Smaller Mean Absolute Error Research Articles

Integrating Radial Basis Networks and Deep Learning for Transportation

Introduction This research focuses on the concept of integrating Radial Basis Function Networks with deep learning models to solve robust regression tasks in both transportation and logistics. Methods It examines such combined models as RNNs with RBFNs, Attention Mechanisms with Radial Basis Function Networks (RBFNs), and Capsule Networks with RBFNs and clearly shows that, in all cases, compared to the others, the former model has a Mean Squared Error (MSE) of 0.010 to 0.013, Mean Absolute Error (MAE) – 0.078 to 0.088, and R-squared (R2) – 0.928 to 0.945, across ten experiments. In the case of Attention Mechanisms with RBFNs, the models also demonstrate strong performance in terms of making predictions. The MSE ranges from 0.012 to 0.015, the MAE from 0.086 to 0.095, and the R2 from 0.914 to 0.933. Results However, it is critical to note that the Capsule Networks with RBFNs outperform other models. In particular, they offer the lowest MSE, which is between 0.009 and 0.012, the smallest MAE, which ranges from 0.075 to 0.083, and the highest R2, from 0.935 to 0.950. Conclusion Overall, the results indicate that the use of RBFNs in combination with different types of deep learning networks can provide highly accurate and reliable solutions for regression problems in the domain of transportation and logistics.

Convolutional variational autoencoder and multi-scale attention convolutional neural network based diagnostics on filament current sensors for mass spectrometers

Fault detection of the filament current sensor of the spaceborne mass spectrometer is of great significance to the safe operation of the mass spectrometer. Neglecting sensor failures may affect the normal operation of the mass spectrometer, which in turn affects the health of the astronauts and the space missions. This paper proposes an enhanced intelligent diagnosis method based on filament current sensor for mass spectrometer via improved deep learning network, aiming to improve the accuracy and reliability of the filament current sensor when the fault samples are insufficient. The improved convolutional variational autoencoder (CVAE) model not only enhances the data for four typical sensor fault signals, namely bias, drift, missing and random, but also solves the problem of insufficient samples when spike, Precision degradation and stuck fault occur in sensors. Short-time Fourier transform (STFT) is utilized to transform one-dimensional data into two-dimensional spectrograms, which gives more fault characteristics to the data set. The improved multi-scale attention mechanism convolutional neural network (MSAM-CNN) model is constructed to perform fault diagnosis on the generated spectrogram, which solves the problem of low accuracy of fault identification in traditional convolutional neural network (CNN) models. The results of ablation and comparison experiments show that the samples generated by CVAE have smaller Mean Absolute Error (MAE), Mean Square Error (MSE), and Root Mean Square Error (RMSE), the enhanced samples improve the classification and clustering of MSAM-CNN, and compared to other diagnostic models, MSAM-CNN achieves the highest accuracy, precision, recall, and F1-score of 99.2%, 99.3%, 98.8%, and 0.991.

Heat Stress Metrics, Trends, and Extremes in the Southeastern United States

Abstract Humid heat and associated heat stress have increased in frequency, intensity, and duration across the globe, particularly at lower latitudes. One of the more robust metrics for heat stress impacts on the human body is wet-bulb globe temperature (WBGT), because it incorporates temperature, humidity, wind speed, and solar radiation. WBGT can typically only be measured using nonstandard instrumentation (e.g., black globe thermometers). However, estimation formulas have been developed to calculate WBGT using standard surface meteorological variables. This study evaluates several WBGT estimation formulas for the southeastern United States using North Carolina Environment and Climate Observing Network (ECONet) and U.S. Military measurement campaign data as verification. The estimation algorithm with the smallest mean absolute error was subsequently chosen to evaluate summer WBGT trends and extremes at 39 ASOS stations with long continuous (1950–2023) data records. Trend results showed that summer WBGT has increased throughout much of the southeastern United States, with larger increases at night than during the day. Although there were some surprisingly large WBGT trends at higher elevation locations far from coastlines, the greatest increases were predominantly located in the Florida Peninsula and Louisiana. Increases in the intensity and frequency of extreme (90th percentile) WBGTs were particularly stark in large coastal urban centers (e.g., New Orleans, Tampa, and Miami). Some locations like New Orleans and Tampa have experienced more than two additional extreme heat stress days and nights per decade since 1950, with an exponential escalation in the number of extreme summer nights during the most recent decade. Significance Statement Humid heat and associated heat stress pose threats to health in the moist subtropical climate of the southeastern United States. Wet-bulb globe temperature (WBGT) is a robust metric for heat stress but must be estimated using complex algorithms. We first evaluated the accuracy of three WBGT algorithms in the southeastern United States, using measured verification data. Subsequently, we used the most accurate algorithm to investigate WBGT trends and extremes since 1950 in 39 cities. Results showed that summer heat stress has increased throughout the region, especially at night. Increases in the intensity and frequency of extreme heat stress were most prevalent at urban coastal locations in Florida and Louisiana, emphasizing the impacts of increased urbanization and evaporation on heat stress.

A framework for modelling whole-lung and regional TLCOusing hyperpolarised129Xe lung MRI

BackgroundPulmonary gas exchange is assessed by the transfer factor of the lungs (TL) for carbon monoxide (TLCO), and can also be measured with inhaled xenon-129 (129Xe) MRI. A model has been proposed to estimate TL from129Xe MRI metrics, but this approach has not been fully validated and does not utilise the spatial information provided by 3D129Xe MRI.MethodsThree models for predicting TL from129Xe MRI metrics were compared; (1) a previously-published physiology-based model, (2) multivariable linear regression and (3) random forest regression. Models were trained on data from 150 patients with asthma and/or chronic obstructive pulmonary disease. The random forest model was applied voxel-wise to129Xe images to yield regional TL maps.ResultsCoefficients of the physiological model were found to differ from previously reported values. All models had good prediction accuracy with small mean absolute error (MAE); (1) 1.24±0.15 mmol·min−1·kPa−1, (2) 1.01±0.06 mmol·min−1·kPa−1, (3) 0.995±0.129 mmol·min−1·kPa−1. The random forest model performed well when applied to a validation group of post-COVID-19 patients and healthy volunteers (MAE=0.840 mmol·min−1·kPa−1), suggesting good generalisability. The feasibility of producing regional maps of predicted TL was demonstrated and the whole-lung sum of the TL maps agreed with measured TLCO(MAE=1.18 mmol·min−1·kPa−1).ConclusionThe best prediction of TLCOfrom129Xe MRI metrics was with a random forest regression framework. Applying this model on a voxel-wise level to create parametric TL maps provides a useful tool for regional visualisation and clinical interpretation of129Xe gas exchange MRI.

Taylor‐based smart flower optimization algorithm with the deep residual network to predict mechanical materials properties

AbstractThe expedience of materials processing is of great significance and increased the industrial interest in meeting the needs of contemporary engineering applications. The inspection of mechanical properties is extensively explored by scientists, but the prediction of properties with the deep model is limited. This article presents an optimized deep residual network (DRN) to predict mechanical properties of materials. The quantile normalization is applied for improved processing. The DRN is pre‐trained with an optimization model for initializing the best set of attributes and tuning the parameters of the model. Here, Taylor‐Smart Flower Optimization Algorithm (Taylor‐SFOA) is adapted for training DRN by tuning optimum weights. The proposed Taylor‐SFOA helps to effectively offer precise mapping amidst mechanical properties and processing parameters. The optimal features are selected with the Ruzicka and Motyka. The selected features are fused with a dice coefficient to choose distinct features for attaining effective performance. The method yielded better outcomes with improved generalization. The Taylor‐SFOA‐based DRN provided better outcomes with smallest Mean absolute error (MAE) of 0.049, Mean square error (MSE) of 0.116, Root Mean square error (RMSE) of 0.340, memory footprint of 37.700 MB, and training time of 9.633 Sec.

An improved limited memory-Sage Husa-cubature Kalman filtering algorithm for the state of charge and state of energy co-estimation of lithium-ion batteries based on hysteresis effect-dual polarization model

The accurate estimation of battery state of charge (SOC) and state of energy (SOE) are the key technologies in researching battery management systems. Due to the uncertainty of prior noise in practical engineering, traditional cubature Kalman filters (CKF) have lower filtering accuracy. This paper proposes an improved limited memory-Sage Husa-cubature Kalman (LM–SH–CKF) algorithm to estimate the SOC and SOE of lithium-ion batteries. During the filtering process, the Sage-Husa estimator is introduced to estimate the noise, correcting the statistical characteristics of the noise in real-time, thereby improving the estimation accuracy. Meanwhile, establish a filtering divergence criterion to determine whether the filtering is abnormal and ensure the real-time performance of the algorithm. To improve the accuracy of parameter identification, the Hysteresis Effect-Dual Polarization (HE-DP) model is established and the limited memory recursive least square (LMRLS) algorithm is proposed. The results show that the LM–SH–CKF has the smallest mean absolute error (MAE) and root mean square error (RMSE). Under the dynamic stress test (DST) at 15 °C, the estimated MAE and RMSE of SOC and SOE are both less than 1 %. The LM–SH–CKF has high accuracy and robustness, and short calculation time, providing a reference for battery status monitoring.

Facilitating self-directed language learning in real-life scene description tasks with automated evaluation

Engaging children in describing real-life scenes provides an effective approach to fostering language production and developing their language skills, enabling them to establish meaningful connections between their language proficiency and authentic contexts. However, for such learning tasks, there has been a lack of research focusing on promoting self-directed language learning by using artificial intelligence techniques, primarily due to the challenges of handling multimodal information involved in such tasks. To address this gap, this study introduced a two-stage automated evaluation method that employed emerging cross-modal matching AI techniques. Firstly, an automated scoring model was developed to evaluate the quality of students' responses to scene description tasks. Compared with manually assigned human scores, our model scored students' descriptions accurately, as evidenced by a small testing mean absolute error of 0.3969 for a total score of 10 points. Based on the scoring results, immediate feedback was then provided to students by generating targeted comments and suggestions. The goal of this feedback was to assist students in progressively improving their descriptions of daily-life scenes, thereby enabling them to practice their language skills independently. To assess the effectiveness of the feedback, a comprehensive investigation was conducted involving 157 students from middle schools in China, and both qualitative and quantitative experimental data were collected from the students. It is found that the quality of students' descriptions was improved significantly with the assistance of immediate feedback. On average, students achieved an increase of 1.48 points in their scores after making revisions based on the feedback. In addition, students reported positive learning experiences and expressed favorable opinions regarding the language learning tasks with the automated evaluation. The findings of this study have significant implications for future research and educational practice. They not only highlighted the potential of emerging cross-modal matching AI techniques in automatically evaluating learning tasks involving multimodal data but also suggested that providing immediate targeted feedback based on automated scoring results can effectively promote students’ self-directed language learning.

Structural seismic responses prediction using the gradient-enhanced hybrid PINN

To rapidly and effectively assess the bridge seismic-resistant capability, it is essential to conduct efficient predictions of bridge seismic responses. Recently, physics informed neural network (PINN) has made great progress and utilized to solve differential equations in different fields. However, how to increase its accuracy and efficiency still remains an open challenge. In this work, a novel gradient-enhanced Fourth-Order Runge-Kutta PINN (gRK4-PINN), as a powerful hybrid PINN, is utilized to achieve this goal. As for gRK4-PINN, the physical information is not simply embedded into the loss function; instead, the RK4 method and the physical model is intricately integrated with the neural network. In addition, to improve the predictive performance, additional gradient equation is directly embedded in loss function. A large-span continuous girder high speed railway (CGHSR) bridge is adopted as numerical experiment to validate the fidelity of the proposed method. Results reveal that the Mean Absolute Error (MAE) of the predicting seismic responses is relatively small, whose value is below 0.014 in most of the time. These small MAE values indicate that the proposed gRK4-PINN performs well in predicting the seismic responses of the CGHSR bridge.

Accuracy of 14 intraocular lens power calculation formulas in extremely long eyes.

To compare the accuracy of 14 formulas in calculating intraocular lens (IOL) power in extremely long eyes with axial length (AL) over 30.0mm. In this retrospective study, 211 eyes (211 patients) with ALs > 30.0mm were successfully treated with cataract surgery without complications. Ocular biometric parameters were obtained from IOLMaster 700. Fourteen formulas were evaluated using the optimized A constants: Barrett Universal II (BUII), Kane, Emmetropia Verifying Optical (EVO) 2.0, PEARL-DGS, T2, SRK/T, Holladay 1, Holladay 2, Haigis and Wang-Koch AL adjusted formulas (SRK/Tmodified-W/K, Holladay 1modified-W/K, Holladay 1NP-modified-W/K, Holladay 2modified-W/K, Holladay 2NP-modified-W/K). The mean prediction error (PE) and standard deviation (SD), mean absolute errors (MAE), median absolute errors (MedAE), and the percentage of prediction errors (PEs) within ± 0.25 D, ± 0.50 D, ± 1.00 D were analyzed. The Kane formula had the smallest MAE (0.43 D) and MedAE (0.34 D). The highest percentage of PE within ± 0.25 D was for EVO 2.0 (37.91%) and the Holladay 1NP-modified-W/K formulas (37.91%). The Kane formula had the highest percentage of PEs in the range of ± 0.50, ± 0.75, ± 1.00, and ± 2.00 D. There was no significant difference in PEs within ± 0.25, ± 0.50 ± 0.75 and ± 1.00 D between BUII, Kane, EVO 2.0 and Wang-Koch AL adjusted formulas (P > .05) by using Cochran's Q test. The Holladay 2modified-W/K formula has the lowest percentage of hyperopic outcomes (29.38%). The BUII, Kane, EVO 2.0 and Wang-Koch AL adjusted formulas have comparable accuracy for IOL power calculation in eyes with ALs > 30.0mm.

Accuracy of Toric Intraocular Lens Formulas With Measured Posterior Corneal Astigmatism of Different Orientations

PurposeTo assess whether the use of measured posterior corneal astigmatism (PCA) values improves the prediction accuracy of toric intraocular lens power formulas, compared to predicted PCA values, when the orientation of the steep axis of PCA is non-vertical. DesignRetrospective observational cohort study Methods418 eyes of 344 patients were included in the study. Prediction errors (PE) for postoperative refractive astigmatism at 4 weeks postoperatively were determined using vector analysis and compared for the following toric intraocular lens power formulas: Barrett Toric with predicted posterior corneal astigmatism (PPCA); Barrett Toric with measured posterior corneal astigmatism (MPCA); EVO Toric PPCA; EVO Toric MPCA; Holladay I with Abulafia-Koch regression. Subgroup analysis compared PEs for eyes with a vertically orientated steep axis of PCA (60-120o) to eyes with a non-vertically orientated steep axis of PCA. SettingCathedral Eye Clinic, Belfast, United Kingdom and Tan Tock Seng Hospital, Singapoore. ResultsStandard keratometry was with-the-rule in 48% of eyes, while the steep PCA axis was vertically orientated in 91% of eyes. For all eyes, EVO PPCA had a smaller mean absolute error than Barrett-MPCA, Barrett-PPCA and Abulafia-Koch (p < 0.01 for all). EVO-PPCA had the highest percentage of eyes within 0.50D of predicted postoperative astigmatism for eyes with vertical PCA (61%), while EVO-MPCA had the highest percentage for eyes with non-vertical PCA (54%). EVO-MPCA had the smallest centroid error for all eyes, and the subgroups (p < 0.01 for all). Eyes with non-vertical PCA had a lower percentage within 0.50D than eyes with vertical PCA when using PPCA (43% vs 61%, p = 0.034), but there was no significant difference between these groups when MPCA is used for eyes with non-vertical PCA (54% vs 61%, p = 0.40). ConclusionsWhen the steep axis of posterior corneal astigmatism is not vertically orientated, the use of measured posterior keratometry values improves prediction accuracy.

A Comparative Analysis of Machine Learning Algorithms for Predicting Paddy Production

For countries with large populations, such as Indonesia, food security is a very important issue. The majority of Indonesia's population depends on rice as their main food, and paddy is one of the most widely cultivated food commodities. The very good and accurate national paddy production prediction results really support decisions regarding national paddy production targets for the coming period. Therefore, to ensure supply and price stability, paddy availability must be predicted. Many studies have used machine learning to predict crop yields. By learning important patterns and relationships from input data, machine learning can combine the advantages of other methods to make better predictions of paddy yields. The aim of this research is to conduct a comparative analysis between three machine learning algorithms, namely, random forest, decision tree, and k-nearest neighbors, in predicting paddy production. To determine which algorithm is the best, a model evaluation is carried out using the coefficient of determination (R2-score), mean absolute error (MAE), and mean squared error (MSE). This research goes through methodological stages, starting from collecting datasets, data preprocessing, training and testing split datasets, applying algorithms, and evaluating the model. From this research, results were obtained for the random forest algorithm with an R2-score of 82.38%, MAE of 261726.20, and MSE of 2.19495E+11. For the decision tree, the R2-score was 79.62%, MAE was 323257.99, and MSE was 2.49304E+11. Meanwhile, k-nearest neighbors obtained an R2-score of 76.25%, MAE of 318433.42, and MSE of 2.90577E+11. The results of this research show that the random forest algorithm is the best for predicting paddy production because it obtains a larger R2-score as well as smaller MAE and MSE results.

Prediction method of stalk content in strip particles after redrying of tobacco leaves

AbstractFour tobacco leaf modules processed by Yunnan Tobacco Redrying Co., Ltd. during the 2022 roasting season were used to investigate the method of stalk content in strip particles after redrying of tobacco leaves, effectively reducing the loss of strip particles. A total of 151 sets of experimental data were used to construct the prediction model for the stalk content in strip particles after redrying using the BP artificial neural network method, the linear regression method, and the support vector machine method. The results show that the prediction model constructed by the BP artificial neural network method has high accuracy and stability, with a relatively small absolute error of prediction (e = 0.0195%) and the root-mean-square error of interactive verification (RMSECV = 0.0227%), as well as a relatively small mean absolute error of production data validation (e = 0.0675%), while the prediction deviation ratio (RPD = 2.2435) is relatively large. Overall, the prediction model established by BP artificial neural network could provide new insight into the non-destructive detection of stalk content in strip particles of redried tobacco leaves after threshing and redrying and potentially leading to a reduction in tobacco leaf crushing by more than 112,500 kg per year.

Machine learning approach for predicting post-intubation hemodynamic instability (PIHI) index values: towards enhanced perioperative anesthesia quality and safety

BackgroundAdequate preoperative evaluation of the post-intubation hemodynamic instability (PIHI) is crucial for accurate risk assessment and efficient anesthesia management. However, the incorporation of this evaluation within a predictive framework have been insufficiently addressed and executed. This study aims to developed a machine learning approach for preoperatively and precisely predicting the PIHI index values.MethodsIn this retrospective study, the valid features were collected from 23,305 adult surgical patients at Peking Union Medical College Hospital between 2012 and 2020. Three hemodynamic response sequences including systolic pressure, diastolic pressure and heart rate, were utilized to design the post-intubation hemodynamic instability (PIHI) index by computing the integrated coefficient of variation (ICV) values. Different types of machine learning models were constructed to predict the ICV values, leveraging preoperative patient information and initiatory drug infusion. The models were trained and cross-validated based on balanced data using the SMOTETomek technique, and their performance was evaluated according to the mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE) and R-squared index (R2).ResultsThe ICV values were proved to be consistent with the anesthetists’ ratings with Spearman correlation coefficient of 0.877 (P < 0.001), affirming its capability to effectively capture the PIHI variations. The extra tree regression model outperformed the other models in predicting the ICV values with the smallest MAE (0.0512, 95% CI: 0.0511–0.0513), RMSE (0.0792, 95% CI: 0.0790–0.0794), and MAPE (0.2086, 95% CI: 0.2077–0.2095) and the largest R2 (0.9047, 95% CI: 0.9043–0.9052). It was found that the features of age and preoperative hemodynamic status were the most important features for accurately predicting the ICV values.ConclusionsOur results demonstrate the potential of the machine learning approach in predicting PIHI index values, thereby preoperatively informing anesthetists the possible anesthetic risk and enabling the implementation of individualized and precise anesthesia interventions.

Spatial heterogeneity of long-range dependence and self-similarity of global sea surface chlorophyll concentration with their environmental impact factors analysis

Understanding the long-range dependence and self-similarity of global sea surface chlorophyll concentration (SSCC) will enrich its characteristics description and analysis with global change patterns. The satellite SSCC products were collected from the European Space Agency during the period from 29 July 1998 to 31 December2020. After resampling the SSCC products into the spatial resolution of 1°, the missing values were interpolated by Bayesian maximum entropy with mean absolute error of cross validation equaling to 0.1295 mg/m3. Generalized Cauchy model was employed to quantitatively determine the long-range dependence and self-similarity of SSCC at a global scale by using the Hurst parameter and fractal dimension. Good fitted results were achieved with an averaged R2 of 0.9141 and a standard deviation of 0.0518 across the 32,281 spatial locations of the entire ocean; the averaged values of Hurst parameter and fractal dimension were 0.8667 and 1.2506, respectively, suggesting strong long-range dependence and weak self-similarity of SSCC in the entire oceans. Univariate and multivariate generalized addictive models (GAM) were introduced to depict the influence of sea surface height anomaly, sea surface salinity, sea surface temperature and sea surface wind on the Hurst parameter and fractal dimension of SSCC; and smaller mean absolute error were achieved for the GAM of Hurst parameter than that of fractal dimension. Sea surface height anomaly showed the strongest influence for the Hurst parameter than the other three factors, and sea surface wind depicted similar influence; the sea surface temperature owned opposite influence on Hurst parameter compared to sea surface salinity.

Accurate Electronic and Optical Properties of Organic Doublet Radicals Using Machine Learned Range-Separated Functionals.

Luminescent organic semiconducting doublet-spin radicals are unique and emergent optical materials because their fluorescent quantum yields (Φfl) are not compromised by the spin-flipping intersystem crossing (ISC) into a dark high-spin state. The multiconfigurational nature of these radicals challenges their electronic structure calculations in the framework of single-reference density functional theory (DFT) and introduces room for method improvement. In the present study, we extended our earlier development of ML-ωPBE [J. Phys. Chem. Lett., 2021, 12, 9516-9524], a range-separated hybrid (RSH) exchange-correlation (XC) functional constructed using the stacked ensemble machine learning (SEML) algorithm, from closed-shell organic semiconducting molecules to doublet-spin organic semiconducting radicals. We assessed its performance for a new test set of 64 doublet-spin radicals from five categories while placing all previously compiled 3926 closed-shell molecules in the new training set. Interestingly, ML-ωPBE agrees with the nonempirical OT-ωPBE functional regarding the prediction of the molecule-dependent range-separation parameter (ω), with a small mean absolute error (MAE) of 0.0197 a0-1, but saves the computational cost by 2.46 orders of magnitude. This result demonstrates an outstanding domain adaptation capacity of ML-ωPBE for diverse organic semiconducting species. To further assess the predictive power of ML-ωPBE in experimental observables, we also applied it to evaluate absorption and fluorescence energies (Eabs and Efl) using linear-response time-dependent DFT (TDDFT), and we compared its behavior with nine popular XC functionals. For most radicals, ML-ωPBE reproduces experimental measurements of Eabs and Efl with small MAEs of 0.299 and 0.254 eV, only marginally different from those of OT-ωPBE. Our work illustrates a successful extension of the SEML framework from closed-shell molecules to doublet-spin radicals and will open the venue for calculating optical properties for organic semiconductors using single-reference TDDFT.

Intelligently Quantifying the Entire Irregular Dental Structure

Quantitative analysis of irregular anatomical structures is crucial in oral medicine, but clinicians often typically measure only several representative indicators within the structure as references. Deep learning semantic segmentation offers the potential for entire quantitative analysis. However, challenges persist, including segmentation difficulties due to unclear boundaries and acquiring measurement landmarks for clinical needs in entire quantitative analysis. Taking the palatal alveolar bone as an example, we proposed an artificial intelligence measurement tool for the entire quantitative analysis of irregular dental structures. To expand the applicability, we have included lightweight networks with fewer parameters and lower computational demands. Our approach finally used the lightweight model LU-Net, addressing segmentation challenges caused by unclear boundaries through a compensation module. Additional enamel segmentation was conducted to establish a measurement coordinate system. Ultimately, we presented the entire quantitative information within the structure in a manner that meets clinical needs. The tool achieved excellent segmentation results, manifested by high Dice coefficients (0.934 and 0.949), intersection over union (0.888 and 0.907), and area under the curve (0.943 and 0.949) for palatal alveolar bone and enamel in the test set. In subsequent measurements, the tool visualizes the quantitative information within the target structure by scatter plots. When comparing the measurements against representative indicators, the tool’s measurement results show no statistically significant difference from the ground truth, with small mean absolute error, root mean squared error, and errors interval. Bland-Altman plots and intraclass correlation coefficients indicate the satisfactory agreement compared with manual measurements. We proposed a novel intelligent approach to address the entire quantitative analysis of irregular image structures in the clinical setting. This contributes to enabling clinicians to swiftly and comprehensively grasp structural features, facilitating the design of more personalized treatment plans for different patients, enhancing clinical efficiency and treatment success rates in turn.

Predicting voxel-level dose distributions of single-isocenter volumetric modulated arc therapy treatment plan for multiple brain metastases.

Brain metastasis is a common, life-threatening neurological problem for patients with cancer. Single-isocenter volumetric modulated arc therapy (VMAT) has been popularly used due to its highly conformal dose and short treatment time. Accurate prediction of its dose distribution can provide a general standard for evaluating the quality of treatment plan. In this study, a deep learning model is applied to the dose prediction of a single-isocenter VMAT treatment plan for radiotherapy of multiple brain metastases. A U-net with residual networks (U-ResNet) is employed for the task of dose prediction. The deep learning model is first trained from a database consisting of hundreds of historical treatment plans. The 3D dose distribution is then predicted with the input of the CT image and contours of regions of interest (ROIs). A total of 150 single-isocenter VMAT plans for multiple brain metastases are used for training and testing. The model performance is evaluated based on mean absolute error (MAE) and mean absolute differences of multiple dosimetric indexes (DIs), including (D max and D mean) for OARs, (D 98, D 95, D 50, and D 2) for PTVs, homogeneity index, and conformity index. The similarity between the predicted and clinically approved plan dose distribution is also evaluated. For 20 tested patients, the largest and smallest MAEs are 3.3% ± 3.6% and 1.3% ± 1.5%, respectively. The mean MAE for the 20 tested patients is 2.2% ± 0.7%. The mean absolute differences of D 98, D 95, D 50, and D2 for PTV60, PTV52, PTV50, and PTV40 are less than 2.5%, 3.0%, 2.0%, and 3.0%, respectively. The prediction accuracy of OARs for D max and D mean is within 3.2% and 1.2%, respectively. The average DSC ranges from 0.86 to 1 for all tested patients. U-ResNet is viable to produce accurate dose distribution that is comparable to those of the clinically approved treatment plans. The predicted results can be used to improve current treatment planning design, plan quality, efficiency, etc.

A Method for Predicting Ground Pressure in Meihuajing Coal Mine Based on Improved BP Neural Network by Immune Algorithm-Particle Swarm Optimization

Based on the background of dynamic mining pressure monitoring and pressure prediction research on the No. 232205 working face of the Meihuajing coal mine, this study systematically investigates the predictive model of mining pressure manifestation on the working face of the Meihuajing coal mine by integrating methods such as engineering investigation, theoretical analysis, and mathematical modeling. A mining pressure manifestation prediction method based on IA-PSO-BP is proposed. The IA-PSO optimization algorithm is applied to optimize the hyperparameters of the BP neural network, and the working face mining pressure prediction model based on IA-PSO-BP is established. The mean absolute error (MAE), mean square error (MSE), and coefficient of determination (R2) are selected as evaluation indicators to compare the prediction performance of the BP model, PSO-BP model, and IA-PSO-BP model. The experimental results of the model show that the convergence speed of the IA-PSO-BP model is about eight times faster than that of the BP model and two times faster than that of the PSO-BP model. Compared with the BP and PSO-BP models, the IA-PSO-BP model has the smallest MAE and MSE and the largest R2 on the three different data sets of the test set, indicating significantly improved prediction accuracy. The predicted results conform to the periodic variation pattern of mining pressure data and are consistent with the actual situation in the coal mine.

Deep learning model based on multi-scale feature fusion for precipitation nowcasting

Abstract. Forecasting heavy precipitation accurately is a challenging task for most deep learning (DL)-based models. To address this, we present a novel DL architecture called “multi-scale feature fusion” (MFF) that can forecast precipitation with a lead time of up to 3 h. The MFF model uses convolution kernels with varying sizes to create multi-scale receptive fields. This helps to capture the movement features of precipitation systems, such as their shape, movement direction, and speed. Additionally, the architecture utilizes the mechanism of discrete probability to reduce uncertainties and forecast errors, enabling it to predict heavy precipitation even at longer lead times. For model training, we use 4 years of radar echo data from 2018 to 2021 and 1 year of data from 2022 for model testing. We compare the MFF model with three existing extrapolative models: time series residual convolution (TSRC), optical flow (OF), and UNet. The results show that MFF achieves superior forecast skills with high probability of detection (POD), low false alarm rate (FAR), small mean absolute error (MAE), and high structural similarity index (SSIM). Notably, MFF can predict high-intensity precipitation fields at 3 h lead time, while the other three models cannot. Furthermore, MFF shows improvement in the smoothing effect of the forecast field, as observed from the results of radially averaged power spectral (RAPS). Our future work will focus on incorporating multi-source meteorological variables, making structural adjustments to the network, and combining them with numerical models to further improve the forecast skills of heavy precipitations at longer lead times.

A Universal Test on Spikes in a High-Dimensional Generalized Spiked Model and Its Applications

This paper aims to test the number of spikes in a generalized spiked covariance matrix, the spiked eigenvalues of which may be extremely larger or smaller than the non-spiked ones. For a high-dimensional problem, we first propose a general test statistic and derive its central limit theorem by random matrix theory without a Gaussian population constraint. We then apply the result to estimate the noise variance and test the equality of the smallest roots in generalized spiked models. Simulation studies showed that the proposed test method was correctly sized, and the power outcomes showed the robustness of our statistic to deviations from a Gaussian population. Moreover, our estimator of the noise variance resulted in much smaller mean absolute errors and mean squared errors than existing methods. In contrast to previously developed methods, we eliminated the strict conditions of diagonal or block-wise diagonal form of the population covariance matrix and extend the work to a wider range without the assumption of normality. Thus, the proposed method is more suitable for real problems.