DNN Model Research Articles

Study on the human-brain confusion levels and corresponding EEG levels

Confusion is a state in which an individual is not clear about the situation at hand, does not understand the logic of the matter, or cannot arrive at a reasoned outcome. When people are confused, they often make poor decisions or even fail to make decisions. In this paper, we tested and validated our two models using the collected EEG signals of ten students while watching online courses of different difficulty levels. Both the LSTM model and the DNN model obtained an accuracy rate close to 70% in the validation part, but they showed different characteristics. In the end, we compare and summarize the results of the two models and try to generalize them to other domains in the future.

A novel and effective method based deep learning model for detecting non-technical electricity losses

<p>This study focused on non-technical electricity loss detection. As mentioned, non-technical losses (NTLs) affect utilities and economies financially. Electricity theft, fraud, and metering issues can create NTLs. NTL generate most distribution losses in electrical power networks, costing utilities a lot. NTL detection approaches are data-focused, network-oriented, or hybrid. Data-oriented writing dominated this analysis. After data collection and cleaning and labeling the unlabeled dataset with a target, a methodology was supplied that used four machine learning techniques random forest, decision tree, KNN, and logistic regression and four neural network models-DNN, CNN, CNN-LSTM, and CNN-GRU. The CNN and DNN model have the best accuracy, stability, fast learning, and training time.</p>

Interpreting Convolutional Layers in DNN Model Based on Time–Frequency Representation of Emotional Speech

Abstract The paper describes the relations of speech signal representation in the layers of the convolutional neural network. Using activation maps determined by the Grad-CAM algorithm, energy distribution in the time–frequency space and their relationship with prosodic properties of the considered emotional utterances have been analysed. After preliminary experiments with the expressive speech classification task, we have selected the CQT-96 time–frequency representation. Also, we have used a custom CNN architecture with three convolutional layers in the main experimental phase of the study. Based on the performed analysis, we show the relationship between activation levels and changes in the voiced parts of the fundamental frequency trajectories. As a result, the relationships between the individual activation maps, energy distribution, and fundamental frequency trajectories for six emotional states were described. The results show that the convolutional neural network in the learning process uses similar fragments from time–frequency representation, which are also related to the prosodic properties of emotional speech utterances. We also analysed the relations of the obtained activation maps with time-domain envelopes. It allowed observing the importance of the speech signals energy in classifying individual emotional states. Finally, we compared the energy distribution of the CQT representation in relation to the regions’ energy overlapping with masks of individual emotional states. In the result, we obtained information on the variability of energy distributions in the selected signal representation speech for particular emotions.

Time and Space Domain Prediction of Water Quality Parameters of Bagmati River Using Deep Learning Methods

Bagmati river is biologically, geologically, religiously and historically significant among the river systems of the Kathmandu Valley. The river is affected by five major tributaries, including Manohara, Dhobi Khola, Tukucha, Bishnumati, and Balkhu Khola, which significantly impact the water chemistry inside the Kathmandu Valley. The data of water quality parameters pH, dissolved oxygen, turbidity, temperature, oxygen reduction potential, conductivity, total dissolved solids, salinity among others was collected using fixed sensors (in period of 5 seconds) and mobile sensors (with latitude and longitude) along the river. The observation is important for two reasons, one because it was collected in real-time and fine scale, which is not normally possible with traditional ways, and next such observation was done for the first time in Bagmati River. The aim of this study was to predict water quality parameters of the Bagmati River using machine learning time series models, specifically ARIMA and LSTM. The LSTM model was designed with one input layer, one encoder layer, one repeat layer, one decoder layer, and one output dense layer to separate the output into temporal slices. Additionally, a DNN model was employed for location-based prediction, utilizing two input layers for latitude and longitude and seven output layers for the seven water quality parameters considered for study. The models demonstrated promising performance, but further data collection and parameter variation are recommended for continued optimization.

Service recommendation based on contrastive learning and multi-task learning

Service recommendation is an efficient method for service-oriented software that can improve software quality. Applications often require the integration of multiple services to create more powerful and complex functionality while saving software development time. However, the vast number of available candidate Web services can impose a heavy burden on software developers’ selection decisions. The existing service recommendation challenges are mainly come from: (1) the development requirements entered by users are too arbitrary (2) the extreme sparsity of invocation records. To address the above challenges, in this paper, we propose a Service Recommendation method based on Contrastive Learning and Multi-task Learning (SRCLML). Specifically, we utilize the Transformer model to extract the development requirements of users, conduct in-depth mining of text descriptions, and extract features of applications. Next, the features are fed into the DNN model to predict the probability that the service will be selected. Moreover, we add a tag judgment task to make it capable of multi-task learning, through which, the training signal information implied can be used as an inductive bias to improve service recommendation capabilities. Additionally, we build three subgraphs based on the global graph, conduct in-depth mining of historical invocation records based on contrastive learning and graph neural network to extract features of applications and services and calculate application preferences for each service. Finally, we combined the above two to obtain the final recommendation service list. Extensive experiments on real-world datasets demonstrate that our method, SRCLML, outperforms several state-of-the-art comparison methods in the domain of service recommendation.

Assessing the impact of borehole field data on AI-based deep learning models for heating and cooling prediction

Sensor readings play a critical role in the prediction performance of AI-based data-driven fault-detection and diagnostics for borehole heat exchangers in heating and cooling systems. The accuracy and algorithm selection of individual deep neural networks are well-investigated research topics but the impact of quality of sensors data on algorithm selection and its further impact on prediction performance has not been investigated and quantified for geothermal heating/cooling systems. The objective of the current work is to investigate the impact of data quality, available as measured time-series data from the borehole field, on the prediction accuracy of the deep neural networks-based models. In the current work four different data sets (D-1, D-2, D-3, D-4) are used for data-driven modeling using four different deep machine learning approaches, i.e., LSTM, BD-LSTM, GRU and CNN. The developed AI models have been trained, validated, and tested using the real-field data sets with same number of independent input parameters i.e., inlet temperature and mass flow rate to predict the fluid outlet temperature. The validated AI models are used to predict the performance of borehole heat exchanger to meet the heating and cooling loads. The prediction results show that the BD-LSTM outperforms other deep neural networks in terms of accuracy. With the data sets D-1, D-2, D-3 and D-4 it was observed that BD-LSTM had lowest MAPE of 2.99 %, 0.22 %, 6.39 % and 2.21 % respectively. Several experiments show that the BD-LSTM and LSTM models are better than the other used approaches in terms of prediction accuracy, but larger number of trainable parameters results in more computation time than CNN and GRU respectively. More data points with higher sampling frequency, results in more accurate prediction as depicted by the results from D-2 and D-4. Although, the results showed that the data collection interval is crucial, but different test runs indicate that avoiding the loss of time-series data during the cleaning process is equally important to accurately capture the Seasonality and Trend of the heating/cooling data. This is realized in case of D-1 where almost 58 % of the values were removed as outliers and these missing values were needed to be approximated which also affected the model prediction accuracy. The presented analysis highlights that the LSTM based DNN model has the capability to predict borehole heat exchanger performance accurately and the predictions are highly dependent on data frequency, the quality as well as data quantity from geothermal installations used to train the model.

Machine learning framework for precise localization of bleached corals using bag-of-hybrid visual feature classification

Corals are sessile invertebrates living underwater in colorful structures known as reefs. Unfortunately, coral’s temperature sensitivity is causing color bleaching, which hosts organisms that are crucial and consequently affect marine pharmacognosy. To address this problem, many researchers are developing cures and treatment procedures to restore bleached corals. However, before the cure, the researchers need to precisely localize the bleached corals in the Great Barrier Reef. The researchers have developed various visual classification frameworks to localize bleached corals. However, the performance of those techniques degrades with variations in illumination, orientation, scale, and view angle. In this paper, we develop highly noise-robust and invariant robust localization using bag-of-hybrid visual features (RL-BoHVF) for bleached corals by employing the AlexNet DNN and ColorTexture handcrafted by raw features. It is observed that the overall dimension is reduced by using the bag-of-feature method while achieving a classification accuracy of 96.20% on the balanced dataset collected from the Great Barrier Reef of Australia. Furthermore, the localization performance of the proposed model was evaluated on 342 images, which include both train and test segments. The model achieved superior performance compared to other standalone and hybrid DNN and handcrafted models reported in the literature.

Enhancing Cardiovascular Disease Prediction Based on AI and IoT Concepts

One-third of all deaths worldwide yearly are attributable to cardiovascular disease (CVD). In contrast to the 7% of the wealthy who experience premature death, 43% of the poor do. Lifestyle diseases like obesity and diabetes are to blame. The importance of early identification of heart disease was demonstrated, and premature mortality was kept to a minimum. Combining clinical and biochemical data is essential for the early diagnosis of heart illness. Numerous IoT-enabled wearable healthcare applications have been created and released in recent years. Although the ability of wearable devices to share patient health data is expanding, it remains challenging to predict and identify health problems. Security, data storage, and patient monitoring are all part of the system. Artificial intelligence (AI) therapies may one day change the face of cardiology by providing doctors with cutting-edge data analysis and therapeutic decision-making resources. As the volume and complexity of data continue to increase, AI tools like machine learning (ML) and deep learning (DL) can assist medical professionals in learning more. Suppose we want to provide medical care to the elderly and those with chronic illnesses in the comfort of their own homes. In that case, we must upgrade our communication and information technology systems. The implemented DNN model's accuracy is amazing at 95.34 % and can yield other noteworthy outcomes when used to identify CVDs. We discuss and suggest the most suitable AI-IoT models for early CVD prediction and detection to reduce computational costs and increase time efficiency.

Time-Constrained Ensemble Sensing With Heterogeneous IoT Devices in Intelligent Transportation Systems

Recently we have witnessed the rise of Artificial Intelligence of Things (AIoT) and the shift of sensing paradigm from cloud-centric to the edge-centric, which effectively improves the sensing capability of intelligence transportation systems. To improve the real-time sensing performance, in this work we propose an ensemble sensing based scheme to solve the time-constraint synchronized inference problem and achieve robust inference with heterogeneous IoT devices in intelligence transportation systems. We design and implement Ensen, which incorporates various novel techniques such as customized DNN model design, KD-based model training, and dynamic deep ensemble management, etc., to achieve improved accuracy and maximize the computational resource usage of the whole sensing group. Extensive evaluations on different types of common IoT devices have shown that Ensen achieves a robust performance and can be easily extended to different types of convolutional neural networks.

Breaking down health fakes: a hybrid DNN model for multi-class classification on a self-constructed dataset

Breaking down health fakes: a hybrid DNN model for multi-class classification on a self-constructed dataset

Deep neural network techniques for monaural speech enhancement and separation: state of the art analysis

AbstractDeep neural networks (DNN) techniques have become pervasive in domains such as natural language processing and computer vision. They have achieved great success in tasks such as machine translation and image generation. Due to their success, these data driven techniques have been applied in audio domain. More specifically, DNN models have been applied in speech enhancement and separation to perform speech denoising, dereverberation, speaker extraction and speaker separation. In this paper, we review the current DNN techniques being employed to achieve speech enhancement and separation. The review looks at the whole pipeline of speech enhancement and separation techniques from feature extraction, how DNN-based tools models both global and local features of speech, model training (supervised and unsupervised) to how they address label ambiguity problem. The review also covers the use of domain adaptation techniques and pre-trained models to boost speech enhancement process. By this, we hope to provide an all inclusive reference of all the state of art DNN based techniques being applied in the domain of speech separation and enhancement. We further discuss future research directions. This survey can be used by both academic researchers and industry practitioners working in speech separation and enhancement domain.

Integration of multimodal data for large-scale rapid agricultural land evaluation using machine learning and deep learning approaches

Rapid and accurate agricultural land evaluation provides essential guidance for the supervision and allocation of agricultural land resources; it also helps to ensure food security. Previous work has mainly evaluated the land quality at the county level by using field sampling data and based on a factor approach. However, it is difficult to achieve uniform, large-scale agricultural land evaluation via conventional approaches because of its spatial heterogeneity, as well as the large temporal and economic costs associated with data acquisition. In this study, we integrated publicly available multimodal data (i.e., satellite remote sensing, environmental, and socioeconomic data) into the Google Earth Engine (GEE) platform, selected the best indicators from each modality using the geodetector, on the basis of which different combinations of input models were designed. And then we developed machine learning (random forest, RF) and deep learning (deep neural network, DNN) models to evaluate the land quality in paddy field and dry land systems in 2013 throughout Guangdong Province, China. The results showed that the performance of our combination of variables decreased in the following order: multimodal > bimodal > unimodal. With the best input combination, the RF model (R2 = 0.91, RMSE = 97.56, and CCC = 0.95) outperformed the DNN model (R2 = 0.89, RMSE = 108.72, and CCC = 0.94) in terms of predicting the quality of paddy field. The RF model (R2 = 0.90, RMSE = 104.27, and CCC = 0.95) also outperformed the DNN model (R2 = 0.86, RMSE = 124.38, and CCC = 0.93) in terms of predicting the quality of dry land. The agricultural land quality estimates obtained using the RF and DNN models were more accurate for paddy field than for dry land systems because of greater land quality homogeneity in paddy fields. This research proposed a simple, low-cost for rapid and accurate agricultural land evaluation at the provincial scale using publicly available multimodal data, which can help to achieve control of the agricultural land grade at multiple spatial and temporal scales.

Data augmentation using continuous conditional generative adversarial networks for regression and its application to improved spectral sensing.

Machine learning-assisted spectroscopy analysis faces a prominent constraint in the form of insufficient spectral samples, which hinders its effectiveness. Meanwhile, there is a lack of effective algorithms to simulate synthetic spectra from limited samples of real spectra for regression models in continuous scenarios. In this study, we introduced a continuous conditional generative adversarial network (CcGAN) to autonomously generate synthetic spectra. The labels employed for generating the spectral data can be arbitrarily selected from within the range of labels associated with the real spectral data. Our approach effectively produced spectra using a small spectral dataset obtained from a self-interference microring resonator (SIMRR)-based sensor. The generated synthetic spectra were subjected to evaluation using principal component analysis, revealing an inability to discern them from the real spectra. Finally, to enhance the DNN regression model, these synthetic spectra are incorporated into the original training dataset as an augmentation technique. The results demonstrate that the synthetic spectra generated by CcGAN exhibit exceptional quality and significantly enhance the predictive performance of the DNN model. In conclusion, CcGAN exhibits promising potential in generating high-quality synthetic spectra and delivers a superior data augmentation effect for regression tasks.

Novel Optimized Strategy Based on Multi-Next-Hops Election to Reduce Video Transmission Delay for GPSR Protocol over VANETs

Reducing transmission traffic delay is one of the most important issues that need to be considered for routing protocols, especially in the case of multimedia applications over vehicular ad hoc networks (VANET). To this end, we propose an extension of the FzGR (fuzzy geographical routing protocol), named MNH-FGR (multi-next-hops fuzzy geographical routing protocol). MNH-FGR is a multipath protocol that gains great extensibility by employing different link metrics and weight functions. To schedule multimedia traffic among multiple heterogeneous links, MNH-FGR integrates the weighted round-robin (WRR) scheduling algorithm, where the link weights, needed for scheduling, are computed using the multi-constrained QoS metric provided by the FzGR. The main goal is to ensure the stability of the network and the continuity of data flow during transmission. Simulation experiments with NS-2 are presented in order to validate our proposal. Additionally, we present a neural network algorithm to analyze and optimize the performance of routing protocols. The results show that MNH-FGR could satisfy critical multimedia applications with high on-time constraints. Also, the DNN model used can provide insights about which features had an impact on protocol performance.

Remaining Useful Life Prediction Method for High Temperature Blades of Gas Turbines Based on 3D Reconstruction and Machine Learning Techniques

Turbine blades are crucial components exposed to harsh conditions, such as high temperatures, high pressures, and high rotational speeds. It is of great significance to accurately predict the life of blades for reducing maintenance cost and improving the reliability of gas turbine systems. A rapid and accurate blade life assessment method holds significant importance in the maintenance plan of gas turbine engines. In this paper, a novel on-line remaining useful life (RUL) prediction method for high-temperature blades is proposed based on 3D reconstruction technology and data-driven surrogate mode. Firstly, the 3D reconstruction technology was employed to establish the geometric model of real turbine blades, and the fluid–thermal–solid analysis under actual operational conditions was carried out in ANSYS software. Six checkpoints were selected to estimate the RUL according to the stress–strain distribution of the blade surface. The maximum equivalent stress was 1481.51 MPa and the highest temperature was 1393.42 K. Moreover, the fatigue-creep lifetime was calculated according to the parameters of the selected checkpoints. The RUL error between the simulation model and commercial software (Control and Engine Health Management (CEHM)) was less than 0.986%. Secondly, different data-driven surrogate models (BP, DNN, and LSTM algorithms) were developed according to the results from numerical simulation. The maximum relative errors of BP, DNN, and LSTM models were 0.030%, 0.019%, and 0.014%. LSTM demonstrated the best performance in predicting the RUL of turbine blades with time-series characteristics. Finally, the LSTM model was utilized for predicting the RUL within a gas turbine real operational process that involved five start–stop cycles.

Predicting VOCs content and roasting methods of lamb shashliks using deep learning combined with chemometrics and sensory evaluation.

A comparison was made between the traditional charcoal-grilled lamb shashliks (T) and four new methods, namely electric oven heating (D), electric grill heating (L), microwave heating (W), and air fryer treatment (K). Using E-nose, E-tongue, quantitative descriptive analysis (QDA), and HS-GC-IMS and HS-SPME-GC-MS, lamb shashliks prepared using various roasting methods were characterized. Results showed that QDA, E-nose, and E-tongue could differentiate lamb shashliks with different roasting methods. A total of 43 and 79 volatile organic compounds (VOCs) were identified by HS-GC-IMS and HS-SPME-GC-MS, respectively. Unsaturated aldehydes, ketones, and esters were more prevalent in samples treated with the K and L method. As a comparison to the RF, SVM, 5-layer DNN and XGBoost models, the CNN-SVM model performed best in predicting the VOC content of lamb shashliks (accuracy rate all over 0.95) and identifying various roasting methods (accuracy rate all over 0.92).

Automatic DNN architecture design using CPSOTJUTT for power system inspection

To quickly and accurately automatically design more high-precision deep neural network models (DNNs), this paper proposes an automatic DNN architecture design ensemble model based on consensus particle swarm optimization-assisted trajectory unified and TRUST-TECH (CPSOTJUTT), called CPSOTJUTT-EM. The proposed model is a three-layer model, and its core is a three-stage method for addressing the sensitivity of the local solver to the initial point and enabling fast and robust training DNN, effectively avoiding missing high-quality DNN models in the process of automatic DNN architecture design. CPSOTJUTT has the following advantages: (1) high-quality local optimal solutions (LOSs) and (2) robust convergence against random initialization. CPSOTJUTT-EM consists of the bottom layer: stable and fast design high-quality DNN architectures, the middle layer: exploration for a diverse set of optimal DNN classification engines, and the top layer: ensemble model for higher performance. This paper tests the performance of CPSOTJUTT-EM on public datasets and three self-made power system inspection datasets. Experimental results show that the CPSOTJUTT-EM has excellent performance in automatic DNN architecture design, DNN model optimization. And the CPSOTJUTT-EM can automatically design high-quality DNN ensemble models, laying a solid foundation for the application of DNN in other fields.

Using Deep Neural Network Methods for Forecasting Energy Productivity Based on Comparison of Simulation and DNN Results for Central Poland—Swietokrzyskie Voivodeship

Forecasting electricity demand is of utmost importance for ensuring the stability of the entire energy sector. However, predicting the future electricity demand and its value poses a formidable challenge due to the intricate nature of the processes influenced by renewable energy sources. Within this piece, we have meticulously explored the efficacy of fundamental deep learning models designed for electricity forecasting. Among the deep learning models, we have innovatively crafted recursive neural networks (RNNs) predominantly based on LSTM and combined architectures. The dataset employed was procured from a SolarEdge designer. The dataset encompasses daily records spanning the past year, encompassing an exhaustive collection of parameters extracted from solar farm (based on location in Central Europe (Poland Swietokrzyskie Voivodeship)). The experimental findings unequivocally demonstrated the exceptional superiority of the LSTM models over other counterparts concerning forecasting accuracy. Consequently, we compared multilayer DNN architectures with results provided by the simulator. The measurable results of both DNN models are multi-layer LSTM-only accuracy based on R2—0.885 and EncoderDecoderLSTM R2—0.812.

Prediction of Equivalence Ratio in Combustion Flame Using Chemiluminescence Emission and Deep Neural Network

A reliable combustion monitoring system is essential to satisfy global carbon neutrality trends. As the concentrations of emissions and flame stability are associated with the air–fuel ratio, the equivalence ratio should be continuously evaluated. In this study, a deep neural network- (DNN-) based regression model is proposed to predict the equivalence ratio of turbulent diffusion flames. Chemiluminescence signals from the OH ∗ , CH ∗ , and C2 ∗ radicals were acquired as input features. In addition, three different optical sensing views were applied to consider the future general measurement conditions. Furthermore, a loss function comparison for model training and hyperparameter tuning techniques, such as random search and Bayesian optimization, were used to improve the prediction performance. Consequently, the enhanced DNN model showed reductions in the mean absolute error and root mean square error of ~17.84% and ~12.06%, respectively, compared with the initial model. In addition, a mean absolute percentage error and R -squared value of ~3.61% and ~0.9311, respectively, were obtained. Thus, a novel sensing method has been proposed for flame monitoring systems to realize future digital transformations in the combustion industry.

Extracting and visualizing hidden activations and computational graphs of PyTorch models with TorchLens

Deep neural network models (DNNs) are essential to modern AI and provide powerful models of information processing in biological neural networks. Researchers in both neuroscience and engineering are pursuing a better understanding of the internal representations and operations that undergird the successes and failures of DNNs. Neuroscientists additionally evaluate DNNs as models of brain computation by comparing their internal representations to those found in brains. It is therefore essential to have a method to easily and exhaustively extract and characterize the results of the internal operations of any DNN. Many models are implemented in PyTorch, the leading framework for building DNN models. Here we introduce TorchLens, a new open-source Python package for extracting and characterizing hidden-layer activations in PyTorch models. Uniquely among existing approaches to this problem, TorchLens has the following features: (1) it exhaustively extracts the results of all intermediate operations, not just those associated with PyTorch module objects, yielding a full record of every step in the model's computational graph, (2) it provides an intuitive visualization of the model's complete computational graph along with metadata about each computational step in a model's forward pass for further analysis, (3) it contains a built-in validation procedure to algorithmically verify the accuracy of all saved hidden-layer activations, and (4) the approach it uses can be automatically applied to any PyTorch model with no modifications, including models with conditional (if–then) logic in their forward pass, recurrent models, branching models where layer outputs are fed into multiple subsequent layers in parallel, and models with internally generated tensors (e.g., injections of noise). Furthermore, using TorchLens requires minimal additional code, making it easy to incorporate into existing pipelines for model development and analysis, and useful as a pedagogical aid when teaching deep learning concepts. We hope this contribution will help researchers in AI and neuroscience understand the internal representations of DNNs.