Image classifiers often report high accuracy on clean benchmark data, but in the real world, their inputs are not always clean. This paper tests how two convolutional models respond when part of the fashion-MNIST test set is deliberately corrupted. A custom four-layer convolutional neural network and an ImageNet-pretrained ResNet-18 were trained only on clean Fashion-MNIST images. A fixed corruption, Gaussian blur followed by contrast enhancement, was then applied to 0%, 10%, 30%, 50%, 70%, and 100% of the official test images. Both models used the same corrupted image set. Accuracy, mean predictive entropy, and t-distributed stochastic neighbor embedding plots were used to compare the runs. The custom model reached 91.31% accuracy on clean data but fell to 10.82% when every test image was corrupted. ResNet-18 started lower, at 88.10%, but reached 20.44% at 100% corruption. Entropy also rose with the corrupted percentage, from 0.1729 to 0.5379 for the custom model and from 0.1517 to 0.3073 for ResNet-18. These suggest that both clean-trained models were fragile under this specific corruption, while ResNet-18 held up better in this run. A possible explanation is that its deeper residual structure and ImageNet-pretrained layers preserved more coarse shape information after blur weakened fine local details.

The current study offers a sentence-level natural language processing approach aimed at detecting competitive discourse and measuring equity orientation in education policy statements. The sample contains 800 policy statements published from 2015 to 2024 by national and provincial educational authorities. After being preprocessed and checked manually, 60,284 sentences from official education policies have been selected for further analysis. Competitive discourse is classified into five categories: selection competition, performance ranking, cultivation of elites, assessment pressures, and school competition. Meanwhile, equity orientation is measured on four levels: access, resource, process, and outcome equity. The BERT-Base model for Chinese was fine-tuned using a training set of 5,000 annotated sentences, and policy balance scores were obtained with the help of discourse and equity measurement indicators. Experiments based on simulated policy statements show that the accuracy of the classifier can be assessed as 0.884 ± 0.017. In addition, the findings confirm that policy balance scores improved starting from 2021 and concerning compulsory education and provincial-level policies.

The UWB is a short-range wireless communication technology using very wide bandwidth RF pulses. UWB communication could provide a high-level accuracy for autonomous tracking and localization. However, the UWB ranging based on the Time Difference of Arrival method (TDoA) would often be compromised by non-ideal environmental conditions, antenna propagation delays and multiple path effects, etc. This paper introduces a method to optimize a UWB localization system based on DW1000, combined with the hardware features of modules and algorithmic approaches. Systematically, it addresses the hardware limitations by simulating and comparing antenna designs and implementing the antenna-delay calibration, maintaining the pulse signal integrity. Also, this study establishes a rigorous TDoA mathematical model to demonstrate three mainstream localization algorithms: Fang, Chan and Taylor through comprehensive Monte Carlo simulations. The evaluations considered different anchor densities, from 3 to 7, with the addition of white Gaussian noise (AWGN) levels and distinct operational boundaries based on macroscopic channel models. The testing result demonstrates that Chan's algorithm achieved the lowest simulated error under AWGN among the tested algorithms, but Taylor's expansion represents a superior robustness in non-Gaussian environments with multipath fading. In conclusion, the hardware experiments verified UWB TWR ranging and two-anchor geometric positioning. TDoA localization algorithms were then studied through MATLAB simulation as a potential extension for multi-anchor localization. Thereby, a hybrid method is required, combined with hardware strategies and software algorithms. The software utilizes the TDoA localization with Taylor-series iterative refinement.

Traditional neural network training based on backpropagation suffers from multiple bottlenecks, including slow convergence rate, susceptibility to local optima, vanishing/exploding gradients, and insufficient generalization performance. To address these issues, this paper deeply integrates convex optimization theory with the backpropagation algorithm and constructs a novel stable and efficient training mechanism for neural networks. Systematical optimization of the conventional training pipeline is realized via convex reconstruction of the loss function, design of an adaptive gradient correction rule under convex optimization constraints, and rigorous theoretical proof of convergence for the integrated algorithm. Experimental results demonstrate that compared with mainstream algorithms such as standard BP, SGD and Adam, the proposed mechanism reduces the number of convergence iterations by over 35%, cuts training time by 28%, improves classification accuracy by 4%-7%, and effectively suppresses gradient anomalies. It achieves favorable adaptability to both shallow fully connected networks and deep convolutional networks. This research complements the theoretical convex optimization framework for non-convex training, and provides methodological support and theoretical references for efficient training and industrial deployment of deep learning models.

The integration of renewable energy, such as solar and wind power, is essential for the development of zero-carbon smart cities. Reliable short-term and medium-term predictions help reduce dependence on fossil-fuel reserves and facilitate more reasonable decisions in urban systems. This study develops a hybrid Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM) model for forecasting multi-source renewable energy. The convolutional layers can extract local temporal patterns from meteorological variables, while the LSTM layers model sequential dependencies over longer horizons. To improve transparency, Local Interpretable Model-agnostic Explanations (LIME) is applied to examine how input variables influence predictions. Experiments are conducted using data from the National Renewable Energy Laboratory, including the National Solar Radiation Database and the Wind Integration National Dataset. Compared with standalone CNNs, LSTMs, and conventional learning models, the hybrid model achieves lower prediction error, with Mean Absolute Percentage Error typically below 3% for forecasting next-day values. Results indicate that higher forecasting accuracy can contribute to reduced renewable curtailment and to more stable, reasonable decisions in smart cities. The proposed framework offers a practical, interpretable approach to renewable energy forecasting in urban systems.

This study applies a Random Forest classifier to the FBI's national Uniform Crime Reporting (UCR) dataset containing 218,069 single-bias hate crimes from 1991to 2020. Within the dataset, crimes were predicted based on bias-driven motives: race, sexual orientation, gender, religion, or disability. Thirteen incident features were analyzed to form the prediction. The primary model chosen for this study oversampled the two smallest categories and tuned hyperparameters. Ultimately, an accuracy of 68.7% with F1 = 0.80 and recall = 0.93 was achieved for the main prediction target: racial bias. Five trials were conducted with different configurations of the train/test model, compared across two metrics. It can be seen that there is a significant trade-off between total balance within all 5 categories versus overall accuracy. It was determined the incident year was the dominant predictive feature, likely due to FBI changes in definition and classification across the 1991-2020 period. Results demonstrate the predictability of indicators within official records and are limited to UCR metrics of data collection.

As corporate crisis communication increasingly relies on public videos, social media statements, and documentary-style visual responses, public judgments of responsibility, sincerity, and remediation capacity are increasingly shaped by narrative credibility. Traditional evaluation of crisis communication mainly focuses on response strategies, responsibility attribution, and reputation repair effects, while giving limited attention to the integrated relationship among video text, visual evidence, and public feedback. To address narrative credibility evaluation in corporate crisis communication, a computational method is developed by combining film narrative analysis with large language models. Public corporate crisis videos, statement texts, and comment data from 2019 to 2025 are used as the corpus. Five evaluation dimensions are established, including factual consistency, responsibility clarity, verifiable action, emotional appropriateness, and narrative coherence. The experiment is completed through human annotation, feature extraction, LLM-based dimensional scoring, and error testing. The results show that LLM-based dimensional scoring outperforms the sentiment lexicon method, TF-IDF plus SVM, and the BERT classifier in Accuracy, Macro-F1, MAE, and Spearman correlation. Its Accuracy reaches 0.781, Macro-F1 reaches 0.742, MAE is 0.46, and Spearman correlation is 0.724. Responsibility clarity and verifiable action show relatively stable performance, while emotional appropriateness still presents higher error. These findings indicate that large language models can provide a computable and explainable technical path for evaluating narrative credibility in corporate crisis videos. They also offer methodological reference for computational film studies, crisis communication assessment, and corporate reputation governance.

Personalized recommender systems may repeatedly expose users to similar content, potentially contributing to information cocoons. Existing research has mainly examined how recommendation algorithms influence content narrowing, while the role of persistent user characteristics remains less explored. This study investigates whether users with different initial levels of interest diversity experience different rates of information cocoon formation in a long-term recommendation environment. Using MovieLens 1M, interest diversity is measured with Shannon entropy over genre distributions. Users are classified into low-, medium-, and high-diversity groups, and 900 users are selected through stratified random sampling. A 20-round UserCF recommendation simulation is then conducted. Mean interest entropy decreases in all three groups, but at different rates. The high-diversity group exhibits the fastest decline, at approximately four times the rate of the low-diversity group. Contrary to the initial expectation, the low-diversity group declines most slowly. This may reflect a floor effect, because its initially concentrated interest distribution leaves less room for further reduction. These findings show that initial interest diversity is associated with the speed of information cocoon formation, shifting attention from whether cocoons form to how quickly they develop.

In response to the contradiction of multi-performance conflict and high cost of high-fidelity simulation evaluation in the integrated design of industrial products, this paper proposes a design-simulation-optimization integration framework that couples a deep surrogate model and multi-objective evolutionary optimization: a deep neural network learns the rapid mapping of parameterized designs to multi-physics performance to replace expensive simulation, and then NSGA-II searches for the Pareto front in the engineering-feasible space. On the public building energy-efficiency benchmark data, the coefficients of determination of the deep surrogate for heating and cooling loads reach 0.994 and 0.982 respectively, and the GPU evaluation speed is about 4.44 million times per second; the surrogate-driven optimization quantitatively portrays the trade-off front of energy consumption and daylighting, and identifies key design drivers such as relative compactness and window-to-wall ratio. The results show that the framework can support the integrated design of data-driven industrial products at a very low computing cost, providing an efficient paradigm for reproducible design decision-making.

Accurate electricity load forecasting is important for power system operation, energy planning and demand management. With the increasing use of data-driven methods, deep learning models have become useful tools for analysing complex electricity demand patterns. This study develops a PyTorch-based LSTM model for short-term electricity load forecasting, using South Australian electricity demand data from 2024. The original dataset consists of chronological five-minute demand records, and total electricity demand is selected as the target variable. Missing values were checked, the demand series was normalised using training-set statistics, and supervised learning samples were generated by a sliding window method. In the proposed model, the previous 24 five-minute demand observations are used to predict the next demand value. The dataset is divided into training, validation and testing subsets in chronological order. The model is trained using mean squared error loss and the Adam optimiser with a step learning-rate schedule. The results show that the LSTM model can follow the main movement of electricity demand, but the prediction curve is smoother than the actual curve and some short-term fluctuations remain difficult to capture. This limitation is mainly caused by the use of a simple univariate input, because weather, calendar, electricity price and renewable generation variables are not included. Future work should introduce multivariate features, compare LSTM against benchmark models such as GRU and ARIMA, and test longer input windows to further improve forecasting accuracy.