxelized peak strains). The TNN-estimated samples (1000-5000 for each data type) are then used to pretrain a CNN, which is further finetuned using directly simulated training samples (250-5000). An independent measured impact dataset considered of complete capture of impact event is used to assess estimation accuracy (N = 191). We find that pretraining can significantly improve CNN accuracy via transfer learning compared to a baseline CNN without pretraining. It is most effective when the finetuning dataset is relatively small (e.g., 2000-4000 pretraining synthetic or augmented samples improves success rate from 0.72 to 0.81 with 500 finetuning samples). When finetuning samples reach 3000 or more, no obvious improvement occurs from pretraining. These results support using the TNN to rapidly generate pretraining samples to facilitate a more efficient training strategy for future deep learning brain models, by limiting the number of costly direct simulations from an alternative baseline model. This study could contribute to a wider adoption of deep learning brain injury models for large-scale predictive modeling and ultimately, enhancing safety protocols and protective equipment.

Improving deep learning acoustic classifiers with contextual information for wildlife monitoring

Bioacoustics, the exploration of animal vocalizations and natural soundscapes, has emerged as a valuable tool for studying species within their habitats, particularly those that are challenging to observe. This approach has broadened the horizons of biodiversity assessment and ecological research. However, monitoring wildlife with acoustic recorders produces large volumes of data that can be labor-intensive to analyze. Deep learning has recently transformed many computational disciplines by enabling the automated processing of large and complex datasets and has gained attention within the bioacoustics community. Despite the revolutionary impact of deep learning on acoustic detection and classification, attaining both high detection accuracy and low false positive rates in bioacoustics remains a significant challenge. An intriguing yet unexplored avenue for enhancing deep learning in bioacoustics involves the utilization of contextual information, such as time and location, to discern animal vocalizations within acoustic recordings. As a first case study, a multi-branch Convolutional Neural Network (CNN) was developed to classify 22 different bird songs using spectrograms as a first input, and spatial metadata as a secondary input. A comparison was made to a baseline model with only spectrogram input. A geographical prior neural network was trained, separately, to estimate the probability of a species occurring at a given location. The output of this network was combined with the baseline CNN. As a second case study, temporal data and spectrograms were used as input to a multi-branch CNN for the detection of Hainan gibbon (Nomascus hainanus) calls, the world’s rarest primate. Our findings demonstrate that adding metadata to the bird song classifier significantly improves classification performance, with the highest improvement achieved using the geographical prior model (F1-score of 87.78% compared to 61.02% for the baseline model). The multi-branch CNNs also proved efficient (F1-scores of 76.87% and 78.77%) and simpler to use than the geographical prior. In the second case study, our findings revealed a decrease in false positives by 63% (94% of the calls were detected) when the metadata was used by the multi-branch CNN, and an increase of 19% in gibbon detection. This study has uncovered an exciting new avenue for improving classifier performance in bioacoustics. The methodology described in this study can assist ecologists, wildlife management teams, and researchers in reducing the amount of time spent analyzing large acoustic datasets obtained from passive acoustic monitoring studies. Our approach can be adapted and applied to other calling species, and thus tailored to other use cases.

Shoe Review Sentiment Analysis Using Machine Learning and Deep Learning with Word2vec

In this modern era, internet use is continually increasing and will run with an increasing amount of existing data, such as text data. Characteristics of unstructured text are a challenge in text processing feature extraction and encourage sentiment analysis research to be carried out. The availability of a lot of text data on the internet is a challenge in sentiment analysis because it requires a complex approach. This study uses the baseline Deep Learning (DL) method, namely Long-Short Term Memory (LSTM) and Convolutional Neural Network (CNN) with word2vec. It uses Machine Learning (ML), namely Random Forest (RF), Support Vector Machine (SVM), and Gaussian Naïve Bayes (GNB) with the proposed method, word2vec, and the dataset used is the shoe review dataset which consists of 389,877 reviews. From the discussion above, to carry out sentiment analysis, a more suitable method is to use the CNN baseline method with word embedding word2vec to get an accuracy value of 91.53%. The novelty of this study is the increase in the classification accuracy value from previous studies.

An overview of ensemble and feature learning in few-shot image classification using siamese networks

Siamese Neural Networks (SNNs) constitute one of the most representative approaches for addressing Few-Shot Image Classification. These schemes comprise a set of Convolutional Neural Network (CNN) models whose weights are shared across the network, which results in fewer parameters to train and less tendency to overfit. This fact eventually leads to better convergence capabilities than standard neural models when considering scarce amounts of data. Based on a contrastive principle, the SNN scheme jointly trains these inner CNN models to map the input image data to an embedded representation that may be later exploited for the recognition process. However, in spite of their extensive use in the related literature, the representation capabilities of SNN schemes have neither been thoroughly assessed nor combined with other strategies for boosting their classification performance. Within this context, this work experimentally studies the capabilities of SNN architectures for obtaining a suitable embedded representation in scenarios with a severe data scarcity, assesses the use of train data augmentation for improving the feature learning process, introduces the use of transfer learning techniques for further exploiting the embedded representations obtained by the model, and uses test data augmentation for boosting the performance capabilities of the SNN scheme by mimicking an ensemble learning process. The results obtained with different image corpora report that the combination of the commented techniques achieves classification rates ranging from 69% to 78% with just 5 to 20 prototypes per class whereas the CNN baseline considered is unable to converge. Furthermore, upon the convergence of the baseline model with the sufficient amount of data, still the adequate use of the studied techniques improves the accuracy in figures from 4% to 9%.

SynthEye: Investigating the Impact of Synthetic Data on Artificial Intelligence-assisted Gene Diagnosis of Inherited Retinal Disease.

Rare disease diagnosis is challenging in medical image-based artificial intelligence due to a natural class imbalance in datasets, leading to biased prediction models. Inherited retinal diseases (IRDs) are a research domain that particularly faces this issue. This study investigates the applicability of synthetic data in improving artificial intelligence-enabled diagnosis of IRDs using generative adversarial networks (GANs). Diagnostic study of gene-labeled fundus autofluorescence (FAF) IRD images using deep learning. Moorfields Eye Hospital (MEH) dataset of 15692 FAF images obtained from 1800 patients with confirmed genetic diagnosis of 1 of 36 IRD genes. A StyleGAN2 model is trained on the IRD dataset to generate 512× 512 resolution images. Convolutional neural networks are trained for classification using different synthetically augmented datasets, including real IRD images plus 1800 and 3600 synthetic images, and a fully rebalanced dataset. We also perform an experiment with only synthetic data. All models are compared against a baseline convolutional neural network trained only on real data. We evaluated synthetic data quality using a Visual Turing Test conducted with 4 ophthalmologists from MEH. Synthetic and real images were compared using feature space visualization, similarity analysis to detect memorized images, and Blind/Referenceless Image Spatial Quality Evaluator (BRISQUE) score for no-reference-based quality evaluation. Convolutional neural network diagnostic performance was determined on a held-out test set using the area under the receiver operating characteristic curve (AUROC) and Cohen's Kappa (κ). An average true recognition rate of 63% and fake recognition rate of 47% was obtained from the Visual Turing Test. Thus, a considerable proportion of the synthetic images were classified as real by clinical experts. Similarity analysis showed that the synthetic images were not copies of the real images, indicating thatcopied real images, meaning the GAN was able to generalize. However, BRISQUE score analysis indicated that synthetic images were of significantly lower quality overall than real images (P < 0.05). Comparing the rebalanced model (RB) with the baseline (R), no significant change in the average AUROC and κ was found (R-AUROC=0.86[0.85-88], RB-AUROC=0.88[0.86-0.89], R-k=0.51[0.49-0.53], and RB-k=0.52[0.50-0.54]). The synthetic data trained model (S) achieved similar performance as the baseline (S-AUROC=0.86[0.85-87], S-k=0.48[0.46-0.50]). Synthetic generation of realistic IRD FAF images is feasible. Synthetic data augmentation does not deliver improvements in classification performance. However, synthetic data alone deliver a similar performance as real data, and hence may be useful as a proxy to real data. Financial Disclosure(s): Proprietary or commercial disclosure may be found after the references.

Entropy-Optimized Fault Diagnosis Based on Unsupervised Domain Adaptation

In practice, the cross-domain transfer of data distribution and the sample imbalance of fault status are inevitable, but one or both are often ignored, which restricts the adaptability and classification accuracy of the generated fault diagnosis (FD) model. Accordingly, an entropy-optimized method is proposed in this paper based on an unsupervised domain-adaptive technique to enhance FD model training. For the training, pseudosamples and labels corresponding to the target samples are generated through data augmentation and self-training strategies to diminish the distribution discrepancy between the source and target domains. Meanwhile, an adaptive conditional entropy loss function is developed to improve the data quality of the semisupervised learning, with which reliable samples are generated for the training. According to the experiment results, compared with other state-of-the-art algorithms, our method can achieve significant accuracy improvement in rolling bearing FD. Typically, the accuracy improvement compared with the baseline Convolutional Neural Network (CNN) is achieved by over 13.23%.

Local Binary Convolutional Neural Networks' Long Short-Term Memory Model for Human Embryos' Anomaly Detection

Accurate selection of embryos with the maximum implementation condition is a necessary step to increase the effectiveness of fertility treatment in in vitro fertilization (IVF). The deep learning algorithms presented high potential for monitoring and visualizing embryo features such as cell numbers and their morphological development in time series manner. Due to the ability of the computer vision and deep learning algorithms, this paper aimed to present a novel deep learning approach to distinguish simultaneous abnormality of embryos in time-lapse systems for detecting live and non-live births in IVF. The approach is composed of local binary convolutional neural network (LBCNN) and long short-term memory (LSTM). The LBCNN improved accuracy of classification by employing deep and local feature sets with lower number of learnable parameters in comparison with a standard convolutional layer. Moreover, LSTM network is employed to analyze temporal information of time-lapse embryos. The results indicate that the proposed approach achieves significant results in ROC analysis (0.98) in 5 days of monitoring compared to state-of-the-art models. In addition, the approach showed compatible results in early diagnosis of abnormality detection (72 hours) with 82.8% accuracy of classification compared to the pretrained well-known convolutional neural network (CNN) models and baseline CNN.

Predicting stock market using natural language processing

PurposePredicting the stock market's prices has always been an interesting topic since its closely related to making money. Recently, the advances in natural language processing (NLP) have opened new perspectives for solving this task. The purpose of this paper is to show a state-of-the-art natural language approach to using language in predicting the stock market.Design/methodology/approachIn this paper, the conventional statistical models for time-series prediction are implemented as a benchmark. Then, for methodological comparison, various state-of-the-art natural language models ranging from the baseline convolutional and recurrent neural network models to the most advanced transformer-based models are developed, implemented and tested.FindingsExperimental results show that there is a correlation between the textual information in the news headlines and stock price prediction. The model based on the GRU (gated recurrent unit) cell with one linear layer, which takes pairs of the historical prices and the sentiment score calculated using transformer-based models, achieved the best result.Originality/valueThis study provides an insight into how to use NLP to improve stock price prediction and shows that there is a correlation between news headlines and stock price prediction.

An Attention-Based Deep Convolutional Neural Network for Brain Tumor and Disorder Classification and Grading in Magnetic Resonance Imaging

This study proposes the integration of attention modules, feature-fusion blocks, and baseline convolutional neural networks for developing a robust multi-path network that leverages its multiple feature-extraction blocks for non-hierarchical mining of important medical image-related features. The network is evaluated using 10-fold cross-validation on large-scale magnetic resonance imaging datasets involving brain tumor classification, brain disorder classification, and dementia grading tasks. The Attention Feature Fusion VGG19 (AFF-VGG19) network demonstrates superiority against state-of-the-art networks and attains an accuracy of 0.9353 in distinguishing between three brain tumor classes, an accuracy of 0.9565 in distinguishing between Alzheimer’s and Parkinson’s diseases, and an accuracy of 0.9497 in grading cases of dementia.

Medical image annotation and classification employing pyramidal feature specific lightweight deep convolution neural network

ABSTRACT Medical image annotation has significant potential to detect multiple tags. To detect specific tags and labels, most of the conventional learning algorithms took difficulty in matching the tag with corresponding region in the medical images. Hence, most of the medical image annotations fail to reproduce discrimination between various classes. In this research work to tackle medical image discrimination, we proposed a lightweight pyramidal feature-specific learning deep learning network. The proposed pyramidal feature-specific Lightweight Convolutional Neural Network (LDCNN) architecture for classifying local visual regions to annotate the classified region. By employing pyramidal learning, the proposed LDCNN align each medical image annotation with its region. The colour conversion makes low computation complexity, since the medical image exhibits degeneration by colour absorption. Hence, the interpretability and classification effectiveness increase. To evaluate the effectiveness and classification accuracy, we compare the proposed LDCNN with AlexNet and EfficientNet on benchmark datasets like MS-COCO, LC25000 and multiclass Kather datasets. Empirical experimental performance index obtained by the proposed LDCNN outperforms baseline convolutional neural network architecture. The proposed LDCNN achieves 99.6% accuracy, 98.4% sensitivity, 97.9% specificity and 99.1% F1 score. Hence, there is a steady improvement in medical annotation efforts and classification by our proposed lightweight feature-specific learning network.

Transfer learning evaluation based on optimal convolution neural networks architecture for bearing fault diagnosis applications

Intelligent fault diagnosis utilizing deep learning algorithms is currently a topic of great interest. When developing a new Convolutional Neural Network (CNN) architecture to address machine diagnosis problem, it is common to use a deep model, with many layers, many feature maps, and large kernels. These models are capable of learning complex relationships and can potentially achieve superior performance on test data. However, not only does a large network potentially impose undue computational complexity for training and eventual deployment, it may also lead to more brittleness—where data outside of the curated dataset used in CNN training and evaluation is poorly handled. Accordingly, this paper will investigate a methodical approach for identifying a quasi-optimal CNN architecture to maximize robustness when a model is trained under one set of operating conditions, and deployed under a different set of conditions. Optuna software will be used to optimize a baseline CNN model for robustness to different rotational speeds and bearing Model #’s. To further improve the network generalization capabilities, this paper proposes the addition of white Gaussian noise to the raw vibration training data. Results indicate that the number of trainable weights and associated multiplications in the optimized model were reduced by almost 95% without jeopardizing the network classification accuracy. Additionally, moderate Additive White Gaussian Noise (AWGN) improved the model adaptation capabilities.

Apple leaf disease identification via improved CycleGAN and convolutional neural network

The identification of apple leaf diseases is crucial to reduce yield reduction and timely take disease control measures. Employing deep learning for apple leaf disease identification is challenging because of the limited availability of samples for supervised training and the serious class imbalance. Hence, this paper proposes an accurate deep learning-based pipeline to solve the problem of limited data sets on farms and reduce the partiality due to serious class imbalance. Firstly, an improved cycle-consistent adversarial networks (CycleGAN) is used to generate synthetic samples to improve the learning of data distribution and solve the problems of small data sets and class imbalance. Secondly, ResNet is trained as a baseline convolutional neural network classifier to classify apple leaf diseases. The experimental results show that ResNet has the highest recognition accuracy on the test set, reaching 97.78%, and the classification accuracy is significantly improved by the generated synthetic samples (+ 14.7%). In addition, the experiment results of t-distributed stochastic neighbor embedding (t-SNE) and visual Turing test visually confirmed that the images generated by improved CycleGAN have much better quality and are more convincing.

CNN parameter design based on fault signal analysis and its application in bearing fault diagnosis

As a representative deep learning network, Convolutional Neural Network (CNN) has been extensively used in bearing fault diagnosis and many good results have been reported. In Prognostics and Health Management (PHM) field, the CNN’s input size is usually designed as a 1D vector or 2D square matrix, and the convolution kernel size is also defined as a square shape like 3 × 3 and 5 × 5, which are directly adopted from the image recognition. Though satisfying results can be obtained, CNN with such parameter specifications is not optimal and efficient. To this end, this paper elaborated the physical characteristics of bearing acceleration signals to guide the CNN design. First, the fault period under different fault types and shaft rotation frequency were used to determine the size of CNN’s input. Next, an exponential function was involved in fitting the envelope of decaying acceleration signal during each fault period, and signal length within different decaying ratios was used to define the CNN’s kernel size. Finally, the designed CNN was validated with the Case Western Reserve University bearing dataset and Paderborn University bearing dataset. Results confirm that the physics-guided CNN (PGCNN) with rectangular input shape and rectangular convolution kernel works better than the baseline CNN with higher accuracy and smaller uncertainty. The feasibility of designing CNN parameters with physics-guided rules derived from bearing fault signal analysis has also been verified.