Search Results (283)

We present a deep learning model for the recognition of speech commands in the English language. The dataset is based on the Google Speech Commands Dataset by Warden P., version 0.01, and it consists of ten distinct commands (“left”, “right”, “go”, “stop”, “up”, “down”, “on”, “off”, “yes”, and “no”) along with additional “silence” and “unknown” classes. The dataset is split in a speaker-independent manner, with 70% of speakers assigned to the training set and 15% to the test set and validation set. All audio clips are sampled at 16 kHz, with a total of 46 146 clips. Audio files are converted into Mel spectrogram representations, which are then used as input to a deep learning model composed of a four-layer convolutional neural network followed by two fully connected layers. The model employs Rectified Linear Unit (ReLU) activation, the Adam optimizer, and dropout regularization to improve generalization. The achieved testing accuracy is 96.05%. Micro- and macro-averaged precision, recall, and F1-score of 95% are reported to reflect class-wise performance, and a confusion matrix is also provided. The proposed model has been deployed on a Raspberry Pi 5 as a Fog computing device for real-time speech recognition applications. Full article

Background: Titanium dental implants are widely used, but their long-term mechanical reliability under fatigue loading remains a key concern. Traditional finite element analysis is accurate but computationally intensive. This study explores the integration of finite element analysis data with neural networks to predict fatigue-related responses efficiently. Methods: A dataset of 200 finite element analysis simulations was generated, varying load intensity, load angle, and implant size. Each simulation provided three outputs: maximum von Mises stress, maximum displacement, and fatigue safety factor. A feedforward neural network with two hidden layers (64 neurons each, ReLU activation) was trained using 160 simulations, with 40 reserved for testing. Results: The neural network achieved high accuracy across all outputs, with R² values of 0.97 for stress, 0.95 for deformation, and 0.92 for the fatigue safety factor. Mean errors across the test set were below 5%, indicating strong predictive performance under diverse conditions. Conclusions: The findings demonstrate that neural networks can reliably replicate finite element analysis outcomes with significantly reduced computational time. This approach offers a promising tool for accelerating implant assessment and supports the growing role of AI in biomechanical design and analysis. Full article

Deep learning-based water body extraction methods generally focus on maximizing accuracy while neglecting inference speed, which can make them challenging to apply in real-time applications. To address this problem, this paper proposes a lightweight u-shaped network (LU-Net), which improves inference speed while maintaining comparable accuracy. To reduce inference latency, a lightweight decoder block (LDB) is designed, which employs a depthwise separable convolution structure to accelerate the decoding process. To enhance accuracy, a lightweight convolutional block attention module (LCBAM) is designed, which effectively captures water-specific spectral and spatial characteristics through a dual-attention mechanism. To improve multi-scale water boundary extraction, a structurally re-parameterized multi-scale fusion prediction module (SRMFPM) is designed, which integrates multi-scale water boundary information through convolutions of different sizes. Comparative experiments are conducted on the GID and LoveDA datasets, with model performance assessed using the MIoU metric and inference latency. The results demonstrate that LU-Net achieves the lowest GPU latency of 3.1 MS and the second-lowest CPU latency of 36 MS in the experiments. On the GID, LU-Net achieves the MIoU of 91.36%, outperforming other tested methods. On the LoveDA datasets, LU-Net achieves the second-highest MIoU of 86.32% among the evaluated models, which is 0.08% lower than the top-performing CGNet. Considering both latency and MIoU, LU-Net demonstrates commendable efficiency on the GID and LoveDA datasets across all compared networks. Full article

The increasing difficulty of handling industrial and agricultural wastes has generated interest in reusing materials such as Cement Kiln Dust (CKD) and Rice Husk Ash (RHA) for sustainable soil stabilization. This study examined the enhancement of lateritic soil with the incorporation of CKD (0–12%) and RHA (0–25%) by weight. An integrated experimental and Artificial Neural Network (ANN) methodology was utilized to evaluate and forecast geotechnical features. Laboratory assessments were conducted to measure Atterberg limits, Maximum Dry Density (MDD), Optimum Moisture Content (OMC), and Unconfined Compressive Strength (UCS) at 0, 7, and 28 days of curing. The results indicated significant enhancements in soil characteristics with CKD-RHA combinations. Artificial Neural Network models, including GELU, LOGSIG-3, and Leaky ReLU activation functions, accurately predicted the UCS, MDD, and OMC, achieving R² values as high as 0.980. This work underscores the efficacy of CKD-RHA mixtures in improving soil stability and the promise of ANN models as excellent prediction instruments, fostering sustainable and economical construction methodologies. Full article

To improve the performance of image super-resolution reconstruction, this paper optimizes the classical SRGAN model architecture. The original SRResNet is replaced with the EDSR network as the generator, which effectively enhances the ability to restore image details. To address the issue of insufficient multi-scale feature extraction in SRGAN during image reconstruction, an LSK attention mechanism is introduced into the generator. By fusing features from different receptive fields through parallel multi-scale convolution kernels, the model improves its ability to capture key details. To mitigate the instability and overfitting problems in the discriminator training, the Mish activation function is used instead of LeakyReLU to improve gradient flow, and a Dropout layer is introduced to enhance the discriminator’s generalization ability, preventing overfitting to the generator. Additionally, a staged training strategy is employed during adversarial training. Experimental results show that the improved model effectively enhances image reconstruction quality while maintaining low complexity. The generated results exhibit clearer details and more natural visual effects. On the public datasets Set5, Set14, and BSD100, compared to the original SRGAN, the PSNR and SSIM metrics improved by 13.4% and 5.9%, 9.9% and 6.0%, and 6.8% and 5.8%, respectively, significantly enhancing the reconstruction of super-resolution images, achieving more refined and realistic image quality improvement. The model also demonstrates stronger generalization ability on complex cross-domain data, such as remote sensing images and medical images. The improved model achieves higher-quality image reconstruction and more natural visual effects while maintaining moderate computational overhead, validating the effectiveness of the proposed improvements. Full article

This study proposes a hybrid deep learning framework for predicting vehicle crash frequency (F_i) using nationwide traffic accident and traffic volume data from the United States (2019–2022). Crash frequency is defined as the product of exposure frequency ($N_a$) and crash risk rate ($\lambda$), a structure widely adopted for its ability to separate physical exposure from the crash likelihood. $N_a$ was computed using an extended Safety Performance Function (SPF) that incorporates roadway traffic volume, segment length, number of lanes, and traffic density, while $\lambda$ was estimated using a multilayer perceptron-based deep neural network (DNN) with inputs such as impact speed, road surface condition, and vehicle characteristics. The DNN integrates rectified linear unit (ReLU) activation, batch normalization, dropout layers, and the Huber loss function to capture nonlinearity and over-dispersion beyond the capability of traditional statistical models. Model performance, evaluated through five-fold cross-validation, achieved R² = 0.7482, MAE = 0.1242, and MSE = 0.0485, demonstrating a strong capability to identify high-risk areas. Compared to traditional regression approaches such as Poisson and negative binomial models, which are often constrained by equidispersion assumptions and limited flexibility in capturing nonlinear effects, the proposed framework demonstrated substantially improved predictive accuracy and robustness. Unlike prior studies that loosely combined SPF terms with machine learning, this study explicitly decomposes F_i into $N_a$ and $\lambda$, ensuring interpretability while leveraging DNN flexibility for crash risk estimation. This dual-layer integration provides a unique methodological contribution by jointly achieving interpretability and predictive robustness, validated with a nationwide dataset, and highlights its potential for evidence-based traffic safety assessments and policy development. Full article

In modern industry, the spring full-lift safety valve is a key device for safe pressure relief of pressure-bearing systems. Its valve seat sealing surface is easily damaged after long-term use, causing internal leakage, resulting in safety hazards and economic losses. Therefore, it is of great significance to quickly and accurately diagnose its internal leakage state. Among the current methods for identifying fluid machinery faults, model-based methods have difficulties in parameter determination. Although the data-driven convolutional neural network (CNN) has great potential in the field of fault diagnosis, it has problems such as hyperparameter selection relying on experience, insufficient capture of time series and multi-scale features, and lack of research on valve internal leakage type identification. To this end, this study proposes a safety valve internal leakage identification method based on high-frequency FPGA data acquisition and improved CNN. The acoustic emission signals of different internal leakage states are obtained through the high-frequency FPGA acquisition system, and the two-dimensional time–frequency diagram is obtained by short-time Fourier transform and input into the improved model. The model uses the leaky rectified linear unit (LReLU) activation function to enhance nonlinear expression, introduces random pooling to prevent overfitting, optimizes hyperparameters with the help of horned lizard optimization algorithm (HLOA), and integrates the bidirectional gated recurrent unit (BiGRU) and selective kernel attention module (SKAM) to enhance temporal feature extraction and multi-scale feature capture. Experiments show that the average recognition accuracy of the model for the internal leakage state of the safety valve is 99.7%, which is better than the comparison model such as ResNet-18. This method provides an effective solution for the diagnosis of internal leakage of safety valves, and the signal conversion method can be extended to the fault diagnosis of other mechanical equipment. In the future, we will explore the fusion of lightweight networks and multi-source data to improve real-time and robustness. Full article

Recent progress in computer vision and embedded systems has facilitated real-time monitoring of bioprocesses; however, lightweight and scalable solutions for resource-constrained settings remain limited. This work presents a modular framework for monitoring Chlorella vulgaris growth by integrating RGB image processing with multimodal sensor fusion. The system incorporates a Logitech C920 camera and low-cost pH and temperature sensors within a compact photobioreactor. It extracts RGB channel statistics, luminance, and environmental data to generate a 10-dimensional feature vector. A feedforward artificial neural network (ANN) with ReLU activations, dropout layers, and SMOTE-based data balancing was trained to classify growth phases: lag, exponential, and stationary. The optimized model, quantized to 8 bits, was deployed on an ESP32 microcontroller, achieving 98.62% accuracy with 4.8 ms inference time and a 13.48 kB memory footprint. Robustness analysis confirmed tolerance to geometric transformations, though variable lighting reduced performance. Principal component analysis (PCA) retained 95% variance, supporting the discriminative power of the features. The proposed system outperformed previous vision-only methods, demonstrating the advantages of multimodal fusion for early detection. Limitations include sensitivity to lighting and validation limited to a single species. Future directions include incorporating active lighting control and extending the model to multi-species classification for broader applicability. Full article

Erythrina edulis, commonly known as pajuro, is a large leguminous plant native to the Amazon region of Peru. Its seeds are valued for their high protein content and their potential to enhance food security in rural communities. However, the current methods of harvesting and sorting are entirely manual, making the process labor-intensive, time-consuming, and subject to high variability, particularly in industrial contexts. A custom lightweight convolutional neural network (CNN) was developed from scratch and optimized specifically for real-time execution on embedded hardware. The model employs ReLU activation, Adam optimization, and a SoftMax output layer to enable efficient and accurate classification. The system employs a fixed-region segmentation strategy to prevent overcounting and utilizes GPIO-based control on a Raspberry Pi 5 to synchronize seed classification with physical sorting in real time. Seeds identified as defective are automatically removed via a servo-controlled ejection mechanism. The integrated system combines object detection, image processing, and real-time actuation, achieving a classification accuracy exceeding 99.6% and an average processing time of 12.4 milliseconds per seed. The proposed solution contributes to the industrial automation of pajuro sorting and provides a scalable framework for color-based grain classification applicable to a wide range of agricultural products. Full article

The compressive strength (Fc) of cement mortar (CM) is a key parameter in ensuring the mechanical reliability and durability of cement-based materials. Traditional testing methods are labor-intensive, time-consuming, and often lack predictive flexibility. With the increasing adoption of machine learning (ML) in civil engineering, data-driven approaches offer a rapid, cost-effective alternative for forecasting material properties. This study investigates a wide range of supervised linear and nonlinear ML regression models to predict the Fc of CM. The evaluated models include linear regression, ridge regression, lasso regression, decision trees, random forests, gradient boosting, k-nearest neighbors (KNN), and twelve neural network (NN) architectures, developed by combining different optimizers (L-BFGS, Adam, and SGD) with activation functions (tanh, relu, logistic, and identity). Model performance was assessed using the root mean squared error (RMSE), coefficient of determination (R²), and mean absolute error (MAE). Among all models, NN_tanh_lbfgs achieved the best results, with an almost perfect fit in training (R² = 0.9999, RMSE = 0.0083, MAE = 0.0063) and excellent generalization in testing (R² = 0.9946, RMSE = 1.5032, MAE = 1.2545). NN_logistic_lbfgs, gradient boosting, and NN_relu_lbfgs also exhibited high predictive accuracy and robustness. The SHAP analysis revealed that curing age and nano silica/cement ratio (NS/C) positively influence Fc, while porosity has the strongest negative impact. The main novelty of this study lies in the systematic tuning of neural networks via distinct optimizer–activation combinations, and the integration of SHAP for interpretability—bridging the gap between predictive performance and explainability in cementitious materials research. These results confirm the NN_tanh_lbfgs as a highly reliable model for estimating Fc in CM, offering a robust, interpretable, and scalable solution for data-driven strength prediction. Full article

Background/Objectives: Parkinson’s disease (PD) is a progressive neurodegenerative disorder that requires early and accurate diagnosis. This study aimed to classify individuals with and without PD using volumetric brain MRI data and to improve model interpretability using explainable artificial intelligence (XAI) techniques. Methods: This retrospective study included 79 participants (39 PD patients, 40 controls) recruited at Inonu University Turgut Ozal Medical Center between 2013 and 2025. A deep neural network (DNN) was developed using a multilayer perceptron architecture with six hidden layers and ReLU activation functions. Seventeen volumetric brain features were used as the input. To ensure robust evaluation and prevent overfitting, a stratified five-fold cross-validation was applied, maintaining class balance in each fold. Model transparency was explored using two complementary XAI techniques: the Contrastive Explanation Method (CEM) and Local Interpretable Model-Agnostic Explanations (LIME). CEM highlights features that support or could alter the current classification, while LIME provides instance-based feature attributions. Results: The DNN model achieved high diagnostic performance with 94.1% accuracy, 98.3% specificity, 90.2% sensitivity, and an AUC of 0.97. The CEM analysis suggested that reduced hippocampal volume was a key contributor to PD classification (–0.156 PP), whereas higher volumes in the brainstem and hippocampus were associated with the control class (+0.035 and +0.150 PP, respectively). The LIME results aligned with these findings, revealing consistent feature importance (mean = 0.1945) and faithfulness (0.0269). Comparative analyses showed different volumetric patterns between groups and confirmed the DNN’s superiority over conventional machine learning models such as SVM, logistic regression, KNN, and AdaBoost. Conclusions: This study demonstrates that a deep learning model, enhanced with CEM and LIME, can provide both high diagnostic accuracy and interpretable insights for PD classification, supporting the integration of explainable AI in clinical neuroimaging. Full article

Micro-motors are essential for power drive systems, and efficient fault diagnosis is crucial to reduce safety risks and economic losses caused by failures. However, the fault signals from micro-motors typically exhibit weak and unclear characteristics. To address this challenge, this paper proposes a novel fault diagnosis method that integrates jump plus AM-FM mode decomposition (JMD), symmetrized dot pattern (SDP) visualization, and an improved convolutional neural network (ICNN). Firstly, we employed the jump plus AM-FM mode decomposition technique to decompose the mixed fault signals, addressing the problem of mode mixing in traditional decomposition methods. Then, the intrinsic mode functions (IMFs) decomposed by JMD serve as the multi-channel inputs for symmetrized dot pattern, constructing a two-dimensional polar coordinate petal image. This process achieves both signal reconstruction and visual enhancement of fault features simultaneously. Finally, this paper designed an ICNN method with LeakyReLU activation function to address the vanishing gradient problem and enhance classification accuracy and training efficiency for fault diagnosis. Experimental results indicate that the proposed JMD-SDP-ICNN method outperforms traditional methods with a significantly superior fault classification accuracy of up to 99.2381%. It can offer a potential solution for the monitoring of electromechanical structures under complex conditions. Full article

Image super-resolution (SR) is essential for enhancing image quality in critical applications, such as medical imaging and satellite remote sensing. However, existing methods were often limited in their ability to effectively process and integrate multi-scales information from fine textures to global structures. To address these limitations, this paper proposes DSCNN, a dynamic snake convolution neural network for enhanced image super-resolution. DSCNN optimizes feature extraction and network architecture to enhance both performance and efficiency: To improve feature extraction, the core innovation is a feature extraction and enhancement module with dynamic snake convolution that dynamically adjusts the convolution kernel’s shape and position to better fit the image’s geometric structures, significantly improving feature extraction. To optimize the network’s structure, DSCNN employs an enhanced residual network framework. This framework utilizes parallel convolutional layers and a global feature fusion mechanism to further strengthen feature extraction capability and gradient flow efficiency. Additionally, the network incorporates a SwishReLU-based activation function and a multi-scale convolutional concatenation structure. This multi-scale design effectively captures both local details and global image structure, enhancing SR reconstruction. In summary, the proposed DSCNN outperforms existing methods in both objective metrics and visual perception (e.g., our method achieved optimal PSNR and SSIM results on the Set5 $\times 4$ dataset). Full article

Unexpected machine failures in industrial environments lead to high maintenance costs, unplanned downtime, and safety risks. This study proposes a proactive predictive model using a hybrid of eXtreme Gradient Boosting (XGBoost) and Neural Networks (NN) to forecast machine failures. A synthetic dataset capturing recent breakdown history and time since last failure was used to simulate industrial scenarios. To address class imbalance, SMOTE and class weighting were applied, alongside a focal loss function to emphasize difficult-to-classify failures. The XGBoost model was tuned via GridSearchCV, while the NN model utilized ReLU-activated hidden layers with dropout. Evaluation using stratified 5-fold cross-validation showed that the NN achieved an F1-score of 0.7199 and a recall of 0.9545 for the minority class. XGBoost attained a higher PR AUC of 0.7126 and a more balanced precision–recall trade-off. Sample predictions demonstrated strong recall (100%) for failures, but also a high false positive rate, with most prediction probabilities clustered between 0.50–0.55. Additional benchmarking against Logistic Regression, Random Forest, and SVM further confirmed the superiority of the proposed hybrid model. Model interpretability was enhanced using SHAP and LIME, confirming that recent breakdowns and time since last failure were key predictors. While the model effectively detects failures, further improvements in feature engineering and threshold tuning are recommended to reduce false alarms and boost decision confidence. Full article

Weed infestation contributes significantly to global agricultural yield loss and increases the reliance on herbicides, raising both economic and environmental concerns. Effective weed detection in agriculture requires high accuracy and architectural efficiency. This is particularly important under challenging field conditions, including densely clustered targets, small weed instances, and low visual contrast between vegetation and soil. In this study, we propose GTDR-YOLOv12, an improved object detection framework based on YOLOv12, tailored for real-time weed identification in complex agricultural environments. The model is evaluated on the publicly available Weeds Detection dataset, which contains a wide range of weed species and challenging visual scenarios. To achieve better accuracy and efficiency, GTDR-YOLOv12 introduces several targeted structural enhancements. The backbone incorporates GDR-Conv, which integrates Ghost convolution and Dynamic ReLU (DyReLU) to improve early-stage feature representation while reducing redundancy. The GTDR-C3 module combines GDR-Conv with Task-Dependent Attention Mechanisms (TDAMs), allowing the network to adaptively refine spatial features critical for accurate weed identification and localization. In addition, the Lookahead optimizer is employed during training to improve convergence efficiency and reduce computational overhead, thereby contributing to the model’s lightweight design. GTDR-YOLOv12 outperforms several representative detectors, including YOLOv7, YOLOv9, YOLOv10, YOLOv11, YOLOv12, ATSS, RTMDet and Double-Head. Compared with YOLOv12, GTDR-YOLOv12 achieves notable improvements across multiple evaluation metrics. Precision increases from 85.0% to 88.0%, recall from 79.7% to 83.9%, and F1-score from 82.3% to 85.9%. In terms of detection accuracy, mAP:0.5 improves from 87.0% to 90.0%, while mAP:0.5:0.95 rises from 58.0% to 63.8%. Furthermore, the model reduces computational complexity. GFLOPs drop from 5.8 to 4.8, and the number of parameters is reduced from 2.51 M to 2.23 M. These reductions reflect a more efficient network design that not only lowers model complexity but also enhances detection performance. With a throughput of 58 FPS on the NVIDIA Jetson AGX Xavier, GTDR-YOLOv12 proves both resource-efficient and deployable for practical, real-time weeding tasks in agricultural settings. Full article