Machine learning has become a central technology for transforming data into intelligent decisions in modern organisations. Recent advances in deep learning and reinforcement learning have greatly expanded the capacity of machines to recognise patterns, forecast outcomes and optimise sequential choices, which has stimulated a wide range of decision-support applications. By adopting a structured review method, this paper reviews the applications and challenges of machine learning in intelligent decision-making. The results show that learning-based models improve the accuracy, speed and scalability of decisions in clinical diagnosis, asset management, supply-chain planning and autonomous driving. The analysis also identified four recurring challenges that limit practical adoption, namely data quality, model interpretability, algorithmic fairness and safety. The paper concluded that reliable deployment depends on transparent model design, fairness auditing and human-in-the-loop governance, and it proposed directions for future research in hybrid human-machine decision systems.

Welding robots are now extensively used in shipbuilding, automobile, and aviation industries. Defects are prone to occur after welding, and traditional non-destructive testing like ultrasonic and radiographic testing, depend on manual judgement, resulting in detection lag and difficulty in achieving online closed-loop control on production lines. This paper focuses on robot-mounted vision-based post-weld inspection, dividing the techniques into non-destructive and destructive testing. It outlines a three-level processing paradigm for vision-based inspection: low-level preprocessing, mid-level defect recognition and segmentation, and high-level quality diagnosis. The paper systematically summarizes three deep-learning technical routes: defect recognition and classification, 3D geometric measurement of weld seams, and multimodal temporal intelligent diagnosis. It also compares the application characteristics and applicable scenarios of mainstream models such as Convolutional Neural Networks (CNNs), You Only Look Once (YOLO), U-Net, and Detection Transformer (DETR). Various interference factors in industrial environments are analyzed. Finally, future technical directions are discussed in four key aspects, providing a comprehensive theoretical reference for the development of intelligent welding inspection algorithms and the engineering integration of robotic systems.

Printed Circuit Boards (PCBs) are necessary parts of current electronic products, and surface defects can cause problems in the operation or lifespan of these devices. Traditional inspection methods are unable to find small defects, are easily affected by a cluttered background, and have low inspection efficiency. Therefore, deep learning and object detection technologies have been applied to automate the inspection of PCB defects. This paper investigates how artificial intelligence (AI) can be used to address the problems of detecting small defects on PCBs and builds a conceptual system that integrates the attributes of PCB defects with AI inspection technology. Based on the review of previous studies, this paper presents an all-encompassing system architecture for defect attributes, AI feature learning, defect detection and localisation, and performance evaluation; four representative case studies are used to verify the system's practicality. Based on the above results, using small-object detection, attention mechanisms, improved feature extraction methods and lightweight network architectures can improve both the detection accuracy and efficiency. In addition, all the above factors should be taken into account in the process of defect detection. A limitation of this study is that it is based on existing literature and case analyses without new experimental verification; in the future, further tests of the framework can be conducted using publicly available PCB defect datasets.

Electroencephalography (EEG)-based emotion recognition has attracted increasing attention in recent years because of its potential applications in brain-computer interfaces, intelligent healthcare, and affective computing. With the rapid development of deep learning, attention mechanisms have gradually become an important component of EEG emotion recognition models, since they are able to improve feature representation by focusing on more informative EEG signals. This paper reviews recent attention-based deep learning methods for EEG emotion recognition, with particular attention to channel attention, temporal attention, and self-attention implemented in Transformer-based architectures. Their main ideas, representative models, as well as their advantages and limitations are discussed and compared. This paper finds that although existing studies have achieved encouraging performance, several challenges, including cross-subject variability, limited EEG datasets, and insufficient model interpretability, still remain. Future research is expected to develop more efficient and interpretable attention mechanisms while exploring larger datasets and multimodal learning frameworks, which may further improve the robustness and practical application of EEG emotion recognition systems.

Skin lesion image classification is important for early skin-cancer screening, but dermoscopic datasets are severely class-imbalanced, which can impair the recognition of minority classes such as melanoma. Using MobileNetV2 as the backbone, this paper studies, through ablation experiments, how channel attention (squeeze-and-excitation, SE) and class-imbalance handling strategies affect seven-class skin lesion classification. On the HAM10000 dataset split by lesion identifier, class-balanced weighted cross-entropy raises the macro-F1 from 0.702 to 0.776, and class-balanced sampling further raises it to 0.797, minority-class recall improves (melanoma 0.514 to 0.577, vascular lesions 0.467 to 0.867). SE alone improves class balance, but its gain does not add up with the imbalance strategies. The results provide empirical guidance for comparing structural enhancements and imbalance-handling strategies in lightweight skin lesion classification.

With the rapid popularization of artificial intelligence (AI) technology in the field of education, personalized learning will become the core direction of digital transformation in education. Learning emotions directly affect learning efficiency, learning motivation, and learning outcomes. In traditional learning systems, only knowledge acquisition is focused on, lacking perception and intervention of learners' emotional states. AI technologies such as computer vision, facial expression recognition, speech emotion recognition, and behavior recognition are becoming increasingly mature, providing the possibility for real-time capture and learning of emotions. Building a personalized learning system that combines emotional perception, intelligent recommendation, and dynamic intervention is an important research gap in the current field of educational technology.

Published comparisons of PM2.5 prediction models reach different conclusions about the value of deep learning. This review examines eleven empirical studies and one methodological review published between 2018 and 2024 to assess those differences. The papers are compared by prediction target, input data, and the alternatives included in each experiment. Support vector regression and linear regression performed well in a Nottingham study using features extracted from recent concentration trends. Other studies reported lower errors with combined convolutional and recurrent networks or Transformers. In these systems, however, changes to the predictor often accompanied changes to feature selection or the representation of spatial information. The reported improvement therefore cannot always be attributed to network architecture alone. Forecast duration presents a further difficulty: estimating the next day's mean concentration and predicting an hourly peak require different evaluations. Conventional regression provides a useful baseline for these comparisons. A proposed replacement needs testing at the intended lead time, including pollution episodes, with inputs restricted to information available when the prediction is issued.

Large language model (LLM) tutors remove exactly the effort that learning requires: students adopt the instant answer, offload the reasoning to the model, and retain less. Friction-based tutors reverse the offloading, yet learners disengage from unbounded friction. This paper tests whether game-structural mechanics—checkpoints, low-stakes failure with retry, and a rescue that guarantees completion—hold the learner inside a tutor's deliberate friction. In a multi-agent simulation of 180 students (three independent draws of 60) across six levels of seeded misconception burden, each student works the same ten-part physics problem under three tutor conditions: instant answer, strict Socratic, and a ten-checkpoint quest in which every part is unlocked only by the student's own attempt. The quest condition removes the measured offloading: students produce 0.999 of the solution steps against 0.302 under the instant answer, self-reported offloading is 1.48 versus 5.58 on a 6-point scale, and every student names the instant-answer condition as the one in which they offloaded most. Structuring the friction also removes its cost: quest sessions complete in 100 percent of cases, against a 14.4 percent abandonment rate in the strict condition, where sessions that run long end when the student's reported motivation reaches the floor. Both friction conditions raise the unaided posttest far above the shared pretest (0.928 and 0.885 versus 0.455); the instant-answer condition reaches only 0.563. Friction that is structured as bounded progression holds the learner inside productive struggle without the attrition that plain withholding carries.

Better energy efficiency is becoming increasingly important as electric vehicles are used more widely in everyday transportation. Artificial intelligence (AI) has become an important technology for improving the energy efficiency of electric vehicles (EVs). This study examines the application of AI in battery management systems (BMS), energy management systems (EMS), thermal management, and intelligent driving strategies. Collectively, these technologies improve battery performance, optimize energy utilization, and extend driving range. Analysis shows that AI can provide a multifaceted approach to improving EV energy efficiency by enabling more accurate battery prediction, dynamic energy allocation, and adaptive thermal control. These algorithmic methods can reduce unnecessary energy consumption while maintaining vehicle performance and battery safety. Prognostic AI models can also help predict battery degradation and support more effective maintenance decisions. Better energy efficiency is becoming increasingly important as electric vehicles are used more widely in everyday transportation. Existing challenges include computational cost, data security, and model training. Prospectively, continued advances in AI are expected to support the development of more efficient and sustainable electric vehicles. In addition, improved interoperability between AI models and vehicle systems may further enhance real-time decision-making and energy optimization.

Current research on multimodal emotion recognition seeks to analyse the relationships among facial, vocal, and bodily expressions in order to recognise different emotions. Although additional data streams are often introduced, they cannot be treated as independent evidence. This conceptual study aims to understand the reliability of multimodal inferences over unimodal classifications. This study covers the theory and practice of affective computing, the study of different modalities, fusion taxonomies and benchmarking. Depending on the statistical dependencies that they learn, fusion can occur in early, intermediate, late or adaptive form. The synthesis suggests that the advantage of multimodal systems is determined by four factors: informational complementarity, partially independent errors, temporally valid alignment, and dynamic reliability estimation. For these factors, fusion of different types of data is a good alternative. Otherwise, it is likely to increase the recording of social displays, bias of annotations, or dependency on a dominant modality. For this reason, a reliability-sensitive framework has been proposed, in which recognition is driven by uncertainty management rather than feature accumulation, while preserving the modality-specific nature of different types of evidence. The resulting account explains both the recurrent gains reported in benchmark studies and their contraction under naturalistic expression, distribution shift, cultural mismatch, strategic display, and systematic missingness. Multimodal systems become defensible not by sensing more behaviour alone, but by estimating when each observation should change, confirm, or weaken an affective inference.