Deep Neural Network Model Training Research Articles

A dataset of drilling site object detection in underground coal mines

Drilling in underground coal mine is an important measure for dealing with gas, water and hidden geological disasters, which can significantly enhance the effectiveness of disaster prevention and control in coal mining operations. In order to monitor the drilling process in real time and improve drilling efficiency, it is necessary to carry out object detection to identify and locate key targets at the drilling site. Compared with traditional object detection method, deep learning-based object detection method can improve accuracy, timeliness and stability of object detection, but it requires high-quality datasets to perform well. At present, research on object detection in underground coal mine drilling sites mainly relies on small-scale private datasets, which are insufficient for providing necessary or reliable data for deep neural network model training. In this study, we constructed a dataset of drilling site object detection using photos taken by intrinsic safety law enforcement recorders. This dataset is developed through several steps, including data cleaning, data labeling, and expert sampling verification. The mainstream YOLO series object detection model is used for data quality assessment. This dataset comprises 70,948 images from drilling sites under different environmental conditions, covering five categories of objects: gripper, chuck, coal miner, mine safety helmet, and drill pipe. It provides annotated files in PASCAL VOC format. This dataset can provide strong data support for object detection research in underground coal mine drilling sites, and plays an important role in promoting intelligent underground coal mine monitoring and early warning.

Robustness analysis and experimental validation of a deep neural network for acoustic source imaging

Deep Neural Network (DNN) models offer an attractive alternative to existing acoustic source imaging techniques, such as acoustic beamforming, due to their ever-growing potential with increasing computational power. Source resolution of acoustic beamforming methods is limited at lower frequencies and their source maps may possess sidelobes at higher frequencies. However, acoustic beamforming methods are typically robust over a wide range of simulation and experimental conditions, such as (i) the number of sources present, (ii) source frequency and (ii) extraneous noise sources. The performance of DNN models, when these conditions are varied from their specific design criteria, is yet to be investigated and much work is needed in this area before DNN models can be utilized in experiments, such as wind tunnel tests. Furthermore, few studies have been conducted on experimental validation of DNN models, predominately due to the difficulty of large sets of experimentally obtained data needed for DNN model training and the sensitivity of DNN model performance when any of the aforementioned experimental conditions are varied. In this paper, a series of studies on the robustness of DNN models based on numerical data and experimental data are presented. Numerical DNN (NDNN) models are trained using in-phase and random-phase pressure data generated from six sources over design frequencies from 500 Hz to 20,000 Hz. The robustness of the NDNN models is tested via (1) inclusion of extraneous Gaussian white noise, (2) inclusion of extraneous tonal noise near the design frequency, (3) using source frequencies that slightly differ from the design frequencies and (4) using a number of sources that differs from the design source number. DNN model performance metrics are introduced that present a promising framework for future DNN model studies and bridging the gap between NDNN and experimentally trained DNN models. A preliminary experimental validation was conducted using a single speaker that was systematically placed over a speaker grid to generate training data via acoustic superposition, from which an experimentally trained DNN (EDNN) model was produced. The EDNN model yields exceptional noise source localization capability of the DNN model, revealing a promising start for a more sophisticated EDNN model.

Advancing Process Control in Fluidized Bed Biomass Gasification Using Model-Based Deep Reinforcement Learning

This study presents a model-based deep reinforcement learning (MB-DRL) controller for the fluidized bed biomass gasification (FBG) process. The MB-DRL controller integrates a deep neural network (DNN) model and a reinforcement learning-based optimizer. The DNN model is trained with operational data from a pilot-scale FBG plant to approximate FBG process dynamics. The reinforcement learning-based optimizer employs a specially designed reward function, determining optimal control policies for FBG. Moreover, the controller includes an online learning component, ensuring periodic updates to the DNN model training. The performance of the controller is evaluated by testing its control accuracy for regulating synthetic gas composition, flow rate, and CO concentration in the FBG. The evaluation also includes a comparison with a model predictive controller. The results demonstrate the superior control performance of MB-DRL, surpassing MPC by over 15% in regulating synthetic gas composition and flow rate, with similar effectiveness observed in synthetic gas temperature control. Additionally, this study also includes systematic investigations into factors like DNN layer count and learning update intervals to provide insights for the practical implementation of the controller. The results, presenting a 50% reduction in control error with the addition of a single layer to the DNN model, highlight the significance of optimizing MB-DRL for effective implementation.

Improving Deep Learning Anomaly Diagnostics with a Physics-Based Simulation Model

Deep learning algorithms often struggle to accurately distinguish between healthy and anomalous states due to the scarcity of high-quality data in real-world applications. However, these data can be obtained through a physics-based simulation model. In this research, the model serves a dual purpose: detecting anomalies in industrial processes and replicating the machine’s operational behavior with high fidelity in terms of a simulated torque signal. When anomalous behaviors are detected, their patterns are utilized to generate anomalous events, contributing to the enhancement of deep neural network model training. This research proposes a method, named Simulation-Enhanced Anomaly Diagnostics (SEAD), to detect anomalies and further create high-quality data related to the diagnosed faults in the machine’s operation. The findings of this study suggest that employing a physics-based simulation model as a synthetic-anomaly signal generator can significantly improve the classification accuracy of identified anomalous states, thereby enhancing the deep learning model’s ability to recognize deviating behavior at an earlier stage when more high-quality data of the identified anomaly has been available for the learning process. This research measures the classification capability of a Long Short-Term Memory (LSTM) autoencoder to classify anomalous behavior in different SEAD stages. The validated results clearly demonstrate that simulated data can contribute to the LSTM autoencoder’s ability to classify anomalies in a peripheral milling machine. The SEAD method is employed to test its effectiveness in detecting and replicating a failure in the support element of the peripheral milling machine.

Energy-Aware, Device-to-Device Assisted Federated Learning in Edge Computing

The surging of deep learning brings new vigor and vitality to shape the prospect of intelligent Internet of Things (IoT), and the rise of edge intelligence enables provisioning real-time deep neural network (DNN) inference services for mobile users. To perform efficient and effective DNN model training in edge computing environments while preserving training data security and privacy of IoT devices, federated learning has been envisioned as an ideal learning paradigm for this purpose. In this paper, we study energy-aware DNN model training in edge computing. We first formulate a novel energy-aware, Device-to-Device (D2D) assisted federated learning problem with the aim to minimize the global loss of a training DNN model, subject to bandwidth capacity on an edge server and energy capacity on each IoT device. We then devise a near-optimal learning algorithm for the problem when the training data follows the i.i.d. data distribution. The crux of the proposed algorithm is to explore using the energy of neighboring devices of each device for its local model uploading, by reducing the problem to a series of weighted maximum matching problems in corresponding auxiliary graphs. We also consider the problem without the assumption of the i.i.d. data distribution, for which we propose an efficient heuristic algorithm. We finally evaluate the performance of the proposed algorithms through experimental simulations. Experimental results show that the proposed algorithms are promising.

FEM-Inclusive Transfer Learning for Bistable Piezoelectric MEMS Energy Harvester Design

In this article, a bistable piezoelectric MEMS energy harvester is presented to operate in a low-frequency range, around 100–200 Hz. The proposed design has an M-shaped structure with a couple of proof masses to not only lower the operating frequencies but also enlarge the frequency bandwidth. This specific structure has multidegrees of freedom, making it fit for a bistable piezoelectric energy harvester on the MEMS scale. An artificial neural network (ANN) is used to tackle this design in order to facilitate the optimization process and determine proper physical dimensions. To improve the accuracy and boost the training process of deep neural network (DNN), we utilize a transfer learning technique in this work. The analytical modeling and finite-element modeling (FEM) simulation data have been used for the DNN model training. Here, a DNN is first trained with a large dataset computed from the lumped-parameter model, and then, the trained network is transferred to a new DNN model for another round of training with a small dataset of highly accurate FEM simulation data samples to further reduce the estimation error. It is shown that the new model can estimate the device features with over 94% accuracy, which is considerably higher than the regular DNN. Next, the trained model is used as a performance estimator in a genetic algorithm (GA) to optimize the topology of the device to improve the operating frequency range and the generated voltage. An optimized design with a total volume of 1.02 mm3 was fabricated by the micromachining process. Our experimental results confirm that the proposed transfer-learning-based method can not only reduce the prototype’s first and second resonant frequencies to 123.8 and 175.7 Hz, respectively, but also enhance the generated power up to <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$2.83 ~\mu \text{W}$ </tex-math></inline-formula> under 0.2-g input acceleration.

Electro-Optical Sensors for Atmospheric Turbulence Strength Characterization with Embedded Edge AI Processing of Scintillation Patterns

This study introduces electro-optical (EO) sensors (TurbNet sensors) that utilize a remote laser beacon (either coherent or incoherent) and an optical receiver with CCD camera and embedded edge AI computer (Jetson Xavier Nx) for in situ evaluation of the path-averaged atmospheric turbulence refractive index structure parameter Cn2 at a high temporal rate. Evaluation of Cn2 values was performed using deep neural network (DNN)-based real-time processing of short-exposure laser-beacon light intensity scintillation patterns (images) captured by a TurbNet sensor optical receiver. Several pre-trained DNN models were loaded onto the AI computer and used for TurbNet sensor performance evaluation in a set of atmospheric propagation inference trials under diverse turbulence and meteorological conditions. DNN model training, validation, and testing were performed using datasets comprised of a large number of instances of scintillation frames and corresponding reference (“true”) Cn2 values that were measured side-by-side with a commercial scintillometer (BLS 2000). Generation of datasets and inference trials was performed at the University of Dayton’s (UD) 7-km atmospheric propagation test range. The results demonstrated a 70–90% correlation between Cn2 values obtained with the TurbNet sensors and those measured side-by-side with the scintillometer.

Multi-Level Functional Connectivity Fusion Classification Framework for Brain Disease Diagnosis.

Brain disease diagnosis is a new hotspot in the cross research of artificial intelligence and neuroscience. Quantitative analysis of functional magnetic resonance imaging (fMRI) data can provide valuable biomarkers that contributes to clinical diagnosis, and the analysis of functional connectivity (FC) has become the primary method. However, previous studies mainly focus on brain disease classification based on the low-order FC features, ignoring the potential role of high-order functional relationships among brain regions. To solve this problem, this study proposed a novel multi-level FC fusion classification framework (MFC) for brain disease diagnosis. We firstly designed a deep neural network (DNN) model to extract and learn abstract feature representations for the constructed low-order and high-order FC patterns. Both unsupervised and supervised learning steps were performed during the DNN model training, and the prototype learning was introduced in the supervised fine-tuning to improve the intra-class compactness and inter-class separability of the feature representation. Then, we combined the learned multi-level abstract FC features and trained an ensemble classifier with a hierarchical stacking learning strategy for the brain disease classification. Systematic experiments were conducted on two real large-scale fMRI datasets. Results showed that the proposed MFC model obtained robust classification performance for different preprocessing pipelines, different brain parcellations, and different cross-validation schemes, suggesting the effectiveness and generality of the proposed MFC model. Overall, this study provides a promising solution to combine the informative low-order and high-order FC patterns to further promote the classification of brain diseases.

Human Action Recognition Algorithm Based on Adaptive Initialization of Deep Learning Model Parameters and Support Vector Machine

Action recognition is a trending topic and key research direction in computer vision, machine learning, artificial intelligence, and other fields. This research seeks to identify human action in image and video data. Its research results have been widely used in the fields of safety monitoring, disability monitoring, understanding multimedia content, human–computer interaction, virtual reality, and so on. However, the existing traditional human action recognition technology has many limitations in practical application, such as low accuracy and weak adaptive ability. Although the action recognition based on deep learning can self-learn and improve the action recognition accuracy, there are many difficulties in training the deep neural network model, such as gradient disappearance, gradient explosion, and overfitting. Therefore, this paper will reduce the abovementioned difficulties in deep neural network model training from the perspective of deep neural network model parameter initialization and then propose a model parameter initialization method based on the multilayer maxout network activation function to solve the difficulties in deep neural network model training. Then, on this basis, a method of learning the temporal and spatial characteristics of human action based on the deep neural network model is proposed. First, the method detects and tracks the human action and uses the restricted Boltzmann machine (RBM) to encode the temporal and spatial features of various parts of the human body. Second, the temporal and spatial feature codes of various parts of the human body are integrated into a global temporal and spatial feature representation method of the action video through a RBM neural network. Finally, the trained SVM classifiers are used to recognize human action. Experiments show that the human action recognition method proposed in this paper not only has high recognition accuracy but also has great adaptability. Thus, this method extracts temporal and spatial features from the shape feature sequences of various parts of the human body, thus opening up a new way to extract human action features and solving the problem of human action recognition in complex scenes. Its proposal provides an exploratory technical method and approach for self-adaptive recognition of human action. It also gives directional enlightenment to the development and improvement of self-adaptive human action methods.