To address the issue that existing platforms focus on equipment access, data aggregation, and business system integration, a five-layer architecture of "object perception - network data - model decision-making - control execution - business management" was constructed. Compared with the typical five-layer technology stack of "field layer - edge layer - IaaS - PaaS - SaaS" and the organizational architecture of ten major business systems, this framework further clarified three types of operating time scales, seven closed-loop links, and two types of feedback channels for production and quality. Analysis shows that intelligent mines are evolving from single-machine automation to cross-process and multi-equipment collaboration, but overall they are still dominated by local closed loops, and the feedback of product quality information to upstream production is insufficient; inconsistent data semantics, insufficient model generalization ability, non-standardized task interfaces, and imperfect cross-equipment safety interlocking remain the main problems restricting large-scale application. To this end, based on a unified object model and digital thread, it was proposed to advance mechanism-data fusion, cloud-edge-terminal collaboration, and a task-driven safe autonomous closed loop, in order to promote the whole industry chain collaboration of planning, production, control, and product quality.

The engineering basis for introducing large models in mines is established through the long-term accumulation of digitalization, automation, and specialized models. However, the integrated understanding of complex risks by safety systems is still hindered by fragmented data, isolated knowledge, and the disconnection between early warning and response. Based on the technological evolution of intelligent mines and engineering practices, the prior foundations, including platform interfaces, specialized tools, multi-source safety evidence, and domain knowledge, were examined, and the differentiated requirements of high-risk scenarios such as goaf, tailings pond, and hoisting system were analyzed. The study suggests that mine safety large models should not replace control systems and on-site responsibility chains, but rather serve as a "cognition-and-proposal layer" to participate in multimodal understanding, risk assessment, mechanism interpretation, and solution generation. Its outputs should be checked by data evidence, professional rules, mechanistic models, protection interlocks, and human review, and routes for rejection and safe fallback shall be retained. Data governance and knowledge base, cognition and proposal, constraint checking, on-site response, and outcome feedback are incorporated into the same accountability chain by the resulting layered collaborative architecture. The engineering value of mine safety large models should possess interpretability, traceability, verifiability, and fallback capability.

To address the problem that the planetary gears in the gearbox of mine hoists operate under long-term low-speed, heavy-load, and alternating-load conditions, and that single acoustic or vibration signals are easily affected by propagation path, measuring point position and background noise, leading to insufficient feature extraction of fault signals, a fault diagnosis method based on point-symmetric acoustic-vibration fusion and ARFM-Net is proposed. Firstly, an acoustic-vibration synchronous acquisition test system for the planetary gearbox was established, and the differences and complementarities of acoustic and vibration signals in expressing fault information were analyzed. Then, Variational Mode Decomposition (VMD) and Feature Mode Decomposition (FMD) were adopted to process the two types of signals according to their respective characteristics. The effective modal components containing major fault information were selected and mapped into a polar coordinate feature map by using the point-symmetric graph method to realize the image-level fusion of multi-scale acoustic-vibration information. Finally, an ARFM-Net diagnostic network composed of Adaptive RepLK and Frequency-aware Moga was constructed. An adaptive large-kernel convolution was used to enhance the extraction capability of local fault details and long-range texture correlations, and a frequency-aware multi-level gated feature aggregation was used to highlight fault-related information and suppress background redundant textures. The experimental results of five types of planetary gear fault states show that the average recognition accuracy of the proposed method reaches 98.72%, and it shows favorable diagnostic performance in both comparative and ablation tests. This method can effectively improve the fault-recognition accuracy of planetary gears in the gearbox of hoists under complex working conditions, and provide technical support for intelligent monitoring of operation status of hoist transmission systems in mines.

To improve the adaptability of the energy management strategy for mining hybrid loaders to variable cyclic operating conditions, an energy management strategy integrating Deep Q-Network (DQN) and Transfer Learning (TL), named DQN-TL, was proposed. The core idea of this method is that the inherent cyclic characteristic of the loader's V-type operation is utilized, and such repeatable patterns are regarded as extractable and reusable high-value knowledge. First, DQN is utilized by the strategy to learn and distill the general control laws for typical operating cycles in a data-rich source domain. Then, through parameter transfer, the relevant law features are introduced into the target domain model, which is then fine-tuned with a small amount of data to adapt to specific operating conditions. Finally, by establishing the intrinsic relationship between the structured cyclic characteristics of the operating conditions and transfer learning, the adaptability to complex underground conditions is improved. The simulation results show that, compared with a rule-based control strategy, the fuel consumption rate of the proposed DQN-TL strategy is 189.21 g/(kW·h), the fuel economy is improved by 13.8%, its power allocation strategy is closer to the global optimal solution, and good potential for engineering application is demonstrated.

Aiming at the problem of difficult multi-protocol wireless access caused by the heterogeneous protocols of underground intelligent mine equipment and the harsh communication environments, a VxLAN over 5G multi-protocol transparent access scheme suitable for mine scenarios was proposed based on the special channel constraints and 5G private network transmission technology in underground mines. A unified two-layer virtual tunnel was constructed over the 5G network to directly encapsulate and transmit raw Ethernet frames of mainstream industrial protocols such as EtherCAT and PROFINET, without secondary modification to the protocol stacks of existing mine equipment, which greatly reduces the cost of upgrading and reconstructing underground communication networks. The actual verification was carried out by building a simulated mine underground 5G private network test platform. The results show that the end-to-end average latency of EtherCAT protocol data is 22.508 ms, and the maximum latency is 44.794 ms. The average latency of PROFINET protocol data is 21.007 ms and the maximum latency is 44.477 ms. The latency of both protocol data can stably meet the real-time requirements for scenarios such as condition monitoring of fixed underground mine equipment (latency≤100 ms) and conventional remote control of auxiliary operation robots (latency≤50 ms). However, for strong real-time control scenarios such as mining faces that require millisecond-level synchronization accuracy (latency≤1 ms), the current scheme cannot support it, and needs to be further optimized in combination with 5G URLLC enhancement features. The research results can provide a feasible technical path with low transformation cost and high versatility for the wireless access of multi-protocol heterogeneous equipment in non-coal mines, and provide engineering reference for the refined service bearer of 5G private network in intelligent mines.

To address prominent problems prevalent at the regulatory end of the dual-prevention mechanism, including data island, risk identification relying on fixed thresholds, and difficulties in penetrating enterprise underlying data, and to explore an effective path for digital-intelligent transformation of the regulatory end, a four-level regulatory framework of "data penetration-early warning grading-closed-loop response-law-enforcement implementation" was proposed. By realizing real-time interconnection between enterprise underlying sensing data and the regulatory end, this framework conducts cross-verification on the authenticity of hazard investigation, rectification management and monitoring data using methods such as spatio-temporal behavior verification, spatial static verification, and data consistency verification. On this basis, an early warning comprehensive index model was established to realize dynamic risk identification and hierarchical response. At the same time, a differentiated and accurate law enforcement path based on a digital evidence chain and "dynamic safety risk profile" was explored to improve the timeliness and accuracy of supervision. The empirical results show that this model improves the efficiency of mine supervision by more than 60% and effectively reduces the frequency of accidents. It provides a scientific scheme for promoting the transformation of the dual prevention mechanism from formal operation to substantive operation, and provides a feasible path for building a more efficient and intelligent safety governance system for non-coal mines.

A comparison and selection study on comminution process schemes was conducted for a super-large low-grade porphyry copper-molybdenum mine in Inner Mongolia (with reserves of 950 million t, a design scale of 70 000 t/d, a Cu grade of 0.165%, and a Mo grade of 0.032%). The ore is moderately hard to hard and contains easily sliming minerals such as chlorite. Based on a systematic review of the application status of three types of comminution processes, including conventional comminution, semi-autogenous grinding (SAG), and high-pressure grinding rolls (HPGR), four schemes were proposed: the semi-autogenous grinding+ball mill+pebble crushing (SABC) process, 50 mm feed high-pressure grinding roll dry process, 30 mm feed high-pressure grinding roll dry process, and high-pressure grinding roll wet process. A technical and economic comparison was conducted from the aspects of equipment configuration, construction investment, energy consumption, operating costs, and total lifecycle costs. The results show that the SABC process has the lowest construction investment, which is 17.89% lower than that of the high-pressure grinding roll wet process, but has the highest production cost per ton of ore (23.55 yuan/t); the high-pressure grinding roll wet process has the lowest production cost per ton of ore (20.07 yuan/t), which is 17.34% lower than that of the SABC process. The high-pressure grinding roll wet process has the lowest present value of the total lifecycle cost, indicating its optimal long-term economic performance. Comprehensively considering the ore characteristics, equipment maturity, and total lifecycle economic performance, the high-pressure grinding roll comminution wet process of "cone medium crushing, closed-circuit screening, high-pressure grinding roll ultra-fine crushing, wet closed-circuit screening, and one-stage closed-circuit grinding" is recommended, to provide a reference for the selection of comminution processes for similar super-large low-grade copper-molybdenum mines.

To address the mining and separation difficulties of a low-grade zircon-hafnium placer in a coastal shallow sea of Hainan, a systematic process mineralogy study was conducted using analytical techniques such as X-ray diffraction, mineral liberation analyzer (MLA), and scanning electron microscopy-energy dispersive spectroscopy (SEM-EDS). The chemical composition of the ore, mineral composition, dissemination size of valuable minerals, and mineral occurrence characteristics were clarified. Furthermore, a feasibility study on pre-screening and centrifugal tailing discarding was carried out using a self-made LX-60 vertical cyclone centrifuge. The results indicate that the ZrO₂ grade of the raw ore is only 0.105 0%, the valuable mineral zircon accounts for only 0.14%, and the content of light gangues such as quartz and feldspar is up to 86.60%. 75% of zircons are monomer with particle size of 0.04-0.20 mm, showing a significant specific density difference from the gangues, which is suitable for gravity separation waste discarding. 25% of zircons contain quartz, feldspar, and ilmenite inclusions with particle sizes less than 0.05 mm. This part of fine inclusions is easy to cause metal loss. The dissemination size of ilmenite is coarser than that of zircon, and it is synchronously enriched during gravity separation. After pre-screening out the +0.15 mm coarse material, the undersize material undergoes one-stage roughing under the conditions of a centrifugal force of 80g, a backwash water flow rate of 4 L/min, and a feeding amount of 80 g, yielding a pre-selected rough concentrate with a ZrO₂ grade of 0.483 6% and an operation recovery of 95.50%. The total zirconium recovery of the combined screening-centrifugation process reaches 93.37%, and the total loss rate to the raw ore is only 6.63%. This process achieves the efficient discarding of massive light gangues and the pre-concentration of zircon minerals, demonstrates potential applicability for offshore in-situ pre-selection tailing discarding, and can provide technical support for the pre-selection tailing discarding of similar extremely low-grade zircon placers.

The recovery rate of micro-fine cassiterite in a complex polymetallic tin ore in Yunnan has long been low, seriously restricting the improvement of comprehensive beneficiation indicators. Through systematic analysis of tin loss characteristics, the causes of tin loss were identified. Analysis of tin loss characteristics shows that the disseminated grain size of cassiterite is mainly concentrated in the range of 5-40 μm, with the proportion of the -20 μm fraction exceeding 55%, and the dissociation degree of cassiterite is only 31.63% at the current grinding fineness. The selectivity of the three-stage series hydrocyclone desliming is poor, the metal distribution rate of +10 μm cassiterite in the desliming overflow reaches 63.62%, and the cumulative tin metal loss from the multi-stage hydrocyclone desliming operation is about 7%. Conventional mechanical agitation flotation machines have insufficient recovery efficiency for -10 μm ultra-fine cassiterite. Technical renovation tests were carried out to break through the technical bottleneck of efficient recovery of micro-fine cassiterite. The results show that by replacing the vertical stirred mill with an IsaMill for fine grinding, the content of the -38 μm fraction increases from 50% to 75%, and the dissociation degree of cassiterite increases from 46.37% to 84.21%. By replacing the three-stage hydrocyclone with a decanter centrifuge for one-stage desliming, the tin recovery rate in the underflow increases by 8.39 percentage points. By applying a microbubble flotation column to replace the conventional mechanical agitation flotation machine, along with simplifying the flowsheet structure, the tin recovery rate is increased by 7.98 percentage points compared with on-site production. Simultaneously, the flotation reagent regime is optimized by adopting the combined collectors YK-Sn, Sn-705, and Sn-F, which increases the recovery rate of the tin rougher concentrate by 6.03 percentage points compared with the original collectors. Through the systematic integration of fine grinding, high-efficiency desliming, enhanced flotation and reagent optimization, the problem of micro-fine cassiterite recovery can be effectively solved, and the pilot combined flowsheet has achieved good recovery results. The research results can provide a reference technical route for the efficient recovery of micro-fine cassiterite from similar complex polymetallic tin ores.

To address the problems of high titanaugite silicate content in ilmenite and the low recovery of the existing gravity separation process from a certain area in Xinjiang, a systematic experimental study on flotation was conducted. Under the conditions of a raw ore TiO₂ grade of 15.64% and an ilmenite distribution rate of 77.37%, using a composite binary organic acid as the titanaugite depressant and adopting the beneficiation process flowsheet of "desulfurization reverse flotation-ilmenite direct flotation", an ilmenite concentrate with a yield of 26.02%, a TiO₂ grade of 45.62%, and a recovery of 76.27% was obtained through the closed-circuit test. Compared with the existing gravity separation process, the TiO₂ recovery rate in ilmenite concentrate by high intensity magnetic separation-flotation process is increased by 8.1 precentage points, and the ilmenite content in the concentrate reached 91.22%, which achieved the effective flotation separation of ilmenite and titanaugite. Zeta potential measurements and adsorption tests indicate that the composite binary organic acid has a greater impact on the surface potential of titanaugite and a higher adsorption amount, which increases the floatability difference between titanaugite and ilmenite, and provides a basis for their flotation separation.