Large model training and pushing all-in-one machine for underground coal mine

Abstract

The utility model relates to a large model training and pushing all-in-one machine for an underground coal mine. The all-in-one machine comprises a heterogeneous calculation module and an explosion-proof box body, the heterogeneous calculation module is arranged in the explosion-proof box body through a damping support, the heterogeneous calculation module is of a layered stacking structure and comprises a reasoning layer GPU acceleration card, a backboard, a CPU mainboard and a training layer NPU acceleration card which are sequentially arranged from top to bottom, the reasoning layer GPU acceleration card and the backboard, the CPU mainboard and the backboard, and the training layer NPU acceleration card and the backboard are respectively in electric connection and communication connection, and the backboard physically separates the reasoning layer GPU acceleration card from the training layer NPU acceleration card. According to the large model training and pushing all-in-one machine, the working efficiency of large model training and pushing service is improved, and the applicability and safety to underground coal mine scenes are enhanced.

Description

Technical Field

[0001] This utility model relates to the field of edge computing and large-scale model engineering deployment technology; specifically, this utility model relates to a large-scale model training and propulsion integrated machine for use in underground coal mines. Background Technology

[0002] In recent years, large-scale language models and multimodal models based on the Transformer architecture have demonstrated powerful cognitive and decision-making capabilities, and are penetrating from general internet scenarios to vertical industrial fields. However, the industrial deployment of large models faces a fundamental contradiction between "centralized cloud deployment" and "real-time edge requirements": although cloud deployment has abundant computing power, network latency is difficult to meet the millisecond-level response requirements of industrial sites; although pure edge inference has low latency, the fixed model cannot adapt to changes in working conditions, and edge computing power is insufficient to support continuous model optimization.

[0003] To address this contradiction, the industry has developed technical approaches such as "cloud-edge collaboration" and "federated learning," attempting to distribute training and inference tasks between the central cloud and the edge. However, these solutions still fundamentally rely on the cloud to dominate model optimization, with the edge only handling data collection and preliminary inference, failing to truly achieve autonomous evolution at the edge. More critically, existing large-scale model training and inference equipment is designed for general data centers or office environments, failing to consider the special constraints of extreme industrial environments such as flammable and explosive materials, strong electromagnetic interference, and limited space, making it virtually impossible to deploy in high-risk scenarios such as underground coal mines.

[0004] Therefore, the intelligent construction of underground coal mines has long been plagued by the dilemma of "having perception but no intelligence" or "weak intelligence relying on the cloud." A large number of sensors and monitoring devices are deployed underground, generating massive amounts of data. However, limited by network bandwidth and computing power distribution, this data is either stored locally and cannot be analyzed in real time, or uploaded to the surface for processing, leading to decision-making delays. Existing large-scale model application technologies and underground computing equipment in coal mines mainly suffer from the following shortcomings:

[0005] First, the training-inference separation architecture restricts the development of edge intelligence. Current large-scale model deployment models generally adopt either "cloud training, edge inference" or "fixed deployment of pre-trained models." The former requires frequent network communication for model updates, which is difficult to implement in the unstable network environment of coal mines; the latter solidifies model capabilities, making it impossible to adaptively optimize based on new equipment failure modes, geological changes, or operational process adjustments underground. Although some edge computing devices have lightweight model training capabilities, limited by computing power, they can only handle traditional machine learning algorithms and cannot support parameter fine-tuning and continuous learning of large models.

[0006] Second, general-purpose computing platforms lack adaptability to underground environments. Existing large-scale model training and propulsion integrated machines are mainly designed for data centers or edge computing rooms, and do not consider the special requirements of underground coal mines in terms of physical form, safety protection, and environmental adaptability. These devices are usually bulky, consume a lot of power, rely on air conditioning systems for heat dissipation, and do not have explosion-proof functions, posing serious safety hazards if deployed directly underground. While existing explosion-proof computers or servers in coal mines meet explosion-proof requirements, their hardware configurations can only support simple logic control or data forwarding. Their CPU computing power, memory bandwidth, and storage performance are insufficient to support the matrix operations and parameter storage requirements of large models, creating a technological gap of "safe but not intelligent, intelligent but unsafe".

[0007] Third, network dependence and data privacy are difficult to reconcile. Underground network conditions in coal mines are complex; fiber optic cables are easily broken due to mining activities, and wireless signals attenuate severely within tunnels, leading to frequent communication interruptions between the surface and underground. Current technologies rely on the cloud for model training and centralized management; once the network is interrupted, the edge devices cannot receive model updates, and the intelligent capabilities degrade over time. Furthermore, underground coal mine data involves core secrets such as production processes and geological information; uploading large amounts to the cloud poses a risk of data leakage, while local processing lacks the necessary computing power.

[0008] Fourth, there is a conflict between the model's generalization ability and specialized needs. General-purpose large models often perform poorly in specialized industrial scenarios, exhibiting limited understanding of coal mine-specific terminology, equipment operating patterns, and disaster symptom characteristics. Existing solutions require extensive fine-tuning of specialized data in the cloud, which is costly and time-consuming. More importantly, different mines have significantly different geological conditions, equipment configurations, and management models, making it difficult for a unified cloud model to adapt to individual needs, while independent cloud training for each mine faces technical and cost hurdles.

[0009] Therefore, existing technologies have not yet solved the core problem of how to deploy large model systems with autonomous training and real-time reasoning capabilities in underground coal mine environments, and there is an urgent need for an "integrated training and reasoning" edge intelligent device designed specifically for this scenario. Utility Model Content

[0010] In view of this, the present invention provides a large-scale model training and pushing integrated machine for underground coal mines, thereby solving or at least alleviating one or more of the above-mentioned problems and other problems existing in the prior art.

[0011] To achieve the aforementioned objectives, this utility model provides a large-scale training and propulsion integrated machine for underground coal mines. The integrated machine includes a heterogeneous computing module and an explosion-proof enclosure. The heterogeneous computing module is installed inside the explosion-proof enclosure via a shock-absorbing bracket. The heterogeneous computing module has a layered stacked structure, including an inference layer GPU accelerator card, a backplane, a CPU motherboard, and a training layer NPU accelerator card arranged sequentially from the top to the bottom. The inference layer GPU accelerator card is electrically and communicatively connected to the backplane, the CPU motherboard is connected to the backplane, and the training layer NPU accelerator card is connected to the backplane. The backplane physically separates the inference layer GPU accelerator card and the training layer NPU accelerator card.

[0012] In the all-in-one machine described above, optionally, the inference layer GPU accelerator card is connected to the backplane via a PCIe interface, the CPU motherboard is connected to the backplane via a connector, and the training layer NPU accelerator card is connected to the backplane via a PCIe interface.

[0013] In the all-in-one machine described above, optionally, a first memory chip is mounted on the substrate of the inference layer GPU accelerator card, and a second memory chip is mounted on the substrate of the training layer NPU accelerator card. The first memory chip and the second memory chip are electrically and communicatively connected through a high-speed inter-board connector disposed on the backplane.

[0014] In the all-in-one machine described above, optionally, both the first memory chip and the second memory chip are HBM chips. The first memory chip is connected to the inference layer GPU accelerator card via an NVLink interface, the second memory chip is connected to the training layer NPU accelerator card via a CXL interface, and the CPU motherboard is connected to the inter-board high-speed connector via a DDR5 memory controller.

[0015] In the all-in-one machine described above, optionally, the explosion-proof enclosure includes an explosion-proof shell, an intrinsically safe electrical interface, an intrinsically safe power supply system, and a heat pipe phase change heat dissipation system disposed on the explosion-proof shell.

[0016] In the all-in-one machine described above, optionally, the intrinsically safe electrical interface is energy-limited by a safety barrier, including but not limited to an Ethernet interface, an RS485 interface, and a CAN bus interface.

[0017] In the all-in-one machine described above, optionally, the explosion-proof enclosure meets the ExdIMb explosion-proof level requirements and is divided into a wiring cavity and a main cavity by an explosion-proof mating surface provided inside the explosion-proof enclosure. The wiring cavity is provided with an explosion-proof cable introduction device, an explosion-proof interface and the intrinsically safe electrical interface on the outside. The heterogeneous computing module is provided inside the main cavity, and a main cavity front door is provided outside the main cavity.

[0018] Optionally, in the all-in-one machine described above, the explosion-proof enclosure further includes an explosion-proof human-machine interface, which includes an explosion-proof observation window embedded in the front door of the main cavity, and the explosion-proof observation window is equipped with a touch screen.

[0019] In the all-in-one machine described above, optionally, the heat pipe phase change heat dissipation system is a split heat pipe, including an evaporation end, a condensation end, and an insulating section connecting the evaporation end and the condensation end. The evaporation end is a heat spreader plate and is closely attached to the surface of the inference layer GPU accelerator card and the training layer NPU accelerator card. The insulating section is an inclined insulating pipe that connects to the condensation end from the evaporation end through a sealing sleeve disposed on the explosion-proof housing. The condensation end is a heat dissipation fin assembly disposed outside the explosion-proof housing.

[0020] In the all-in-one machine described above, optionally, an explosion-proof axial fan is provided outside the heat dissipation fin assembly, and the explosion-proof axial fan is connected to the intrinsically safe power system through an explosion-proof junction box.

[0021] This utility model discloses a large-scale model training and inference integrated machine for underground coal mines. At the hardware level, it integrates a training layer processor and an inference layer processor, enabling the large-scale model training and inference equipment to have autonomous evolution capabilities independent of the cloud. This reduces reliance on network bandwidth and stability, and shortens the response time for critical intelligent decisions. Furthermore, the physical separation of the training layer processor and the inference layer processor avoids resource contention in training and inference tasks, improving the real-time performance and accuracy of the large-scale model inference service. Simultaneously, this utility model employs an explosion-proof design, enhancing the applicability and safety of the large-scale model training and inference equipment in underground coal mine scenarios.

[0022] In optional embodiments, this invention uses a first memory chip and a second memory chip to form a unified memory pool, achieving cache-consistent access and further improving the real-time performance and efficiency of large-scale model training and propulsion in underground coal mines. In other optional embodiments, this invention employs a separate heat pipe and an external explosion-proof axial flow fan to physically isolate the heat source from the heat dissipation end, allowing heat to dissipate into the air of the underground roadway. Under explosion-proof and sealed conditions, this improves the heat dissipation efficiency of the high-power large-scale model training and propulsion equipment, thereby further enhancing the working efficiency of the integrated large-scale model training and propulsion machine. Attached Figure Description

[0023] The disclosure of this utility model will become more apparent with reference to the accompanying drawings. It should be understood that these drawings are for illustrative purposes only and are not intended to limit the scope of protection of this utility model. In the drawings:

[0024] Figure 1 This is a structural schematic diagram of one embodiment of the present invention.

[0025] Figure 2 This is a schematic block diagram illustrating the structure of one embodiment of the heterogeneous computing module in this utility model.

[0026] Figure 3 This is a structural schematic diagram of one embodiment of the explosion-proof enclosure in this utility model.

[0027] Reference numerals: 11-Inference layer GPU accelerator card; 12-Backplane; 13-CPU motherboard; 14-Training layer NPU accelerator card; 15-First memory chip; 16-Second memory chip; 17-High-speed connector between boards; 21-Explosion-proof enclosure; 22-Intrinsically safe electrical interface; 23-Explosion-proof cable entry device; 24-Heat pipe phase change cooling system; 25-Explosion-proof human-machine interface; 26-Lifting lug; 27-Explosion-proof interface; 28-Main cavity front door; 29-Intrinsically safe power system; 30-Explosion-proof axial flow fan; 31-Indicator light. Detailed Implementation

[0028] Referring to the accompanying drawings and specific embodiments, the structure, composition, features, and advantages of the large-scale training and pushing integrated machine for underground coal mines of this utility model will be described below by way of example. However, all descriptions should not be used to limit this utility model in any way.

[0029] Furthermore, for any single technical feature described or implied in the embodiments mentioned herein, or any single technical feature shown or implied in the various drawings, the present invention still allows for any combination or deletion of these technical features (or their equivalents) without any technical obstacle. Therefore, these further embodiments according to the present invention should also be considered within the scope of the description herein.

[0030] It should also be noted that the terms "internal", "top", "bottom", "external", etc., indicate the orientation or positional relationship based on the orientation or positional relationship of the large-scale training and pushing integrated machine used in underground coal mines as shown in the attached drawings. They are only for the convenience of describing this disclosure and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this disclosure.

[0031] Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of those features.

[0032] Figure 1 This is a schematic diagram of the structure of one embodiment of the present utility model. Figure 2The heterogeneous computing modules shown and such Figure 3 The explosion-proof enclosure shown together constitute the large-scale training and pushing integrated machine for underground coal mines of this utility model.

[0033] Combination Figure 1 and Figure 2 As can be seen, the heterogeneous computing module has a layered stacked structure, consisting of an inference layer GPU accelerator card 11, a backplane 12, a CPU motherboard 13, and a training layer NPU accelerator card 14 from top to bottom. The inference layer GPU accelerator card 11 and the training layer NPU accelerator card 14 are respectively inserted into the first and third interfaces on the backplane 12 via PCIe connectors. PCIe (Peripheral Component Interconnect Express) connectors are a computer expansion bus interface capable of simultaneously achieving electrical, communication, and mechanical connections, and are available in different electrical specifications such as PCIe x4, PCIe x8, and PCIe x16. In optional embodiments, the inference layer GPU accelerator card 11 can be a full-height, double-width design connected to the backplane 12 via a PCIe x16 interface, while the training layer NPU accelerator card 14 can be a half-height, single-width design connected to the backplane 12 via a PCIe x8 interface. In other optional embodiments, the CPU motherboard 13 can integrate a CPU, chipset, and expansion interfaces, and the CPU motherboard 13 is inserted into the second interface on the backplane 12 via a connector on the expansion interface. Therefore, the backplane 12 functions as a bus interface for connecting the inference layer GPU accelerator card 11, the CPU motherboard 13, and the training layer NPU accelerator card 14.

[0034] In this all-in-one machine, the CPU motherboard 13 is used for overall system scheduling, data preprocessing, task orchestration, and heterogeneous resource management of the heterogeneous computing modules. It can optimize memory access patterns and caching strategies for large model workloads to improve data locality in large models under the Transformer architecture. The inference layer GPU accelerator card 11 is used for forward inference computation of large models and can support dynamic batch processing technology and concurrent loading of multiple models. The training layer NPU accelerator card 14 is used for training large models and provides hardware acceleration for backpropagation, gradient calculation, and parameter updates during model training. Since the inference task and the training task have different requirements for numerical accuracy and memory bandwidth, the inference layer GPU accelerator card 11 and the training layer NPU accelerator card 14 are physically separated by the backplane 12. This not only achieves efficient edge training while controlling power consumption budget, but also avoids resource contention between training and inference tasks, ensuring that the real-time performance of the inference service is not affected by the training task.

[0035] like Figure 2As shown, the inference layer GPU accelerator card 11 is connected to the first memory chip 15 via an NVLink interface, and the training layer NPU accelerator card 14 is connected to the second memory chip 16 via a CXL interface. The first memory chip 15 and the second memory chip 16 are connected via a high-speed inter-board connector 17 mounted on the backplane 12, thus forming a unified memory pool. Simultaneously, the CPU motherboard 13 is connected to the high-speed inter-board connector 17 via a DDR5 memory controller. Both the NVLink (Nvidia Link) interface and the CXL (Compute Express Link) interface are high-speed interconnect interfaces, enabling both electrical and communication connections. The DDR5 (Double Data Rate 5) memory controller is a data interaction module integrated into the CPU.

[0036] In an optional embodiment, both the first memory chip 15 and the second memory chip 16 are 16GB HBM chips, respectively mounted on the substrates of the inference layer GPU accelerator card 11 and the training layer NPU accelerator card 14 using multi-chip packaging technology. Logical unification is achieved through the inter-board high-speed connector 17, thereby providing a shared data exchange space for the inference layer GPU accelerator card 11 and the training layer NPU accelerator card 14. HBM (High Bandwidth Memory) chips are high-bandwidth memory chips. Furthermore, the CPU motherboard 13 is connected to the unified memory pool via the inter-board high-speed connector 17, with the DDR5 memory controller providing 51.2GB / s bandwidth; the inference layer GPU accelerator card 11 is directly connected to the unified memory pool via the NVLink 3.0 interface, with a bandwidth of 600GB / s, and supports unified memory addressing; the training layer NPU accelerator card 14 is directly connected to the unified memory pool via the CXL 1.1 interface, with a bandwidth of 32GB / s, and supports cache-coherent access.

[0037] This heterogeneous memory architecture enables the CPU motherboard 13, the inference layer GPU accelerator card 11, and the training layer NPU accelerator card 14 to access the same model parameters and dataset with zero copies, eliminating the data transfer bottleneck between the CPU motherboard and the accelerator card in traditional architectures. Furthermore, the unified memory pool supports unified addressing of training and inference data, allowing high-quality samples collected during inference to be directly used for subsequent training, reducing additional data copying overhead.

[0038] In a further optional embodiment, model lifecycle management software can be configured in the CPU motherboard 13. This software runs in a containerized environment and provides the all-in-one machine of this utility model with functions such as model repository, version control, automated training process and model evaluation.

[0039] Based on such Figure 2The heterogeneous computing module shown in this utility model integrates the training and inference capabilities of a large model at the hardware level, realizing the full lifecycle management of a large model within a single device at the coal mine site. This enables the large model to evolve autonomously without relying on the cloud, thereby reducing dependence on network bandwidth and stability and shortening the response time of key intelligent decisions from seconds to milliseconds.

[0040] Furthermore, this utility model sets the heterogeneous computing module as described above in such a way as... Figure 3 The explosion-proof enclosure shown enhances the applicability of the large-scale training and propulsion integrated machine to enclosed industrial environments such as underground coal mines, ensuring the safe and stable operation of high-computing-power equipment even in high-risk and special environments. In an optional embodiment, the heterogeneous computing module can be fixed in the explosion-proof enclosure by vibration-damping brackets at the four corners. The vibration-damping brackets use silicone rubber shock absorbers to absorb the vibration and impact generated by underground blasting operations.

[0041] from Figure 3 As can be seen from the diagram, the explosion-proof enclosure may include an explosion-proof shell 21 and an intrinsically safe electrical interface 22, an intrinsically safe power supply system 29, and a heat pipe phase change cooling system 24 disposed on the explosion-proof shell 21. Intrinsically safe is a protection level for explosion-proof electrical equipment.

[0042] Optionally, the explosion-proof enclosure 21 is integrally cast from high-strength cast aluminum alloy with alloy code ZL114A, and the wall thickness is 15-20mm. In this embodiment, the explosion-proof enclosure 21 has a rectangular horizontal structure, and the dimensions can be optionally set to 600mm×450mm×300mm (length×width×height), meeting the ExdIMb explosion-proof rating requirements. The ExdIMb explosion-proof rating indicates that the explosion-proof type of the explosion-proof enclosure 21 is d (explosion-proof type), suitable for Class I environments (coal mine environment), and the equipment protection level is Mb (high), capable of withstanding the internal explosion pressure and preventing the explosion from propagating to the external gas environment. In other optional embodiments, the explosion-proof enclosure 21 can also be of other shapes, sizes, and explosion-proof standards that meet the requirements of the underground environment.

[0043] In a further optional embodiment, the interior of the explosion-proof enclosure 21 can be divided into a wiring cavity and a main cavity by an explosion-proof mating surface, wherein the dimensions of the wiring cavity are optionally set to 150mm × 450mm × 300mm, and the dimensions of the main cavity are optionally set to 430mm × 450mm × 300mm. Figure 3As shown, a raised explosion-proof cable entry device 23 is provided on the side of the explosion-proof housing 21 for introducing power cables and optical fibers; four raised explosion-proof interfaces 27 are provided on each of the two side walls corresponding to the wiring cavity positions for leading out intrinsically safe signal lines. Optionally, the explosion-proof cable entry device 23 is a compression nut type with an M36×2 thread specification, and the explosion-proof interface 27 has an M20×1.5 thread specification. This utility model does not limit the number, position, or specific specifications of the explosion-proof cable entry device 23 and the explosion-proof interface 27. Additionally, four guide rail grooves can optionally be cast on the inner side walls of the main cavity for installing heterogeneous computing modules, and the spacing of the guide rail grooves meets the international structural standard IEC 60297.

[0044] from Figure 3 As can be seen, the exterior of the explosion-proof enclosure 21 is also equipped with a main cavity front door 28 with a quick-opening structure. The main cavity front door 28 has handles on both sides in the horizontal direction for easy and quick opening and closing. In an optional embodiment, the door cover of the main cavity front door 28 is fastened to the shell of the explosion-proof enclosure 21 using M12 stainless steel bolts. The width of the explosion-proof mating surface is ≥25mm, and the gap is ≤0.15mm, meeting the explosion-proof parameter requirements of GB 3836.2, thereby ensuring the reliability of the explosion-proof design.

[0045] In this embodiment, the explosion-proof housing 21 is further provided with an explosion-proof human-machine interface 25, namely an explosion-proof observation window embedded in the middle of the front door 28 of the main cavity, and a touch screen display screen thereon. Optionally, the explosion-proof observation window adopts a sealed structure of double-layer tempered glass and sintered metal, with a light transmission diameter of 120mm, which can observe the status of indicator lights inside the explosion-proof housing 21, and is connected through an intrinsically safe circuit to support local parameter configuration and emergency operation, thereby meeting the on-site maintenance needs of underground coal mines. Further optionally, three indicator lights 31 are also provided on the side of the explosion-proof human-machine interface 25, which respectively indicate the normal or abnormal status of power supply, equipment operation, and network communication.

[0046] like Figure 3As shown, an intrinsically safe electrical interface 22 and an intrinsically safe power supply system 29 are provided on the side of the explosion-proof enclosure 21. The intrinsically safe electrical interface 22 is used to achieve a safe connection between the integrated machine and underground peripherals. Energy is limited by a safety barrier to ensure that the output voltage and current are insufficient to ignite an explosive gas mixture in a fault condition. The interface configuration of the intrinsically safe electrical interface 22 may include, but is not limited to: multiple intrinsically safe gigabit Ethernet interfaces for accessing the underground industrial ring network and supporting communication between large-scale model applications and the upper-level system; an intrinsically safe RS485 interface for connecting various environmental sensors and equipment controllers to collect multi-source data required for training; and an intrinsically safe CAN bus interface for real-time linkage with the underground PLC control system to achieve rapid execution of inference results. These interfaces enable the integrated machine of this invention to integrate into existing underground monitoring networks, becoming an intelligent data processing hub.

[0047] The intrinsically safe power system 29 is used to convert the voltage of the underground AC power grid to various DC voltage levels required by the integrated machine. Optionally, the power module adopts an explosion-proof and intrinsically safe design, with input terminals having the ability to resist power grid fluctuations and surges, adapting to the characteristics of poor power quality underground; the output terminals provide safe power to the intrinsically safe electrical interface 22 through multiple isolation and energy limiting circuits. At the same time, the intrinsically safe power system 29 can be equipped with an explosion-proof uninterruptible power supply to maintain the computing platform to complete the current inference task, save the training state, and safely shut down when the external power grid is interrupted, avoiding data loss and model damage.

[0048] from Figure 3 As can be seen, a heat pipe phase change heat dissipation system 24 is provided on the top of the explosion-proof enclosure 21. The heat pipe phase change heat dissipation system 24 is a split heat pipe, including a condenser end heat dissipation fin assembly visible on the outside of the explosion-proof enclosure 21, an evaporator end not shown in the figure, and an insulation section connecting the evaporator end and the condenser end.

[0049] In an optional embodiment, the evaporation end uses a copper vapor chamber with dimensions of 200mm × 150mm × 8mm, which is closely attached to the surfaces of the GPU accelerator card 11 in the inference layer and the NPU accelerator card 14 in the training layer, and the contact thermal resistance is reduced by an indium-based thermal interface material (thermal conductivity ≥ 80W / m·K). The interior of the vapor chamber is a vacuum cavity filled with acetone working fluid, and liquid reflux is achieved through a capillary copper powder sintered core. The insulation section uses multiple copper-water insulation tubes, for example, six with a diameter of 10mm, which pass from the evaporation end through a sealing sleeve set on the side wall of the explosion-proof housing 21 and connect to the condensation end. The sealing sleeve adopts a three-layer brazed structure of stainless steel-ceramic-stainless steel, which ensures both explosion-proof sealing and heat conduction. The length of the insulation tube inside the explosion-proof housing 21 can optionally be 80mm, and the extension length outside is 150mm, with the overall arrangement at an angle of, for example, 15°, to facilitate gravity reflux of the condensate.

[0050] Optionally, the condenser end heat sink assembly can be a total of 50 aluminum heat sink fins, with dimensions of 250mm × 200mm × 60mm and a fin spacing of 2.5mm, connected to the end of the heat insulation pipe via thermally conductive epoxy resin. Further, an explosion-proof axial fan 30 can be installed outside the heat sink assembly and connected to the intrinsically safe power system 29 via an explosion-proof junction box. The explosion-proof axial fan 30 optionally meets the ExdI explosion-proof standard, with an airflow of 120 CFM, an air pressure of 80 Pa, and a noise level ≤65 dB(A). The speed of the explosion-proof axial fan 30 can be steplessly adjusted via a PWM signal (duty cycle 0-100%) according to the chip temperature of the inference layer GPU accelerator card 11 and the training layer NPU accelerator card 14.

[0051] The heat pipe phase change heat dissipation system 24 is designed to physically isolate the heat source from the heat dissipation end, dissipating the heat into the tunnel air. This not only meets the explosion-proof sealing requirements but also achieves the same efficiency as open heat dissipation, enabling the high-power chips inside the explosion-proof enclosure to operate continuously at full load underground.

[0052] Optionally, a device such as [missing information] is provided on the top of the explosion-proof enclosure. Figure 3 The lifting lug 26 shown is used to transport and place the integrated machine of this utility model in the underground roadway; a symmetrical support frame is also provided at the bottom of the explosion-proof box to adaptably fix the integrated machine in the underground roadway environment, thereby improving its safety and stability during use.

[0053] Based on the aforementioned hardware structure of heterogeneous computing modules and explosion-proof enclosures, this all-in-one machine can realize a large-scale model training and pushing process and operating mechanism adapted to the underground coal mine environment. The following example illustrates a feasible large-scale model training and pushing process.

[0054] Real-time inference phase: Multi-source data from underground coal mines is connected to the integrated machine via intrinsically safe electrical interface 22. After preprocessing by the CPU motherboard 13, it is sent to the GPU accelerator card 11 in the inference layer and analyzed in real time based on the current version of the large model. Various decision information is output and sent to the actuator or pushed to the underground monitoring terminal via the intrinsically safe electrical interface 22. At this time, the NPU accelerator card 14 in the training layer is in a low-power standby state, and the concentrated computing power ensures low latency inference.

[0055] Data accumulation phase: During inference, the all-in-one machine automatically identifies high-value samples, de-identifies these samples, stores them in a unified memory pool's data cache, and synchronizes them to local storage to build an incremental training dataset. This process is entirely local, without uploading to the cloud, thus protecting data privacy and saving network bandwidth.

[0056] Local training phase: When preset trigger conditions are met, the training layer NPU accelerator card 14 is activated, and incremental learning algorithms are used to fine-tune the parameters of the basic large model. Training tasks can be performed during off-peak periods of downhole operations to avoid affecting real-time inference performance.

[0057] Model update phase: After training is completed, the new model is tested and hot-updated to the inference layer GPU accelerator card 11, replacing the old model. At the same time, model parameter summary information can be selectively uploaded to the ground center via the intrinsically safe electrical interface 22.

[0058] Through the structural design of this utility model, not only is the integrated capability of large models to perform continuous training and real-time inference at the edge enhanced, but the applicability of the large model training and inference integrated machine to extreme environments such as flammable and explosive environments, limited space, and unstable networks in underground coal mines is also improved, while meeting the high requirements for computing power and safety.

[0059] The integrated machine of this invention has completed laboratory verification of its core modules and underground environment adaptability testing. Regarding the performance verification of large-scale model training and inference, tests were conducted based on an open-source 7B-parameter general-purpose large-scale model and a self-built coal mine safety professional dataset. Results show that the GPU accelerator card in the inference layer has a single token generation latency of less than 50 milliseconds in INT8 quantization mode, meeting real-time interaction requirements. The NPU accelerator card in the training layer uses LoRA fine-tuning technology, and 100-step incremental training on the local dataset takes approximately 30 minutes. The model improves accuracy in understanding professional terminology and identifying fault modes by 12%, verifying the effectiveness of edge training. The training and inference task switching latency is less than 1 second, and the inference service is uninterrupted during model hot updates. In terms of environmental adaptability testing, type tests were conducted on the explosion-proof enclosure prototype, and its explosion-proof performance met the standards of GB 3836.1-2021 and GB 3836.2-2021. The heat pipe phase change heat dissipation system operated continuously for 72 hours at an ambient temperature of 40℃ and a total power consumption of 800W, with the core chip temperature remaining stable within a safe range. The seismic test, damp heat test, and electromagnetic compatibility test all met the requirements of coal mine equipment standards such as MT / T 899.

[0060] This utility model's all-in-one machine can serve as the core computing infrastructure for realizing unmanned and minimally manned mining in underground coal mines. It is suitable for intelligent management and control of key areas such as fully mechanized mining faces, tunneling roadways, main transportation systems, central substations, and pump rooms in underground coal mines. It can support various large-scale model application scenarios, including intelligent diagnosis of equipment faults, prediction and early warning of roof pressure, intelligent identification of gas anomalies, safety analysis of personnel behavior, intelligent inference of geological structures, optimization decision-making for ventilation systems, and autonomous handling of emergency scenarios.

[0061] The technical scope of this utility model is not limited to the contents of the above description. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this utility model, and all such modifications and variations should fall within the scope of this utility model.

Claims

1. A large-scale training and pushing integrated machine for underground coal mines, characterized in that, The all-in-one machine includes a heterogeneous computing module and an explosion-proof enclosure. The heterogeneous computing module is installed inside the explosion-proof enclosure via a shock-absorbing bracket. The heterogeneous computing module has a layered stacked structure, including an inference layer GPU accelerator card (11), a backplane (12), a CPU motherboard (13), and a training layer NPU accelerator card (14) arranged sequentially from the top to the bottom. The inference layer GPU accelerator card (11) is electrically connected to the backplane (12), the CPU motherboard (13) is connected to the backplane (12), and the training layer NPU accelerator card (14) is connected to the backplane (12). The backplane (12) physically separates the inference layer GPU accelerator card (11) from the training layer NPU accelerator card (14).

2. The all-in-one machine as described in claim 1, characterized in that, The inference layer GPU accelerator card (11) is connected to the backplane (12) via a PCIe interface, the CPU motherboard (13) is connected to the backplane (12) via a connector, and the training layer NPU accelerator card (14) is connected to the backplane (12) via a PCIe interface.

3. The all-in-one machine as described in claim 1, characterized in that, The inference layer GPU accelerator card (11) has a first memory chip (15) mounted on its substrate, and the training layer NPU accelerator card (14) has a second memory chip (16) mounted on its substrate. The first memory chip (15) and the second memory chip (16) are electrically and communicatively connected through a high-speed inter-board connector (17) provided on the backplane (12).

4. The all-in-one machine as described in claim 3, characterized in that, Both the first memory chip (15) and the second memory chip (16) are HBM chips. The first memory chip (15) is connected to the inference layer GPU accelerator card (11) via an NVLink interface. The second memory chip (16) is connected to the training layer NPU accelerator card (14) via a CXL interface. The CPU motherboard (13) is connected to the inter-board high-speed connector (17) via a DDR5 memory controller.

5. The all-in-one machine as described in claim 1, characterized in that, The explosion-proof enclosure includes an explosion-proof shell (21), an intrinsically safe electrical interface (22), an intrinsically safe power supply system (29), and a heat pipe phase change heat dissipation system (24) disposed on the explosion-proof shell (21).

6. The all-in-one machine as described in claim 5, characterized in that, The intrinsically safe electrical interface (22) is energy limited by a safety barrier, including but not limited to an Ethernet interface, an RS485 interface and a CAN bus interface.

7. The all-in-one machine as described in claim 5, characterized in that, The explosion-proof enclosure (21) meets the ExdIMb explosion-proof level requirements and is divided into a wiring cavity and a main cavity by an explosion-proof mating surface inside the explosion-proof enclosure (21). The wiring cavity is provided with an explosion-proof cable introduction device (23), an explosion-proof interface (27) and an intrinsically safe electrical interface (22) on the outside. The heterogeneous computing module is provided inside the main cavity, and a main cavity front door (28) is provided outside the main cavity.

8. The all-in-one machine as described in claim 7, characterized in that, The explosion-proof enclosure also includes an explosion-proof human-machine interface (25), which includes an explosion-proof observation window embedded in the front door (28) of the main cavity, and the explosion-proof observation window is equipped with a touch screen.

9. The all-in-one machine as described in claim 5, characterized in that, The heat pipe phase change heat dissipation system (24) is a split heat pipe, including an evaporation end, a condensation end, and an insulating section connecting the evaporation end and the condensation end. The evaporation end is a heat spreader plate and is closely attached to the surface of the inference layer GPU accelerator card (11) and the training layer NPU accelerator card (14). The insulating section is an inclined insulating pipe that is connected from the evaporation end to the condensation end through a sealing sleeve set on the explosion-proof shell (21). The condensation end is a heat dissipation fin group set outside the explosion-proof shell (21).

10. The all-in-one machine as described in claim 9, characterized in that, An explosion-proof axial flow fan (30) is provided outside the heat dissipation fin assembly. The explosion-proof axial flow fan (30) is connected to the intrinsically safe power system (29) through an explosion-proof junction box.

Images