Meng Yang

In the field of coal mine safety, traditional large language models (LLMs) face issues such as poor knowledge timeliness, lack of traceable evidence, and susceptibility to hallucinations when handling complex knowledge-intensive tasks. To align with the human-centric principles of Industry 5.0 and meet the strict compliance requirements of Industrial Regulatory Technologies (RegTechs), this study aims to enhance the performance of a Retrieval-Augmented Generation (RAG)-based coal mine safety compliance system. A comprehensive coal mine knowledge dataset (CMK) is constructed to serve as a dynamic RegTech knowledge base. Based on the RAG framework, a two-stage retrieval structure is designed and optimized, transforming the generative AI into a safe, constrained reasoning core. Also, a hard-negative ranking dataset (CMK-R) is developed to support the evaluation of the reranking model and mitigate the risk of semantic hallucinations. The proposed approach adopts Dmeta-embedding-base and BGE-rerank-base as backbone models and applies fine-tuning, teacher–student structured distillation, and Matryoshka Representation Learning (MRL) to improve efficiency and reduce energy consumption. Experimental results demonstrate that the proposed CMS-base model achieves FAHR, MRR@10, and mAP scores of 93.48, 95.59, and 95.63, respectively, indicating that the system can retrieve the correct safety regulation or operational clause at the top ranks with high reliability. Furthermore, the lightweight variant reduces inference time by over 60%, and accuracy degradation remains below 1% even when the vector dimension is compressed to 64, significantly facilitating Edge Intelligence deployments. This robust, human-centric framework provides a scalable industrial AI solution for real-time safety compliance and decision support.

Although fine-tuning pretrained language models has brought remarkable progress to zero-shot named entity recognition (NER), current generative approaches still suffer from inherent limitations. Their autoregressive decoding mechanism requires token-by-token generation, resulting in low inference efficiency, while the massive parameter scale leads to high computational and deployment costs. In contrast, span-based methods avoid autoregressive decoding but often face large candidate spaces and severe noise redundancy, which hinder efficient entity localization in long-text scenarios. To overcome these challenges, we propose an efficient Embedding-based NER framework that achieves an optimal balance between performance and efficiency. Specifically, the framework first introduces a lightweight dynamic feature matching module for coarse-grained entity localization, enabling rapid filtering of potential entity regions. Then, a hierarchical progressive entity filtering mechanism is applied for fine-grained recognition and noise suppression. Experimental results demonstrate that the proposed model, which is trained on a single RTX 5090 GPU for only 24 h, attains approximately 90% of the performance of the SOTA GNER-T5 11B model while using only one-seventh of its parameters. Moreover, by eliminating the redundancy of autoregressive decoding, the proposed framework achieves a 17× faster inference speed compared to GNER-T5 11B and significantly surpasses traditional span-based approaches in efficiency.

Ensuring safety in coal mining operations requires an accurate understanding and execution of complex regulatory and procedural knowledge. However, existing general-purpose language models often lack domain-specific adaptation, limiting their effectiveness in high-risk industrial scenarios. This study presents MS-GLM, a domain-adaptive language model tailored for coal mine safety applications. A multisource and multitask dataset, Mine Safety Intelligence (MSI), was constructed, comprising of 10GB plain text and 10,000 high-quality instances that cover safety protocols, emergency response, and hazard prevention. The data set integrates human-annotated content and filtered ChatGPT output, ensuring both diversity and factual accuracy.MS-GLM is trained using a two-stage framework that combines domain-adaptive pretraining and LoRA-based supervised fine-tuning. Experimental evaluations on a custom coal mine QA benchmark (CMS-QED) and general-purpose NLP tasks demonstrate that MS-GLM significantly improves factual accuracy and safety-related reasoning while maintaining competitive inference efficiency. Ablation studies confirm that pretraining and fine-tuning contribute complementary gains in both precision and generalization.These results suggest that domain-specific language modeling strategies can effectively enhance the decision-support capabilities of expert systems in safety-critical industries.

Adequate CO₂ is essential for vegetation, but industrial chimneys and land, space and oceanic vehicles exert tons of excessive CO₂ and are mostly responsible for the greenhouse effect, global warming and climate change. Due to COVID-19, CO₂ emission was in 2020 at its lowest level compared to prior decades. However, it is unknown how long it will take to reduce CO₂ emission to a tolerable point. Furthermore, it is also unknown to what extent it can increase or change in the future. Accurate forecasting of CO₂ emissions has real significance for choosing the optimum ways of reducing CO₂ emissions. Although some existing models have noticeable CO₂ emission forecasting accuracy, the models implemented in this work have more efficacy in prediction due to incorporating COVID-19’s effect on CO₂ emission. This paper implements four prediction models using SARIMA (SARIMAX) based on ARIMA. The four models are based on the time period of the surge of the COVID-19 pandemic. The main objective of this work is to compare these four models to suggest an effective model to predict the total CO₂ emissions for the future. The study forecasts global total CO₂ emission from 2022 to 2027 for near future prediction, 2022 to 2054 for future prediction and 2022 to 2072 for far future prediction. Among the various error measures, mean absolute percentage error (MAPE) is chosen for accuracy comparison. The calculation yields different accuracy for the four SARIMAX models. The MAPEs for the four methods are: pre-COV (MAPE: 0.32), start-COV (MAPE: 0.28), trans-COV (MAPE: 0.19), post-COV (MAPE: 0.09). The MAPE value is relatively low for post-COV (MAPE: 0.09). Hence, it can be inferred that post-COV are suitable models to forecast the global total CO₂ emission for the future. The post-COV predictions for the global total CO₂ emission for the years 2022 to 2027 are: 36,218.59, 36,733.69, 37,238.29, 37,260.88, 37,674.01 and 37,921.47 million tons (MT). This study successfully predicts CO₂ emission either for the COVID-19 period or the post-COVID-19 normal periods. The Machine Learning (ML) method used in this study has shown good agreement with the IPCC model in predicting the past emissions, the current emissions due to COVID-19 and the emissions of the upcoming future. These prediction results can be an asset for the decision support system to develop a suitable policy for global CO₂ emission reduction. For future research, a number of other external influence variables responsible for CO₂ emission can be added for finer forecasts. This research is an original work in predicting COVID-19-affected CO₂ emission using AI through the ML methodology.

Silent speech decoding is a novel application of the Brain–Computer Interface (BCI) based on articulatory neuromuscular activities, reducing difficulties in data acquirement and processing. In this paper, spatial features and decoders that can be used to recognize the neuromuscular signals are investigated. Surface electromyography (sEMG) data are recorded from human subjects in mimed speech situations. Specifically, we propose to utilize transfer learning and deep learning methods by transforming the sEMG data into spectrograms that contain abundant information in time and frequency domains and are regarded as channel-interactive. For transfer learning, a pre-trained model of Xception on the large image dataset is used for feature generation. Three deep learning methods, Multi-Layer Perception, Convolutional Neural Network and bidirectional Long Short-Term Memory, are then trained using the extracted features and evaluated for recognizing the articulatory muscles’ movements in our word set. The proposed decoders successfully recognized the silent speech and bidirectional Long Short-Term Memory achieved the best accuracy of 90%, outperforming the other two algorithms. Experimental results demonstrate the validity of spectrogram features and deep learning algorithms.