Deep Learning Theory Research Articles

Research on fault diagnosis method of electromechanical transmission system based on one-dimensional convolutional neural network with variable learning rate

As an important part of many mechanical equipment, the mechanical transmission system is very important to carry out efficient and accurate fault monitoring and diagnosis. Compared with traditional fault diagnosis techniques, such as spectrum analysis, deep learning has been widely used in the field of mechanical system fault diagnosis due to its powerful data expression ability, and has achieved certain research results. One-dimensional convolutional neural network is a widely used model for deep learning, so in this paper, the one-dimensional convolutional neural network (1D-CNN) in the deep learning theory and the vibration signal analysis method are integrated and applied to the fault identification of mechanical transmission system to achieve accurate diagnosis and classification of faults. The experiment is mainly to collect the vibration signal data of different fault states such as broken teeth, cracking, shaft unbalance, bearing wear, and excessive friction of the driven wheel of the mechanical transmission system, it was divided into training set and testing set according to an appropriate proportion, and 1D-CNN was built using Python. The deep learning model deeply analyzed the influence of different data sample sizes and different model parameters on the recognition accuracy, and obtained an ideal diagnostic model based on variable learning rate through parameter adjustment and comparative analysis. This experimental results show that the recognition method based on one-dimensional convolutional neural network can be effectively applied to the fault diagnosis of related mechanical transmission, and has a high diagnosis accuracy.

18F-FDG PET/CT Image Deep Learning Predicts Colon Cancer Survival.

Colon cancer is a type of cancer that begins in the large intestine. In the process of efficacy evaluation, postoperative recurrence prediction and metastasis monitoring of colon cancer, traditional medical image analysis methods are highly dependent on the personal ability of the doctors. In the process of patient treatment, it not only increases the workload and work pressure for doctors, but also has some problems with traditional medical image analysis methods. Moreover, the traditional medical image analysis methods have problems such as insufficient prediction accuracy, slow prediction speed, and the risk of errors in prediction. When analyzing 18F-FDG PET/CT images by traditional medical image analysis methods, it is easy to cause problems such as untimely treatment plans and errors in diagnosis, which will adversely affect the survival of colon cancer patients. Although 18F-FDG PET/CT images have certain advantages in image clarity and accuracy compared with traditional medical imaging methods, the analysis method based on 18F-FDG PET/CT images also has certain effects in predicting the survival of colon cancer patients, but there are still many shortcomings: the 18F-FDG PET/CT image analysis method overly relies on the technical advantages of 8F-FDG PET/CT images; in the analysis and prediction of image data, it has not gotten rid of the dependence on the personal medical quality of the doctors; traditional medical image analysis methods are still used when analyzing and predicting images; there is no breakthrough in image analysis effects. In order to solve these problems, this paper combined deep learning theory, using three algorithms of the improved RBM algorithm, image feature extraction method based on deep learning, and regression neural network to analyze and predict 18F-FDG PET/CT images, and applied some algorithms to analyze and predict 18F-FDG PET/CT images, and also established a deep learning-based 18F-FDG PET/CT image survival analysis prediction model. Four aspects survival prediction accuracy, survival prediction speed, survival prediction precision, and physician satisfaction were studied through this model. The research results have shown that compared with traditional medical image analysis methods, the prediction accuracy of 18F-FDG PET/CT image survival analysis prediction model based on deep learning is improved by 0.83%, and the prediction speed is improved by 3.42%, as well as the prediction precision increased by 6.13%. The research results show that the deep learning-based 18F-FDG PET/CT image survival analysis prediction model established in this paper is of great significance to improve the survival rate of colon cancer patients, and also promotes the development of the medical industry.

A novel one-dimensional convolutional neural network-based method for emotion recognition of electric power industry workers

When the workers of electric power industry are going to perform high-risk or dispatching operations, it is important to assess the emotional state of the workers. Emotion recognition is helpful to justify whether the works are suitable to grasp and handle the operation status of the power system in time. In this paper, we develop a simple and efficient automatic emotion recognition method based on deep learning theory using ECG signals. Firstly, since Z-score method which is quite popular for emotion recognition exists high missed outlier detection rate, we define a basic beat and removed the abnormal beats according to the similarities between the truncated beats and the basic beat. Secondly, in order to make it easily for Convolutional Neural Network (CNN) to extract the periodicity features of one-dimensional signal in time domain, we present stacking operation of the truncated segments to increase the dimension of ECG data from one to three. Finally, aiming at decreasing the computation burden, a one-dimensional CNN only consisted of two convolution blocks is designed. The results of applying the proposed approach to the benchmark Wearable Stress and Affect Data set (WESAD) demonstrate that our proposal is capable of classifying four different kinds of emotions with an averaged accuracy of 98.57% and F1 value of 98.73%, respectively. Compared against some literature, the performance of our proposal is superior to them.

Real-time rubber quality model based on CNN-LSTM deep learning theory

Predicting the quality of a rubber compound is a necessity for intelligent mixing. However, the conventional quality prediction based on the power curve has low precision and significant dispersion, making it challenging to satisfy the need for real-time rubber mixing quality inspection. In this research, the vibration signals were collected for the first time as an input feature of the mixing quality prediction model, with carbon black dispersion is utilized as a quality index. On the basis of the theory of deep learning, the online quality prediction model of mixing was constructed using a variety of featured extraction methods and neural network structures, and the models were compared and tested. After 5 experiments, the average root mean square error (RMSE) of convolutional neural network-long short-term memory network (CNN-LSTM) model is 0.1160, which is 21.46% higher than that of LSTM model after time-domain featured extraction, demonstrating the efficacy and superiority of CNN-LSTM end-to-end model. This study is essential for the progression and breakthrough of the real-time mixing quality optimization issue.

Educational Perspective: AI, Deep Learning, and Creativity

Can artificial intelligence (AI) teach and learn more creatively than humans? The article analyses deep learning theory, which follows a deterministic model of learning, since every intellectual procedure of an artificial agent is supported by concrete neural connections in an artificial neural network. Meanwhile, human creative reasoning follows a non-deterministic model. The article analyses Bayes’ theorem, in which a reasoning system makes judgments about the probability of future events based on events that have happened to it. Meillassoux’s open probability and M. A. Boden’s three types of creativity are discussed. A comparison is made between the a priori algorithm of the Turing machine and a playing child, who invents new a posteriori algorithms while playing. The Heideggerian perspective on the co-creativity of humans and thinking machines is analyzed. The authors conclude that humans have an open horizon for teaching and learning, and that makes them superior with respect to creativity in an educational perspective.

TRFM-LS: Transformer-Based Deep Learning Method for Vessel Trajectory Prediction

In the context of the rapid development of deep learning theory, predicting future motion states based on time series sequence data of ship trajectories can significantly improve the safety of the traffic environment. Considering the spatiotemporal correlation of AIS data, a trajectory time window panning and smoothing filtering method is proposed for the abnormal values existing in the trajectory data. The application of this method can effectively deal with the jump values and outliers in the trajectory data, make the trajectory smooth and continuous, and ensure the temporal order and integrity of the trajectory data. In this paper, for the features of spatiotemporal data of trajectories, the LSTM structure is integrated on the basis of the deep learning Transformer algorithm framework, abbreviated as TRFM-LS. The LSTM module can learn the temporal features of spatiotemporal data in the process of computing the target sequence, while the self-attention mechanism in Transformer can solve the drawback of applying LSTM to capture the sequence information weakly at a distance. The advantage of complementarity of the fusion model in the training process of trajectory sequences with respect to the long-range dependence of temporal and spatial features is realized. Finally, in the comparative analysis section of the error metrics, by comparing with current state-of-the-art methods, the algorithm in this paper is shown to have higher accuracy in predicting time series trajectory data. The research in this paper provides an early warning information reference for autonomous navigation and autonomous collision avoidance of ships in practice.

An Intelligent Algorithm for USVs Collision Avoidance Based on Deep Reinforcement Learning Approach with Navigation Characteristics

Many achievements toward unmanned surface vehicles have been made using artificial intelligence theory to assist the decisions of the navigator. In particular, there has been rapid development in autonomous collision avoidance techniques that employ the intelligent algorithm of deep reinforcement learning. A novel USV collision avoidance algorithm based on deep reinforcement learning theory for real-time maneuvering is proposed. Many improvements toward the autonomous learning framework are carried out to improve the performance of USV collision avoidance, including prioritized experience replay, noisy network, double learning, and dueling architecture, which can significantly enhance the training effect. Additionally, considering the characteristics of the USV collision avoidance problem, two effective methods to enhance training efficiency are proposed. For better training, considering the international regulations for preventing collisions at sea and USV maneuverability, a complete and reliable USV collision avoidance training system is established, demonstrating an efficient learning process in complex encounter situations. A reward signal system in line with the USV characteristics is designed. Based on the Unity maritime virtual simulation platform, an abundant simulation environment for training and testing is designed. Through detailed analysis, verification, and comparison, the improved algorithm outperforms the pre-improved algorithm in terms of stability, average reward, rules learning, and collision avoidance effect, reducing 26.60% more accumulated course deviation and saving 1.13% more time.

Machine-guided path sampling to discover mechanisms of molecular self-organization

Molecular self-organization driven by concerted many-body interactions produces the ordered structures that define both inanimate and living matter. Here we present an autonomous path sampling algorithm that integrates deep learning and transition path theory to discover the mechanism of molecular self-organization phenomena. The algorithm uses the outcome of newly initiated trajectories to construct, validate and—if needed—update quantitative mechanistic models. Closing the learning cycle, the models guide the sampling to enhance the sampling of rare assembly events. Symbolic regression condenses the learned mechanism into a human-interpretable form in terms of relevant physical observables. Applied to ion association in solution, gas-hydrate crystal formation, polymer folding and membrane-protein assembly, we capture the many-body solvent motions governing the assembly process, identify the variables of classical nucleation theory, uncover the folding mechanism at different levels of resolution and reveal competing assembly pathways. The mechanistic descriptions are transferable across thermodynamic states and chemical space.

InSAR Phase Unwrapping Algorithm Based on Deep GAN

At present, the traditional phase unwrapping algorithm is difficult to balance the accuracy and unwrapping efficiency. The traditional phase unwrapping algorithm is difficult to balance the accuracy and efficiency in the phase unwrapping experiments of simulated and measured topographic interferograms. In this paper, the phase unwrapping technology will be studied under the framework of deep learning theory according to the development trend of InSAR intelligence. A phase unwrapping algorithm based on deep GAN is proposed. The model structure includes dense convolution layer blocks and series structure, so that the network can achieve a better balance between feature extraction and detail preservation of interferogram. This is helpful to improve the phase unwrapping accuracy and training efficiency of the network. The experimental results show that the algorithm has a good expansion effect on the interferogram with high signal-to-noise ratio. The synthesis algorithm makes full use of the advantages of the branch cutting method and the finite element method. The phase reliable region and unreliable region are determined, and the transmission of phase error from the unreliable region to reliable region is effectively avoided. The accuracy of the phase unwrapping results in the reliable region is ensured, and the overall phase unwrapping convergence accuracy is greatly improved.

Implicit Bias of Deep Learning in the Large Learning Rate Phase: A Data Separability Perspective

Previous literature on deep learning theory has focused on implicit bias with small learning rates. In this work, we explore the impact of data separability on the implicit bias of deep learning algorithms under the large learning rate. Using deep linear networks for binary classification with the logistic loss under the large learning rate regime, we characterize the implicit bias effect with data separability on training dynamics. From a data analytics perspective, we claim that depending on the separation conditions of data, the gradient descent iterates will converge to a flatter minimum in the large learning rate phase, which results in improved generalization. Our theory is rigorously proven under the assumption of degenerate data by overcoming the difficulty of the non-constant Hessian of logistic loss and confirmed by experiments on both experimental and non-degenerated datasets. Our results highlight the importance of data separability in training dynamics and the benefits of learning rate annealing schemes using an initial large learning rate.

An Improved VGG16 and CNN-LSTM Deep Learning Model for Image Forgery Detection

As the field of image processing and computer vision continues to develop, we are able to create edited images that seem more natural than ever before. Identifying real photos from fakes has become a formidable obstacle. Image forgery has become more common as the multimedia capabilities of personal computers have developed over the previous several years. This is due to the fact that it is simpler to produce fake images. Since image object fabrication might obscure critical evidence, techniques for detecting it have been intensively investigated for quite some time. The publicly available datasets are insufficient to deal with these problems adequately. Our work recommends using a deep learning based image inpainting technique to create a model to detect fabricated images. To further detect copy-move forgeries in images, we use an CNN-LSTM and Improved VGG adaptation network. Our approach could be useful in cases when classifying the data is impossible. In contrast, researchers seldom use deep learning theory, preferring instead to depend on tried-and-true techniques like image processing and classifiers. In this article, we recommend the CNN-LSTM and improved VGG-16 convolutional neural network for intra-frame forensic analysis of altered images.

Intelligent Method to Optimize the Frequency Modulation for Beam Pumping System Based on Deep Reinforcement Learning.

A mathematical simulation model of a beam pumping system with frequency conversion control is established, considering the influence of the real-time frequency variation on the motion law of a pumping unit, the longitudinal vibration of a sucker rod string, the crankshaft torque, and the motor power. On this basis, the key links such as state space, action space, and reward function are defined by using deep reinforcement learning theory, and an intelligent model to optimize the frequency modulation for a beam pumping system based on deep reinforcement learning is constructed. The simulation and field application results show that the frequency optimization model can significantly reduce the fluctuation amplitude of the polished rod load, crankshaft torque, motor power, and input power of the system, making the operation of the pumping system more stable and energy-saving. More importantly, the model can realize the independent learning and control of the corresponding parameters without manual intervention to ensure the normal operation of the system and improve the level of information and intelligent management of oil wells.

Analysis and construction of the coal and rock cutting state identification system in coal mine intelligent mining

The recognition of cutting state of coal-rock is the key technology to realize “unmanned” mining in coal face. In order to realized real-time perception and accurate judgment of coal-rock cutting state information, this paper combined the field test sampling, construction technology of complex coal seam, virtual prototype technology, bidirectional coupling technology, data processing theory, image fusion method, and deep learning theory to carry out multi domain deep fusion experimental research on multi-source heterogeneous data of coal and rock cutting state. The typical complex coal seam containing gangue, inclusion, and minor fault in Yangcun mine of Yanzhou mining area was taken as the engineering object. The high-precision three-dimensional simulation model of the complex coal seam that can update and replace particles was constructed. Based on the simulation results of Discrete Element Method-Multi Flexible Body Dynamics (DEM-MFBD), the one-dimensional original vibration acceleration signals of the key components of the shearer cutting part were determined, including spiral drum, rocker arm shell, and square head. After transforming one-dimensional original signal data into two-dimensional time–frequency images by Short-time Fourier Transform, morphological wavelet image fusion technology was used to realize the effective fusion of characteristic information of spiral drum, rocker arm shell, and square head under different working conditions. Based on the deep learning theory, the DCGAN-RFCNN (Deep Convolutional Generative Adversarial Networks-Random Forest Convolutional Neural Networks) coal and rock cutting state recognition network model was constructed. Combining convolution neural network with random forest recognition classifier, RFCNN coal and rock cutting state recognition classification model was constructed, and the recognition network model was trained to obtain the model recognition results. Through the comparative experimental analysis of the RFCNN network model with different recognition network models and different synthetic sample numbers in the recognition network, the effectiveness of the recognition network model was verified. The results show that: When synthetic samples are not included in each working condition in the RFCNN model, the average recognition rate is 90.641%. With the increase of the number of synthetic samples, the recognition rate of coal and rock cutting state increases. When the number of synthetic samples added to each working condition reaches 5000, the recognition effect is the best, and the average recognition rate reaches 98.344%, which verifies the superiority of enriching the data set by using the improved DCGAN network. Also, the RFCNN outperformed the other variants: it obtained higher recognition accuracy by 25.085, 21.925 and 19.337%, respectively, over SVW, CNN, and AlexNet. Also, the experimental platform of shearer cutting coal and rock was built, where the coal and rock cutting state recognition network was trained and tested based on the migration learning theory. Through the statistical test results, the accuracy of coal and rock cutting state recognition is 98.64%, which realizes the accurate recognition of coal and rock cutting state.

Blockchain-Based Resource Trading in Multi-UAV-Assisted Industrial IoT Networks: A Multi-Agent DRL Approach

With the Industrial Internet of Things (IIoT), mobile devices (MDs) and their demands for low-latency data communication are increasing. Due to the limited resources of MDs, such as energy, computation, storage, and bandwidth, IIoT systems cannot meet MDs’ quality of service (QoS) and security requirements. Recently, UAVs have been deployed as aerial base stations in the IIoT network to provide connectivity and share resources with MDs. We consider a resource trading environment where multiple resource providers compete to sell their resources to MDs and maximize their profit by continually adjusting their pricing strategies. Multiple MDs, on the other hand, interact with the environment to make purchasing decisions based on the prices set by resource providers to reduce costs and improve QoS. We propose a novel intelligent resource trading framework that integrates multi-agent deep reinforcement Learning (MADRL), blockchain, and game theory to manage dynamic resource trading environments. A consortium blockchain with a smart contract is deployed to ensure the security and privacy of the resource transactions. We formulated the optimization problem using a Stackelberg game. However, the formulated optimization problem in the multi-agent IIoT environment is complex and dynamic, making it difficult to solve directly. Thus, we transform it into a stochastic game to solve the dynamics of the optimization problem. We propose a dynamic pricing algorithm that combines the Stackelberg game with the MADRL algorithm to solve the formulated stochastic game. The simulation results show that our proposed scheme outperforms others to improve resource trading in UAV-assisted IIoT networks.

CNN-based Model for Face Expression Recognition

Face recognition is a biometric technique that uses data on facial features to identify individuals. It is also a key area of study for computer vision researchers. CNN is a subclass of feedforward neural networks with convolutional processing and depth structure and one of the illustrative deep learning techniques. Since the deep learning theory was put forth and computational power increased, CNN has rapidly advanced and is now utilized in computer vision, natural language processing, and other fields. Our research is focused on face recognition, and because the mini-Xception model has a condensed volume and few parameters, it is used in this study. The dataset we used is fer2013, which is a classical dataset among CNN algorithms and is used in many studies. We also used data augmentation methods, and Keras’ ImageDataGenerator image generator was the optimal data augmentation method we came up with after reading the paper. Finally, we came up with a final model with 61% accuracy, which we are satisfied with and within the error results of the papers we reviewed.

Feature Extraction and Prediction of Water Quality Based on Candlestick Theory and Deep Learning Methods

In environmental hydrodynamics, a research topic that has gained popularity is the transmission and diffusion of water pollutants. Various types of change processes in hydrological and water quality are directly related to meteorological changes. If these changing characteristics are classified effectively, this will be conducive to the application of deep learning theory in water pollution simulation. When periodically monitoring water quality, data were represented with a candlestick chart, and different classification features were displayed. The water quality data from the research area from 2012 to 2019 generated 24 classification results in line with the physics laws. Therefore, a deep learning water pollution prediction method was proposed to classify the changing process of pollution to improve the prediction accuracy of water quality, based on candlestick theory, visual geometry group, and gate recurrent unit (CT-VGG-GRU). In this method, after the periodic changes of water quality were represented by candlestick graphically, the features were extracted by the VGG network based on its advantages in graphic feature extraction. Then, this feature and other scenario parameters were fused as the input of the time series network model, and the pollutant concentration sequence at the predicted station constituted the output of the model. Finally, a hybrid model combining graphical and time series features was formed, and this model used continuous time series data from multiple stations on the Lijiang River watershed to train and validate the model. Experimental results indicated that, compared with other comparison models, such as the back propagation neural network (BPNN), support vector regression (SVR), GRU, and VGG-GRU, the proposed model had the highest prediction accuracy, especially for the prediction of extreme values. Additionally, the change trend of water pollution was closer to the real situation, which indicated that the process change information of water pollution could be fully extracted by the CT-VGG-GRU model based on candlestick theory. For the water quality indicators DO, CODMn, and NH3-N, the mean absolute errors (MAE) were 0.284, 0.113, and 0.014, the root mean square errors (RMSE) were 0.315, 0.122, and 0.016, and the symmetric mean absolute percentage errors (SMAPE) were 0.022, 0.108, and 0.127, respectively. The established CT-VGG-GRU model achieved superior computational performance. Using the proposed model, the classification information of the river pollution process could be obtained effectively and the time series information could also be retained, which made the application of the deep learning model to the transmission and diffusion process of river water pollution more explanatory. The proposed model can provide a new method for water quality prediction.

Separation of scales and a thermodynamic description of feature learning in some CNNs

Deep neural networks (DNNs) are powerful tools for compressing and distilling information. Their scale and complexity, often involving billions of inter-dependent parameters, render direct microscopic analysis difficult. Under such circumstances, a common strategy is to identify slow variables that average the erratic behavior of the fast microscopic variables. Here, we identify a similar separation of scales occurring in fully trained finitely over-parameterized deep convolutional neural networks (CNNs) and fully connected networks (FCNs). Specifically, we show that DNN layers couple only through the second cumulant (kernels) of their activations and pre-activations. Moreover, the latter fluctuates in a nearly Gaussian manner. For infinite width DNNs, these kernels are inert, while for finite ones they adapt to the data and yield a tractable data-aware Gaussian Process. The resulting thermodynamic theory of deep learning yields accurate predictions in various settings. In addition, it provides new ways of analyzing and understanding DNNs in general.

Inception Convolution and Feature Fusion for Person Search.

With the rapid advancement of deep learning theory and hardware device computing capacity, computer vision tasks, such as object detection and instance segmentation, have entered a revolutionary phase in recent years. As a result, extremely challenging integrated tasks, such as person search, might develop quickly. The majority of efficient network frameworks, such as Seq-Net, are based on Faster R-CNN. However, because of the parallel structure of Faster R-CNN, the performance of re-ID can be significantly impacted by the single-layer, low resolution, and occasionally overlooked check feature diagrams retrieved during pedestrian detection. To address these issues, this paper proposed a person search methodology based on an inception convolution and feature fusion module (IC-FFM) using Seq-Net (Sequential End-to-end Network) as the benchmark. First, we replaced the general convolution in ResNet-50 with the new inception convolution module (ICM), allowing the convolution operation to effectively and dynamically distribute various channels. Then, to improve the accuracy of information extraction, the feature fusion module (FFM) was created to combine multi-level information using various levels of convolution. Finally, Bounding Box regression was created using convolution and the double-head module (DHM), which considerably enhanced the accuracy of pedestrian retrieval by combining global and fine-grained information. Experiments on CHUK-SYSU and PRW datasets showed that our method has higher accuracy than Seq-Net. In addition, our method is simpler and can be easily integrated into existing two-stage frameworks.

Strategies for Environmental Protection and Optimization of Ecological Business Economic Growth from the Perspective of Sustainable Development

The concept of ecological commercial economy refers to the use of ecological economics principles and system engineering methods to change production and consumption patterns within the scope of the carrying capacity of the ecosystem in order to tap into all of the available resource potential. It develops some economically developed and ecologically efficient industries and builds a culture with reasonable systems, a harmonious society, and a healthy ecological environment. This paper aims to use deep learning algorithms to study environmental protection and the optimization of ecological business economic growth from the perspective of sustainable development. In this regard, this paper proposes a theoretical model of environmental regulation, which aids in the study of the sustainable development of the ecological economy. Through experimental analysis, this study determined that the non-renewable resources of the two cities designated M and N dropped from 82% and 99% in 2017 to 78% and 79% in 2021, a decrease of 3% and 20%, respectively. This shows that the non-renewable resources of the four cities in area A generally showed a downward trend. The experimental results show that the deep learning theory and the environmental regulation model play a specific and effective role in the researching of the ecological business economy.

English Poetry Teaching based on Deep Learning Theory in Junior High School

English poetry is a special form of language with concise language, beautiful rhythm and profound connotation. In junior high school English teaching practice, English poetry is an indispensable form of discourse. However, in practical English teaching pedagogy, many teachers only focus on the language and structure of English poetry, without digging into the profound meaning of poetry, which is not conducive to cultivate students' key competence. The purpose of deep learning is to promote the development of students' key competence, so that students can construct knowledge and improve their ability of deep thinking. Therefore, this article based on literature research method, from the perspective of deep learning, on the basis of the English curriculum standard requirement to poetry teaching and the practical problems existing in the poetry teaching, combined with the specific teaching cases, explores the junior high school English poetry teaching strategy. It can be concluded as the way that in poetry teaching, teachers need to focus on poetry content from three aspects of language, culture and thought to promote deep learning process.