Machine room operation and maintenance method and device and electronic equipment

Abstract

This invention provides a data center operation and maintenance method, device, and electronic equipment. It trains a neural radiation field model using multimodal data streams to construct a spatiotemporally continuously updated current physical field twin model. When a data center operation and maintenance event occurs, it reduces the dimensionality of the high-dimensional target tensor data extracted from the current physical field twin model into a physical tensor summary. This summary, supplemented by retrieved decision reference information, is then input into a large language model to generate automated data center operation and maintenance decisions. This forms a full-link adaptive closed-loop mechanism integrating continuous scene reconstruction and knowledge retrieval generation. It overcomes the dual bottlenecks of existing 3D GIS relying on static modeling and the difficulty of large language models in handling high-dimensional physical field data. It can accurately reflect the data center's thermal distribution and environmental evolution, automatically outputting safety-verified optimization strategies, significantly reducing the intensity of manual inspections, shortening the location time of major faults to minutes, and deeply releasing the space for data center energy efficiency optimization.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for data center computer rooms, and in particular to a method, apparatus and electronic equipment for the operation and maintenance of a computer room. Background Technology

[0002] With the continuous expansion of data center scale and the increase in computing power density, the thermal distribution, airflow organization, and energy consumption status within the data center exhibit highly dynamic complexity. To ensure the safe and stable operation of the data center and improve energy efficiency management, achieving continuous and precise perception and intelligent automated decision-making of the data center spatial environment has become a crucial overall technical requirement in the current data center operation and maintenance field. To meet these technical requirements, existing data center operation and maintenance technologies typically employ static modeling based on 3D GIS (Geographic Information System) combined with rule-driven monitoring logic. When formulating operation and maintenance control strategies, they mainly rely on traditional expert systems to match preset threshold alarms with human experience; or, when introducing large language models, they directly input scattered, discrete structured sensor log text into the model for basic text-level reasoning to output corresponding energy-saving or fault-response suggestions.

[0003] The aforementioned existing technical solutions have significant shortcomings in practical applications. On the one hand, traditional static 3D model architectures cannot continuously and accurately reflect the real-time thermal distribution of data center equipment and the dynamic evolution of the 3D physical environment at design boundaries. On the other hand, existing intelligent inference model architectures are mainly designed for one-dimensional or discrete text logic. When faced with complex, high-dimensional continuous physical field data in data centers, their original computational paradigms cannot effectively extract features and achieve deep logical coupling for such high-dimensional spatiotemporal physical situations. Consequently, when facing complex operation and maintenance scenarios across devices, it is difficult to perform global coordination and accurate automated inference, ultimately severely limiting the energy efficiency optimization space of data centers, slowing fault location, and still requiring frequent manual intervention. Summary of the Invention

[0004] This invention provides a data center operation and maintenance method, device, and electronic equipment to address the shortcomings of existing technologies, such as static three-dimensional models failing to continuously and accurately reflect the dynamic evolution of the data center's physical environment, and existing intelligent reasoning models struggling to effectively extract features and couple deep logic, resulting in difficulties in achieving global coordination and accurate automated reasoning in complex scenarios.

[0005] This invention provides a method for the operation and maintenance of a data center, including: Acquire multimodal sensing data from the data center; The multimodal sensing data is input into the neural radiation field model to reconstruct the previous physical field twin model and construct the current physical field twin model of the computer room. In response to a detected data center maintenance event, target tensor data of the target region corresponding to the data center maintenance event is extracted from the current physical field twin model, and the target tensor data is integrated to generate a physical tensor summary of the target region. Decision reference information associated with the target area is retrieved, and the decision reference information and the physical tensor summary are input into the large language model to obtain the data center operation and maintenance decision output by the large language model.

[0006] According to a data center operation and maintenance method provided by the present invention, after obtaining the data center operation and maintenance decision output by the large language model, the method includes: Determine and execute the target action sequence corresponding to the data center operation and maintenance decision; Collect feedback status data of the computer room after the target action sequence is completed; Obtain the predicted state data generated based on the current physical field twin model for the target action sequence; Calculate the state deviation between the feedback state data and the predicted state data; If the state deviation exceeds a preset state deviation threshold, a reconstruction operation for the target region is triggered to obtain an optimized current physical field twin model. Based on the state deviation, the large language model is fine-tuned to obtain an optimized large language model.

[0007] According to a data center operation and maintenance method provided by the present invention, the step of determining and executing the target action sequence corresponding to the data center operation and maintenance decision includes: Extract the initial operation parameters corresponding to the data center operation and maintenance decision, and determine whether the initial operation parameters meet the preset data center security boundary specifications; Detect the execution timing relationship and resource mutual exclusion relationship among different operation parameters in the initial operation parameters; The target action sequence is obtained based on the initial operation parameters that meet the data center security boundary specifications and do not have execution timing and resource conflicts.

[0008] According to a data center operation and maintenance method provided by the present invention, the target tensor data includes the three-dimensional temperature tensor, airflow vector field tensor, and power consumption tensor of the target area at the current moment; The process of integrating the target tensor data to generate a physical tensor summary of the target region includes: Spatial dimension reduction projection is performed on the three-dimensional temperature tensor, the airflow vector field tensor, and the power consumption tensor to extract the environmental features and equipment load features of the target area; Based on the environmental characteristics and the load characteristics, the physical tensor summary is constructed.

[0009] According to a data center operation and maintenance method provided by the present invention, the step of inputting the multimodal sensing data into a neural radiation field model to reconstruct the previous physical field twin model and construct the current physical field twin model of the data center includes: The sensor sequence data is extracted from the multimodal sensor data by a preset sliding window, and the sensor sequence data is input into the neural radiation field model. The current spatial density and current color parameters of the previous physical field twin model are updated based on incremental training. The high-resolution network parameters in the previous physical twin model are converted into a sparse voxel mesh; Based on the sparse voxel mesh, the current spatial density, and the current color parameters, a lookup table index for spatial location and rendering attributes is constructed, and the current physical field twin model is reconstructed.

[0010] According to a data center operation and maintenance method provided by the present invention, the step of converting the high-resolution network parameters in the previous physics twin model into a sparse voxel mesh includes: Obtain the spatial location of the equipment in the computer room; Based on the spatial location of the equipment, the computer room space is divided into an adjacent area and a background area. The sparse voxel mesh is obtained by sampling the high-resolution network parameters according to the first sampling density corresponding to the device's adjacent area and the second sampling density corresponding to the device's background area. The first sampling density is higher than the second sampling density.

[0011] According to a data center operation and maintenance method provided by the present invention, the decision reference information associated with the target area includes at least one of the following: topology connection associated with the target area, historical operation and maintenance records, energy quota, and security threshold. The topology connection includes at least one of spatial location topology, power supply cascade topology, and cooling coverage topology.

[0012] According to a data center operation and maintenance method provided by the present invention, the step of acquiring multimodal sensing data of the data center includes: Acquire initial multi-source sensor data for the computer room; Perform a rendering consistency check on the image data and point cloud data in the initial multi-source sensor data to obtain the rendering consistency check result; Based on the rendering consistency test results, the initial multi-source sensor data is filtered to obtain the multimodal sensing data.

[0013] The present invention also provides a data center operation and maintenance device, comprising: The acquisition unit acquires multimodal sensing data from the computer room; The reconstruction unit inputs the multimodal sensing data into the neural radiation field model to reconstruct the previous physical field twin model and build the current physical field twin model of the computer room. The tensor summary generation unit, in response to a detected data center operation and maintenance event, extracts target tensor data of the target region corresponding to the data center operation and maintenance event from the current physical field twin model, integrates the target tensor data, and generates a physical tensor summary of the target region. The decision-making unit retrieves decision reference information associated with the target area, inputs the decision reference information and the physical tensor summary into the large language model, and obtains the data center operation and maintenance decision output by the large language model.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the operation and maintenance method of the computer room as described above.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the operation and maintenance method of the computer room as described above.

[0016] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the operation and maintenance method of the computer room as described above.

[0017] The data center operation and maintenance method, device, and electronic equipment provided by this invention train a neural radiation field model driven by multimodal data streams to construct a spatiotemporally continuously updated current physical field twin model. When a data center operation and maintenance event occurs, the high-dimensional target tensor data extracted from the current physical field twin model is reduced to a physical tensor summary, supplemented by retrieved decision reference information, and jointly input into a large language model to generate automated data center operation and maintenance decisions. This forms a full-link adaptive closed-loop mechanism integrating continuous scene reconstruction and knowledge retrieval generation, breaking through the dual bottlenecks of existing 3D GIS relying on static modeling and the difficulty of large language models in handling high-dimensional physical field data. It can accurately reflect the heat distribution and environmental evolution of the data center, automatically output optimized strategies verified by safety, significantly reduce the intensity of manual inspections, shorten the location time of major faults to the minute level, and deeply release the space for data center energy efficiency optimization. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is one of the flowcharts illustrating the operation and maintenance method for a data center provided by the present invention; Figure 2 This is the second flowchart illustrating the operation and maintenance method for a data center provided by this invention; Figure 3 This is a schematic diagram of the structure of the computer room operation and maintenance device provided by the present invention; Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0021] It should be understood that the terms "first" and "second" 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. Therefore, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0022] Furthermore, all actions involving the acquisition of signals, information, or data in this application are carried out in compliance with the relevant data protection laws and policies of the country where the application is located, and with authorization from the owner of the relevant device.

[0023] To address the aforementioned problems, this invention provides a data center operation and maintenance method to achieve efficient and accurate data center operation and maintenance. The execution entity of this method can be an electronic device with data processing and model inference capabilities, such as an edge computing server deployed locally in the data center, a data center operation and maintenance management node, or a cloud server cluster. Figure 1 This is one of the flowcharts illustrating the data center operation and maintenance method provided by the present invention, such as... Figure 1 As shown, the method includes: Step 110: Obtain multimodal sensing data for the computer room.

[0024] Here, multimodal sensing data refers to a collection of information collected from different physical dimensions and device interfaces to comprehensively reflect the real-time operating status and spatial environment of the data center.

[0025] Specifically, raw data is continuously acquired first through various sensing devices deployed at the edge of the computer room. This data, ranging from macroscopic to microscopic, can include visual image frames captured by high-resolution cameras, three-dimensional point clouds output by LiDAR scanning, and various raw environmental quantities such as temperature, humidity, energy consumption, power, and airflow reported synchronously by environmental sensors.

[0026] After acquisition, a unified time synchronization mechanism can be used to perform millisecond-level time-series alignment of multiple heterogeneous data streams, and a denoising algorithm can be invoked to eliminate abnormal noise caused by dust reflection or lens distortion. The unified data stream after cleaning, alignment, and structured packaging constitutes the multimodal sensing data.

[0027] It should be noted that by aligning and cleaning multi-source heterogeneous data in a spatiotemporal manner, noise and redundancy in the underlying hardware acquisition are effectively eliminated. This reduces the bandwidth usage and provides a data foundation for building high-precision, spatiotemporally continuous digital twin scenarios.

[0028] Step 120: Input the multimodal sensing data into the neural radiation field model, reconstruct the previous physical field twin model, and construct the current physical field twin model of the computer room.

[0029] Here, the Neural Radiance Field (NeRF) model is a continuous modeling network that implicitly represents the density and color of complex 3D scenes through deep learning.

[0030] Furthermore, the previous physical twin model here refers to the historical four-dimensional digital twin state of the data center before the current time window, reflecting the physical field distribution of the data center at the previous moment. Therefore, the current physical twin model is obtained by dynamically updating over time after integrating the latest multimodal sensing data. It not only includes spatial geometric coordinates and color density, but also deeply integrates physical fields such as temperature, airflow, and power consumption into the latest four-dimensional data center scene model.

[0031] Specifically, the integrated multimodal sensing data stream can be continuously input into the training pipeline of the cloud-based incremental neural radiation field model using a sliding window mechanism. This allows for rapid optimization and iteration of the local parameters of the previous physical field twin model based on newly acquired images and point clouds. After training is complete, the feature mapping in the neural radiation field model is immediately refreshed, thereby constructing a current physical field twin model that is synchronized with the real physical world in real time.

[0032] It should be noted that, here, based on real-time multimodal sensing data and neural radiation field model, the previous physical field twin model is reconstructed to obtain the current physical field twin model of the computer room. This achieves continuous and refined physical field twin model construction, ensuring that the digital space always maintains a high degree of consistency with the dynamic environment of the physical computer room.

[0033] Step 130: In response to the detected data center operation and maintenance event, extract the target tensor data of the target area corresponding to the data center operation and maintenance event from the current physical field twin model, integrate the target tensor data, and generate a physical tensor summary of the target area.

[0034] Here, data center operation and maintenance events cover various dynamic changes that require intervention or attention during the daily operation of the data center, such as equipment migration detected by the monitoring platform, abnormal energy consumption alarms, temperature hotspots, or energy-saving optimization requests initiated by operation and maintenance personnel.

[0035] Furthermore, the target area here refers to the specific location or cluster of server racks in the actual three-dimensional space of the data center where the aforementioned operational events occurred. Here, target tensor data refers to the set of numerical values ​​derived from the current physical twin model, used to characterize the high-dimensional physical properties of the target space at the latest time point, such as the temperature spatial distribution matrix, airflow vector field, and power consumption three-dimensional tensor. Here, physical tensor summaries are text or lightweight structured data that can be directly read and understood by natural language models after dimensionality reduction and feature extraction of these high-dimensional, obscure tensor data.

[0036] Specifically, upon detecting a trigger signal for a data center maintenance event, the system immediately locates the target area based on the alarm coordinates and issues a query command to the continuous twin model to retrieve the latest target tensor data for that area. Subsequently, this high-dimensional physical data undergoes dimensionality reduction projection and semantic transformation to generate a concise physical tensor summary. This successfully breaks down the barrier between high-dimensional complex physical field data and the underlying language reasoning model. Through dimensionality reduction and abstraction, not only is the memory usage and computational overhead of subsequent algorithms greatly reduced, but it also provides accurate physical status input for intelligent diagnosis.

[0037] Step 140: Retrieve decision reference information associated with the target area, input the decision reference information and the physical tensor summary into the large language model, and obtain the data center operation and maintenance decision output by the large language model.

[0038] Here, decision reference information refers to the structured prior rules and historical operational background extracted from the data center's professional knowledge base, including the data center's overall energy allocation and security warning thresholds, the specific equipment topology connections within the target area, and historical maintenance work orders. Furthermore, the data center operation and maintenance decision here is the final inference output action guide for responding to the aforementioned operation and maintenance events, typically including the target variables to be adjusted, specific action instruction sequences, execution timing, and mandatory security constraints.

[0039] Specifically, firstly, a retrieval interface can be invoked based on the device identifier and event type to accurately match and retrieve the corresponding decision reference information from the knowledge base. Next, this prior decision reference information is contextually concatenated with the real-time physical tensor summary generated in the preceding steps and embedded into a preset prompt template.

[0040] Finally, the constructed prompt text is input into a large language model with retrieval and enhanced generation capabilities. The large language model performs deep reasoning within a context that fully integrates the current physical status and historical operation and maintenance rules, generating data center operation and maintenance decisions that include specific control strategies.

[0041] It should be noted that by deeply integrating real-time 3D physical situation and static expert knowledge base with large language model, the automated generation of structured control strategies with constraints is realized, which gets rid of the heavy reliance on traditional manual experience and greatly enhances the speed of computer room fault response and the scientific nature of decision-making.

[0042] The method provided in this invention drives the training of a neural radiation field model through multimodal data streams, constructs a spatiotemporally continuously updated current physical field twin model, and when a data center operation and maintenance event occurs, reduces the high-dimensional target tensor data extracted from the current physical field twin model into a physical tensor summary, supplemented by retrieved decision reference information, and inputs it into a large language model to generate automated data center operation and maintenance decisions. This forms a full-link adaptive closed-loop mechanism that integrates continuous scene reconstruction and knowledge retrieval generation, breaking through the dual bottlenecks of existing 3D GIS relying on static modeling and the difficulty of large language models in handling high-dimensional physical field data. It can accurately reflect the heat distribution and environmental evolution of the data center, automatically output optimized strategies that have been verified by safety, significantly reduce the intensity of manual inspections, shorten the location time of major faults to the minute level, and deeply release the space for data center energy efficiency optimization.

[0043] Based on any of the above embodiments, after step 140, the following steps are included: Determine and execute the target action sequence corresponding to the data center operation and maintenance decision; Collect feedback status data of the computer room after the target action sequence is completed; Obtain the predicted state data generated based on the current physical field twin model for the target action sequence; Calculate the state deviation between the feedback state data and the predicted state data; If the state deviation exceeds a preset state deviation threshold, a reconstruction operation for the target region is triggered to obtain an optimized current physical field twin model. Based on the state deviation, the large language model is fine-tuned to obtain an optimized large language model.

[0044] Specifically, first, the target action sequence corresponding to the data center operation and maintenance decision is determined and executed. Here, the target action sequence refers to the set of instructions parsed from the data center operation and maintenance decision that can be directly recognized and executed by the underlying physical equipment of the data center, such as the temperature reduction instruction for a specific air conditioner, the power limiting instruction for the power distribution unit, etc.

[0045] In practice, based on the data center operation and maintenance decisions output by the large language model, specific action sequences for the target variables can be extracted and mapped into standard industrial protocol instructions. These instructions are then sent to subsystems such as air conditioning, UPS (Uninterruptible Power Supply), and monitoring via the control bus for automated execution.

[0046] Then, the feedback status data of the data center after executing the target action sequence is collected. Here, the feedback status data refers to the real physical readings reported in real time by environmental sensors after the physical data center receives and executes the control commands, such as the measured cold aisle temperature and the actual current of the equipment. In addition, the predicted status data here refers to the expected change results given by the current physical field twin model after performing physical field simulation calculations in virtual space based on the same action sequence input.

[0047] In practice, on the one hand, feedback information from actuators and sensors can be collected in real time; on the other hand, a simulation request is simultaneously sent to the NeRF (Neural Radiance Field) continuous modeling module to obtain its predicted value for this regulation within the current physical twin model. Understandably, this step simultaneously obtains the response results from the real physical world and the prediction results from the digital twin space, providing a rigorous data comparison basis for subsequent evaluation of model accuracy and strategy effectiveness.

[0048] Next, the state deviation between the feedback state data and the predicted state data is calculated. If the state deviation exceeds a preset state deviation threshold, a reconstruction operation of the target region is triggered to obtain an optimized current physical field twin model. Based on the state deviation, the large language model is fine-tuned to obtain an optimized large language model.

[0049] Here, state bias is a numerical indicator that quantifies the prediction accuracy of the digital twin model and the effectiveness of policy execution. The preset state bias threshold is the tolerance boundary for determining whether the model or policy has failed.

[0050] Furthermore, the reconstruction operation here specifically refers to a rapid neural radiation field relearning process performed only on localized spatial hotspot regions that have undergone changes. It is understandable that the optimized current physical field twin model and the optimized large language model are evolved versions that have incorporated the latest operational rules after bias correction.

[0051] In practice, the difference can be calculated by automatically comparing the measured sensor readings with the predicted values ​​of the twin model. Once the difference exceeds a pre-set threshold, it indicates a significant disconnect between the current digital world and the physical data center. At this point, a local reconstruction process is immediately invoked, extracting the latest multimodal data only in the local space where the alarm is located to update the model weights. Simultaneously, the execution effect, including state deviations, is recorded as a tag in the policy database, triggering incremental fine-tuning of the large language model to absorb new operational data and experience.

[0052] It should be noted that by using residual strategy learning to perform closed-loop correction of the deviation between measured and predicted values, the computational power wasted on relearning the entire model was avoided, and model self-correction at the minute level was achieved.

[0053] The method provided in this invention collects actual feedback data and predicted data from a twin model after the execution of data center operation and maintenance decisions. By comparing the state deviations of the two, when the deviation exceeds the limit, it adaptively triggers the rapid reconstruction of the local NeRF model and the incremental fine-tuning of the large language model, forming a truly minute-level self-correction and continuous optimization adaptive closed loop. This continuously eliminates the difference between the model and the actual measurement, enabling the data center's energy efficiency optimization and fault response capabilities to continuously evolve and improve after multiple iterations.

[0054] Based on any of the above embodiments, determining and executing the target action sequence corresponding to the data center operation and maintenance decision includes: Extract the initial operation parameters corresponding to the data center operation and maintenance decision, and determine whether the initial operation parameters meet the preset data center security boundary specifications; Detect the execution timing relationship and resource mutual exclusion relationship among different operation parameters in the initial operation parameters; The target action sequence is obtained based on the initial operation parameters that meet the data center security boundary specifications and do not have execution timing and resource conflicts.

[0055] Specifically, firstly, the initial operating parameters corresponding to the data center operation and maintenance decisions are extracted, and it is determined whether the initial operating parameters meet the preset data center safety boundary specifications. Here, initial operating parameters refer to the original control variables that have just been generated by the large language model and have not yet undergone rigorous physical feasibility review, such as the target absolute value of temperature and the expected voltage adjustment range. Furthermore, the data center safety boundary specifications here are hard protection rules formulated based on the data center's physical hardware constraints and industry safety standards, covering hard threshold limits for parameters such as temperature, current, and voltage.

[0056] In practice, the built-in symbolic security checker can be used to receive the initial operating parameters corresponding to the data center operation and maintenance decisions output by the large language model. First, all initial operating parameters undergo formal analysis for limit value verification to ensure that the temperature control or current distribution parameters generated by the large language model strictly fall within the safe operating range of the equipment, avoiding the risk of overloading. Understandably, based on this step, the output configuration rules of the large language model, which have a certain degree of illusionary risk, are checked, preventing the issuance of dangerous commands at the source and ensuring the security of the underlying hardware of the data center.

[0057] Next, the execution timing relationships and resource mutual exclusion relationships among different operation parameters in the initial operation parameters are detected. Based on the initial operation parameters that meet the data center security boundary specifications and do not have execution timing and resource conflicts, the target action sequence is obtained.

[0058] Here, execution sequence refers to the logical order in which multiple control actions must be followed, such as turning on the backup cooling equipment before turning off the main cooling equipment. Resource mutual exclusion refers to conflicting control commands issued for the same physical resource within the same time period.

[0059] In practice, after passing the boundary threshold verification, a conflict resolution algorithm is further used to examine the action combinations in the strategy. The logical coherence and resource preemption between multi-step operations are cross-checked, eliminating redundant / erroneous parameters that are logically mutually exclusive or have reversed timing. After the above formal conflict analysis and limit value dual verification, the final legal and compliant parameter set is translated and assembled to form the final deployable target action sequence.

[0060] It should be noted that by deeply analyzing the timing of actions and mutual exclusion of resources, the smoothness and stability of the multi-device collaborative control process are ensured, avoiding system oscillations or command deadlocks caused by complex strategies in actual implementation.

[0061] The method provided in this invention introduces boundary specification verification and timing and mutual exclusion conflict resolution based on a symbolic security checker before executing the operation and maintenance decisions generated by the large language model. Only operation parameters that pass the rigorous verification are allowed to form the final action sequence. This makes up for the logical loopholes that may exist in generative artificial intelligence in the field of industrial control, provides financial-grade and industrial-grade security and compliance protection, ensures the high availability of the data center during the fully automatic unattended adjustment process, and avoids the occurrence of data center operation accidents.

[0062] Based on any of the above embodiments, the target tensor data includes the three-dimensional temperature tensor, airflow vector field tensor, and power consumption tensor of the target region at the current moment; In step 130, the target tensor data is integrated to generate a physical tensor summary of the target region, including: Spatial dimension reduction projection is performed on the three-dimensional temperature tensor, the airflow vector field tensor, and the power consumption tensor to extract the environmental features and equipment load features of the target area; Based on the environmental characteristics and the load characteristics, the physical tensor summary is constructed.

[0063] Specifically, firstly, spatial dimension reduction projection is performed on the three-dimensional temperature tensor, airflow vector field tensor, and power consumption tensor to extract the environmental features and equipment load features of the target area.

[0064] Here, the target tensor data specifically covers three core physical dimensions. The three-dimensional temperature tensor is used to characterize the real-time temperature distribution at various points in the target area. The airflow vector field tensor is used to reflect the flow direction and wind speed of the cooling airflow in three-dimensional space. The power consumption tensor represents the real-time energy consumption distribution of each cabinet or IT device in the area.

[0065] Furthermore, spatial dimensionality reduction projection here refers to the process of using data processing algorithms to map and compress the aforementioned complex three-dimensional high-dimensional matrix or tensor data into low-dimensional feature vectors or two-dimensional feature maps. Here, environmental features and equipment load features are the core semantic information extracted after dimensionality reduction; the former includes, for example, the average temperature of the hot and cold aisles and local hot spot extremes, while the latter includes, for example, the real-time power ratio of a specific cabinet.

[0066] In the specific implementation, the three-dimensional temperature tensor, the airflow vector field tensor, and the power consumption tensor of the target area at the latest time point are directionally derived from the current physical field twin model. Then, a dimensionality reduction algorithm is used to perform spatial dimensionality reduction projection on these high-dimensional data, remove redundant spatial background noise, and accurately extract the environmental features and equipment load features that can reflect the current situation of the target area. This effectively reduces the computational complexity and storage overhead of high-dimensional physical field data, while accurately preserving the key spatial distribution information that characterizes the operating status of the computer room.

[0067] Next, a physical tensor summary is constructed based on environmental and load features. Specifically, after successfully extracting environmental and device load features, these feature values ​​are concatenated and transformed into descriptive text containing clear contextual meaning according to preset semantic transformation rules and prompt word templates. This constructs the physical tensor summary, which can then be directly embedded into prompt templates for use by large language models. This successfully transforms the obscure high-dimensional physical tensor across modalities into structured prompt input that large language models can directly understand, achieving a seamless connection between the physical world state and the semantic reasoning space.

[0068] The method provided in this invention clarifies the specific three-dimensional multiphysics attributes contained in the target tensor data, and accurately extracts environmental and load features through spatial dimensionality reduction projection technology. Then, it constructs a physical tensor summary specifically for large language model inference across modalities, solving the industry problem that large language models have difficulty directly processing and understanding high-dimensional continuous physical field data. Under the premise of ensuring that key thermodynamic and energy consumption information in three-dimensional space is not lost, it greatly improves the data throughput efficiency of the model inference process and the scientific accuracy of the final operation and maintenance decision.

[0069] Based on any of the above embodiments, step 120 includes: The sensor sequence data is extracted from the multimodal sensor data by a preset sliding window, and the sensor sequence data is input into the neural radiation field model. The current spatial density and current color parameters of the previous physical field twin model are updated based on incremental training. The high-resolution network parameters in the previous physical twin model are converted into a sparse voxel mesh; Based on the sparse voxel mesh, the current spatial density, and the current color parameters, a lookup table index for spatial location and rendering attributes is constructed, and the current physical field twin model is reconstructed.

[0070] Specifically, firstly, perceptual sequence data is extracted from the multimodal perceptual data using a preset sliding window. This perceptual sequence data is then input into the neural radiation field model, and the current spatial density and color parameters of the previous physical field twin model are updated based on incremental training. Here, the preset sliding window refers to a dynamically advancing time interval set on the time axis, used to capture the latest continuous data segment. The perceptual sequence data refers to the set of image frames and point clouds, etc., arranged chronologically within this time interval. Furthermore, the incremental training here refers to a machine learning method that fine-tunes and optimizes the model using only newly input data based on the existing parameters of the previous physical field twin model, rather than retraining from scratch. Here, the current spatial density and current color parameters are the core physical quantities output by the neural radiation field model, representing the geometric shape and surface color of each point in the three-dimensional space of the computer room, respectively.

[0071] In the specific implementation, the latest arriving perception sequence data is continuously extracted from the integrated multimodal perception data stream according to a preset sliding window. The consistency of image and point cloud rendering within the window is checked, and segments with drastic lighting changes or large occlusions are removed. It should be noted that the combination of sliding window and incremental training avoids the enormous computational overhead of relearning from the entire dataset, ensuring the timeliness of model updates and significantly reducing computational resource consumption. Subsequently, valid frames are fed into the incremental training task, enabling rapid optimization of the current spatial density and color parameters in 3D space based on the previous physics twin model.

[0072] Next, the high-resolution network parameters in the previous physics twin model are converted into a sparse voxel mesh. Here, the high-resolution network parameters refer to the massive multilayer perceptron weight matrix used to characterize the dense computer room scene within the neural radiation field model. The sparse voxel mesh, on the other hand, is a discretized three-dimensional spatial data structure that divides continuous space into numerous tiny voxels, retaining only those containing valid object information and discarding blank regions.

[0073] In practice, after incremental training is completed, voxel distillation can be performed immediately to extract and compress scene information that was originally implicitly stored in dense, high-resolution network parameters into an explicit sparse voxel mesh. It is understandable that voxel distillation successfully decouples the computational process of the complex neural network from the final 3D representation, greatly reducing the model's memory usage during runtime.

[0074] Then, based on the sparse voxel mesh, the current spatial density, and the current color parameters, a lookup table index for spatial location and rendering attributes is constructed to reconstruct the physical twin model. Here, the lookup table index refers to a fast retrieval data structure that directly maps three-dimensional spatial coordinates to corresponding physical attributes, such as density, color, and temperature.

[0075] In the specific implementation, after generating the sparse voxel mesh, the updated spatial density and color parameters, among other rendering attributes, are bound to the corresponding voxel nodes, thus constructing a lookup table index that maps spatial location to rendering attributes. This method achieves a lightweight reconstruction of the entire data center digital scene, ultimately reconstructing the current physical twin model. It should be noted that the complex rendering process, which originally required multiple neural network forward propagations, has been simplified to a direct memory read operation. This reduces subsequent inference query latency from seconds to milliseconds, meeting the extremely low latency requirements of real-time data center operations.

[0076] The method provided in this invention uses a preset sliding window to extract data for incremental training of a neural radiation field model, and uses voxel distillation technology to convert high-resolution network parameters into sparse voxel grids and construct lookup table indexes. This cleverly balances the extremely high geometric accuracy and rapid response capability of the digital twin model, achieving high-concurrency query speed in milliseconds while completing local reconstruction in minutes, ensuring that the twin model always maintains efficient synchronization with the real data center status.

[0077] Based on any of the above embodiments, converting the high-resolution network parameters in the previous physics twin model into a sparse voxel mesh includes: Obtain the spatial location of the equipment in the computer room; Based on the spatial location of the equipment, the computer room space is divided into an adjacent area and a background area. The sparse voxel mesh is obtained by sampling the high-resolution network parameters according to the first sampling density corresponding to the device's adjacent area and the second sampling density corresponding to the device's background area. The first sampling density is higher than the second sampling density.

[0078] Specifically, the first step is to obtain the spatial location of the equipment within the data center. Here, the spatial location refers to the specific geometric coordinate range of core IT and cooling assets such as server racks, servers, air conditioning equipment, and power distribution units within the data center's three-dimensional coordinate system. This is achieved by first analyzing the data center's topology and physical inventory to determine the spatial location of each type of equipment.

[0079] Next, based on the spatial location of the equipment, the computer room space is divided into the equipment proximity area and the equipment background area. Here, the equipment proximity area refers to the three-dimensional space within a certain distance of the surface of these core devices, which contains complex geometric details and drastic temperature and airflow changes. On the other hand, the equipment background area refers to areas far away from these devices, such as ceilings, open corridors, or plain white walls, which lack details and are relatively static.

[0080] In the specific implementation, using the spatial location of the equipment as a reference point or bounding box, a spatial partitioning algorithm is used to accurately segment the continuous three-dimensional space of the entire computer room into the adjacent area of ​​the equipment requiring primary attention and the background area of ​​the secondary equipment. It should be noted that by introducing a spatial attention mechanism based on the importance of physical assets, a precise spatial partitioning basis is provided for the subsequent efficient differential model compression.

[0081] Next, the high-resolution network parameters are sampled according to the first sampling density corresponding to the device's adjacent area and the second sampling density corresponding to the device's background area to obtain a sparse voxel mesh. In this step, the first sampling density refers to the higher resolution used when dividing the voxels to capture rich physical details. The second sampling density refers to the lower resolution used to cover large areas of blank or featureless regions.

[0082] In practice, a non-uniform spatial sampling strategy is employed during the voxel distillation process of the high-resolution network parameters. For the device neighborhood containing critical IT assets, an extremely high first sampling density is used for refined feature extraction; while for the open device background area, a lower second sampling density is used for coarse feature aggregation, thereby generating a sparse voxel mesh with varying density. It should be noted that through an adaptive spatial downsampling strategy, while maintaining centimeter-level modeling accuracy for critical equipment in the data center, redundant background data is eliminated to the maximum extent, further compressing the size of the twin model.

[0083] The method provided in this invention differentiates the three-dimensional space based on the spatial location of the physical equipment in the data center, and uses a dual sampling density of near-dense and far-sparse to perform voxel distillation on high-resolution network parameters. This achieves precise allocation of computing power and storage resources to high-value areas of the data center. Without sacrificing the perception accuracy of key physical characteristics such as temperature and airflow of core equipment, it achieves extreme compression of the volume of the physical field twin model, significantly improving resource utilization.

[0084] Based on any of the above embodiments, the decision reference information associated with the target area includes at least one of the following: topological connections associated with the target area, historical operation and maintenance records, energy quotas, and security thresholds. The topology connection includes at least one of spatial location topology, power supply cascade topology, and cooling coverage topology.

[0085] Specifically, topology connections reflect the physical and logical dependencies between components within the data center. Spatial location topology records the three-dimensional adjacency relationships between racks, rows, and rooms. Power supply cascading topology maps the power distribution links from mains power and UPS to specific servers. Cooling coverage topology defines the correspondence between the air conditioning vent direction and the covered hotspot areas.

[0086] Here, the historical maintenance log includes records of similar alarms that occurred in the target area in the past, along with manual handling experience. Here, the energy quota refers to the maximum allowable power consumption allocated to this area or rack. Here, the safety threshold refers to the minimum standards that must be adhered to, such as temperature, voltage, and current.

[0087] In practice, when prompt words need to be generated, the structured interface can be called to accurately retrieve the above-mentioned related information from the data center knowledge base based on the device identifier of the target area, the current time window, and the triggered alarm information.

[0088] It should be noted that by introducing multi-dimensional physical topology, historical experience, and hard constraints, strict professional boundaries are set for the open generation of large language models, completely eliminating the illusionary risk of generation strategies deviating from physical reality.

[0089] The method provided in this invention clarifies that the decision reference information required for reasoning by the large language model covers multi-dimensional topological connections, historical operation and maintenance records, energy quotas, and security thresholds. This endows the large language model with in-depth data center domain expert knowledge, enabling the generated operation and maintenance decisions to not only possess intelligent reasoning capabilities but also strictly conform to the complex power supply and cooling physical architecture of the data center, significantly improving the industrial-grade feasibility of the strategy.

[0090] Based on any of the above embodiments, step 110 includes: Acquire initial multi-source sensor data for the computer room; Perform a rendering consistency check on the image data and point cloud data in the initial multi-source sensor data to obtain the rendering consistency check result; Based on the rendering consistency test results, the initial multi-source sensor data is filtered to obtain the multimodal sensing data.

[0091] Specifically, the first step is to acquire initial multi-source sensor data from the data center. In this step, initial multi-source sensor data refers to the raw signal stream directly collected and reported by the edge hardware of the data center without any filtering or deep cleaning. This includes two-dimensional raw frames captured by high-resolution cameras, three-dimensional raw point arrays scanned by LiDAR, and misaligned measurement values ​​output by various types of environmental sensors.

[0092] In practice, the system can continuously receive initial multi-source sensor data synchronously reported by various sensing devices deployed in the computer room at a unified clock frequency, and use the built-in time alignment algorithm to synchronize these multi-channel data with millisecond-level timestamps to obtain the initial multi-source sensor data.

[0093] Then, a rendering consistency check is performed on the image data and point cloud data from the initial multi-source sensor data to obtain the rendering consistency check result. Here, the rendering consistency check can be used to evaluate whether the image features and point cloud spatial structure at the same moment match when projected onto the same viewpoint. Here, the rendering consistency check result is a score or Boolean judgment value that quantifies this degree of spatial and optical matching, used to identify whether there are anomalies in the data segment.

[0094] In practice, before 3D reconstruction, a geometric projection matrix is ​​used to perform cross-modal alignment of image data and point cloud data, and a volume rendering consistency check is performed. The overlap deviation between the two is strictly calculated, thereby outputting a rendering consistency check result that determines whether the current frame is affected by drastic lighting changes or large-area occlusion. It should be noted that through cross-modal consistency verification, dirty data caused by lens distortion, dust reflection, or human occlusion is accurately identified and blocked, ensuring the purity of the data sent to the training pipeline.

[0095] Finally, based on the rendering consistency check results, the initial multi-source sensor data is filtered to obtain multimodal perception data. For example, based on the rendering consistency check results obtained in the previous steps, keyframe discrimination logic can be configured to directly remove invalid segments that fail the check, contain drastic lighting changes, or are occluded, retaining only valid frames with large view coverage, significant changes, and highly consistent multimodal features. The aligned stream packaged and output after this rigorous filtering mechanism constitutes the final multimodal perception data.

[0096] It should be noted that by intelligently filtering out redundant and low-quality frames, the bandwidth consumption of invalid data in the cloud is significantly reduced, and the divergence or artifacts of the neural radiation field model during incremental training are fundamentally avoided.

[0097] The method provided in this invention introduces a rendering consistency check between images and point clouds during the data access stage, and performs strict filtering and screening on the initial multi-source sensor data based on the check results, outputting pure multimodal perception data, thus constructing a robust data quality firewall. This not only saves valuable transmission bandwidth and computing power, but also ensures efficient convergence and extreme accuracy of the downstream NeRF modeling process, guaranteeing the reliability of the data center adaptive operation and maintenance system.

[0098] Based on any of the above embodiments Figure 2 This is the second flowchart illustrating the data center operation and maintenance method provided by this invention, as shown below. Figure 2 As shown, the method includes: First, in the multimodal data acquisition and preprocessing stage, cameras, LiDAR, and various sensors deployed at the edge constitute the multimodal acquisition entry point, continuously acquiring raw environmental and equipment status data of the data center. Subsequently, the acquired raw data enters the time-series alignment and noise reduction processing stage. During this process, outliers and redundant information are removed through multi-frame filtering. The processed data further passes through a unified data stream output module, where keyframe filtering and compression are performed to form a standardized and spatiotemporally aligned data stream for uploading to the cloud.

[0099] After entering the cloud, the incremental training NeRF Twin module is initiated to perform continuous twin modeling, generating a four-dimensional data center model containing multiple physical fields in real time. When local changes occur in the data center, local hotspot area updates are triggered, and a lookup table index is built through voxel distillation to support millisecond-level queries. Based on the updated local model, tensor summarization is performed to extract physical features such as temperature, airflow, and load; simultaneously, a knowledge base retrieval is conducted to extract relevant topological relationships and historical operation and maintenance information, thus obtaining decision-making reference information.

[0100] During the decision-making and reasoning phase, the generated physical tensor summary is fused with retrieved knowledge base information, and contextual hints are concatenated to generate input for the large language model. After receiving the input, the large model performs reasoning to generate multiple candidate control strategies. Subsequently, these candidate strategies enter the risk scoring and screening stage and undergo symbolic verification to ensure they do not violate the data center's hard security thresholds and resource mutual exclusion constraints, ultimately selecting the optimal security strategy.

[0101] During the execution and closed-loop optimization phase, the validated strategies are translated into industry protocol instructions and distributed by the industry protocol execution module to the BMS (Building Management System) or DCIM (Data Center Infrastructure Management) system for execution. During execution, continuous monitoring of the execution status and deviation detection are performed, collecting sensor feedback in real time and calculating deviations based on the twin model's prediction results. If the deviation exceeds a preset threshold, automatic local reconstruction and language model fine-tuning are triggered, forming a minute-level self-correcting closed loop to continuously optimize the data center's operation and maintenance decisions and control effectiveness.

[0102] The method provided in this invention generates a unified spatiotemporally aligned data stream through distributed cameras, LiDAR, and environmental sensors in the data center. It employs incremental neural radiation field training to construct a dynamically updated four-dimensional data center model over time, and compresses the model using voxel distillation to meet real-time query requirements. The system extracts physical summaries such as temperature, airflow, and load within a specified spatial range, and combines this with topology and historical maintenance records retrieved from the knowledge base to construct a large language model for enhanced retrieval prompts, enabling it to output structured control strategies. After symbolic security verification, the strategies are sent to the power and monitoring subsystem for execution. When the execution result deviates significantly from the model prediction, local model reconstruction and language model fine-tuning are immediately triggered, forming a minute-level self-correction closed loop, thereby achieving automated operation and maintenance for energy efficiency optimization, fault location, and risk prevention.

[0103] In one embodiment, the operation and maintenance process of the data center includes the following stages: Phase 1: Data Acquisition and Spatiotemporal Fusion Process. Specifically, firstly, cameras and LiDAR at the edge of the data center continuously output image frames and point clouds according to a unified clock, while environmental sensors simultaneously report raw data such as temperature, humidity, and energy consumption. The acquisition gateway uses a built-in time-series alignment algorithm to synchronize multiple data streams at millisecond levels and calls a denoising model to clean the images and point clouds, eliminating anomalies caused by dust reflections and lens distortion. To avoid redundant bandwidth consumption, the gateway uses keyframe discrimination logic to retain only frames with large viewing angle coverage and significant changes, while the remaining frames are compressed and cached locally.

[0104] Subsequently, the filtered data is packaged into a unified spatiotemporally aligned stream and uploaded to the cloud-based stream processing cluster. The cluster performs another integrity check on the arriving data and supplements metadata such as sensor identifiers, rack numbers, and camera poses. At this point, the multimodal data has completed cleaning, synchronization, and structuring, and then enters the continuous modeling phase.

[0105] Phase Two: Continuous Digital Twin Modeling and Maintenance Process. In detail, the cloud-based continuous modeling service extracts the latest sequence from the integrated data stream using a sliding window, updating the neural radiation field model in real time. The system first checks the consistency of image and point cloud volume rendering within the window, removing segments with drastic lighting changes or large occlusions. Then, the valid frames are fed into an incremental training task to achieve rapid optimization of spatial density and color. After training, the service immediately triggers voxel distillation, compressing the high-resolution network into a sparse voxel mesh and building a lookup table index, reducing inference query latency from seconds to milliseconds.

[0106] When localized changes occur in the data center, such as equipment relocation or a sudden temperature rise, the model maintenance sub-thread locates hotspot areas based on real-time sensor alarms and performs rapid reconstruction only in those localized areas, eliminating the need for overall relearning. This approach reduces the model update cycle from hours to minutes, ensuring that the digital twin always maintains a high degree of consistency with the physical data center.

[0107] Phase Three: Enhanced Search Prompt Generation and Strategy Decision-Making Process. Specifically, when the monitoring platform detects abnormal energy consumption, alarm events, or optimization requests initiated by maintenance personnel, the decision service first invokes the structured search interface to extract information related to the target area from the knowledge base, such as topological connections, historical maintenance orders, energy quotas, and safety thresholds. Simultaneously, the model summarization interface requests the temperature distribution, airflow vector, and power consumption tensor of the target area from the continuous twin model at the latest time point, and then reduces and integrates this high-dimensional data into a readable summary text.

[0108] The decision service concatenates the two types of context into the prompt template and then passes it to the large language model for reasoning. The language model provides multiple feasible solutions based on the context. The server-side custom scorer combines energy consumption prediction, risk coefficient, and operational complexity to rank the candidate solutions and finally selects the strategy with the highest comprehensive score. The result is returned in JSON format, including the target variable, action sequence, expected threshold, and execution window.

[0109] Phase Four: Policy Verification and Execution Process. In detail, after receiving the policy, the verification engine first performs boundary checks on all actions according to the data center safety regulations to ensure that no temperature exceedances or current overloads occur. Then, it uses a conflict resolution algorithm to check the timing and resource mutual exclusion relationships between different actions. If the verification passes, the engine translates the policy into industrial protocol instructions and sends them to devices such as air conditioners, UPS, and power distribution units via the control bus.

[0110] During execution, the instruction executor collects feedback information in real time and reports it to the cloud for monitoring. The platform displays temperature, power consumption, and device status changes in a visual interface using color blocks or curves, allowing maintenance personnel to intuitively monitor the execution progress and intervene manually when necessary.

[0111] Phase 5: Adaptive Closed-Loop Optimization Process. In detail, after the strategy execution is complete, the closed-loop engine automatically compares the predictions of the continuous twin model with the actual sensor readings and calculates the difference. If the difference exceeds a pre-set threshold, it indicates a deviation in the model or strategy. The system will immediately invoke the local reconstruction process to update the model and simultaneously trigger incremental fine-tuning of the language model to absorb new operational data and maintenance experience.

[0112] Next, the closed-loop engine tags historical records of the same scenario in the policy database based on the collected execution effects, which are then used by subsequent decision-making services during retrieval. This allows the system to accumulate a more accurate action-effect mapping after multiple iterations. Through this continuous learning and self-correction mechanism, the data center's energy efficiency and fault response capabilities will continuously improve, while the frequency of manual intervention will decrease, forming a truly self-evolving operation and maintenance system.

[0113] It should be noted that this invention forms a closed-loop adaptive operation and maintenance system by introducing continuous digital twins and large-scale model collaborative reasoning in a data center scenario. Compared with existing methods that rely on static 3D models and expert rules, this solution has the following significant advantages: First, it balances the accuracy and real-time performance of digital twins; that is, incremental training supplemented by voxel distillation can complete local reconstruction in minutes, maintaining centimeter-level geometric accuracy and millisecond-level query speed, ensuring that the model is always synchronized with the actual data center status. Second, it automates and interprets operation and maintenance decisions; that is, the retrieval-enhanced large language model integrates physical summaries and historical knowledge bases, which can generate constrained structured control strategies; the strategies are automatically labeled after symbolic verification and comparison with actual execution results, providing traceable evidence for subsequent auditing and continuous optimization. Third, it improves both energy efficiency and reliability; that is, the adaptive closed loop continuously eliminates the difference between the model and actual measurements, dynamically optimizes cooling and power distribution parameters, and the measured energy saving rate can reach 5% to 15%, shortening the time for locating major faults from hours to minutes. Fourth, security and privacy are guaranteed; that is, end-to-end encryption and hierarchical access control are adopted throughout the entire chain, and policy and model updates are written to hash logs; in multi-campus deployment scenarios, federated fine-tuning is supported, and only gradient statistics are uploaded, effectively preventing the leakage of sensitive data. Fifth, it is easy to expand and deploy; that is, the system is based on a microservice and service mesh architecture, which can be deployed independently in a single data center or upgraded across clusters, and is compatible with mainstream air conditioning, power supply and monitoring interfaces, reducing transformation costs.

[0114] Based on any of the above embodiments Figure 3 This is a structural schematic diagram of the data center operation and maintenance device provided by the present invention, as shown below. Figure 3 As shown, the device includes: Acquisition unit 310 acquires multimodal sensing data from the computer room; The reconstruction unit 320 inputs the multimodal sensing data into the neural radiation field model to reconstruct the previous physical field twin model and construct the current physical field twin model of the computer room. Tensor summary generation unit 330, in response to a detected data center operation and maintenance event, extracts target tensor data of the target region corresponding to the data center operation and maintenance event from the current physical field twin model, integrates the target tensor data, and generates a physical tensor summary of the target region. The decision unit 340 retrieves decision reference information associated with the target area, inputs the decision reference information and the physical tensor summary into the large language model, and obtains the data center operation and maintenance decision output by the large language model.

[0115] The device provided in this invention drives the training of a neural radiation field model through multimodal data streams, constructs a spatiotemporally continuously updated current physical field twin model, and when a data center operation and maintenance event occurs, reduces the high-dimensional target tensor data extracted from the current physical field twin model into a physical tensor summary, supplemented by retrieved decision reference information, and inputs it into a large language model to generate automated data center operation and maintenance decisions. This forms a full-link adaptive closed-loop mechanism that integrates continuous scene reconstruction and knowledge retrieval generation, breaking through the dual bottlenecks of existing 3D GIS relying on static modeling and the difficulty of large language models in handling high-dimensional physical field data. It can accurately reflect the heat distribution and environmental evolution of the data center, automatically output optimized strategies that have been verified by safety, significantly reduce the intensity of manual inspections, shorten the location time of major faults to the minute level, and deeply release the space for data center energy efficiency optimization.

[0116] Based on any of the above embodiments, an execution and feedback unit is included after the decision-making unit, and the execution and feedback unit is specifically used for: Determine and execute the target action sequence corresponding to the data center operation and maintenance decision; Collect feedback status data of the computer room after the target action sequence is completed; Obtain the predicted state data generated based on the current physical field twin model for the target action sequence; Calculate the state deviation between the feedback state data and the predicted state data; If the state deviation exceeds a preset state deviation threshold, a reconstruction operation for the target region is triggered to obtain an optimized current physical field twin model. Based on the state deviation, the large language model is fine-tuned to obtain an optimized large language model.

[0117] Based on any of the above embodiments, the execution and feedback unit is further specifically used for: Extract the initial operation parameters corresponding to the data center operation and maintenance decision, and determine whether the initial operation parameters meet the preset data center security boundary specifications; Detect the execution timing relationship and resource mutual exclusion relationship among different operation parameters in the initial operation parameters; The target action sequence is obtained based on the initial operation parameters that meet the data center security boundary specifications and do not have execution timing and resource conflicts.

[0118] Based on any of the above embodiments, the target tensor data includes the three-dimensional temperature tensor, airflow vector field tensor, and power consumption tensor of the target region at the current moment; The tensor summation generation unit is specifically used for: Spatial dimension reduction projection is performed on the three-dimensional temperature tensor, the airflow vector field tensor, and the power consumption tensor to extract the environmental features and equipment load features of the target area; Based on the environmental characteristics and the load characteristics, the physical tensor summary is constructed.

[0119] Based on any of the above embodiments, the reconstruction unit is specifically used for: The sensor sequence data is extracted from the multimodal sensor data by a preset sliding window, and the sensor sequence data is input into the neural radiation field model. The current spatial density and current color parameters of the previous physical field twin model are updated based on incremental training. The high-resolution network parameters in the previous physical twin model are converted into a sparse voxel mesh; Based on the sparse voxel mesh, the current spatial density, and the current color parameters, a lookup table index for spatial location and rendering attributes is constructed, and the current physical field twin model is reconstructed.

[0120] Based on any of the above embodiments, the reconstruction unit is further specifically used for: Obtain the spatial location of the equipment in the computer room; Based on the spatial location of the equipment, the computer room space is divided into an adjacent area and a background area. The sparse voxel mesh is obtained by sampling the high-resolution network parameters according to the first sampling density corresponding to the device's adjacent area and the second sampling density corresponding to the device's background area. The first sampling density is higher than the second sampling density.

[0121] Based on any of the above embodiments, the decision reference information associated with the target area includes at least one of the following: topological connections associated with the target area, historical operation and maintenance records, energy quotas, and security thresholds. The topology connection includes at least one of spatial location topology, power supply cascade topology, and cooling coverage topology.

[0122] Based on any of the above embodiments, the acquisition unit is specifically used for: Acquire initial multi-source sensor data for the computer room; Perform a rendering consistency check on the image data and point cloud data in the initial multi-source sensor data to obtain the rendering consistency check result; Based on the rendering consistency test results, the initial multi-source sensor data is filtered to obtain the multimodal sensing data.

[0123] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4As shown, the electronic device may include a processor 410, a communication interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communication interface 420, and the memory 430 communicate with each other through the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute a data center operation and maintenance method, which includes: acquiring multimodal sensing data of the data center; inputting the multimodal sensing data into a neural radiation field model to reconstruct the previous physical field twin model and construct the current physical field twin model of the data center; in response to a detected data center operation and maintenance event, extracting target tensor data of the target area corresponding to the data center operation and maintenance event from the current physical field twin model, integrating the target tensor data to generate a physical tensor summary of the target area; retrieving decision reference information associated with the target area, and inputting the decision reference information and the physical tensor summary into a large language model to obtain the data center operation and maintenance decision output by the large language model.

[0124] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0125] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the data center operation and maintenance method provided by the above methods. The method includes: acquiring multimodal sensing data of the data center; inputting the multimodal sensing data into a neural radiation field model to reconstruct a previous physical field twin model and construct a current physical field twin model of the data center; in response to a detected data center operation and maintenance event, extracting target tensor data of the target area corresponding to the data center operation and maintenance event from the current physical field twin model, integrating the target tensor data to generate a physical tensor summary of the target area; retrieving decision reference information associated with the target area, and inputting the decision reference information and the physical tensor summary into a large language model to obtain the data center operation and maintenance decision output by the large language model.

[0126] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the data center operation and maintenance method provided by the above methods. The method includes: acquiring multimodal sensing data of the data center; inputting the multimodal sensing data into a neural radiation field model to reconstruct a previous physical field twin model and construct a current physical field twin model of the data center; in response to a detected data center operation and maintenance event, extracting target tensor data of a target region corresponding to the data center operation and maintenance event from the current physical field twin model, integrating the target tensor data to generate a physical tensor summary of the target region; retrieving decision reference information associated with the target region, and inputting the decision reference information and the physical tensor summary into a large language model to obtain a data center operation and maintenance decision output by the large language model.

[0127] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0128] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for the operation and maintenance of a data center, characterized in that, include: Acquire multimodal sensing data from the data center; The multimodal sensing data is input into the neural radiation field model to reconstruct the previous physical field twin model and construct the current physical field twin model of the computer room. In response to a detected data center maintenance event, target tensor data of the target region corresponding to the data center maintenance event is extracted from the current physical field twin model, and the target tensor data is integrated to generate a physical tensor summary of the target region. Decision reference information associated with the target area is retrieved, and the decision reference information and the physical tensor summary are input into the large language model to obtain the data center operation and maintenance decision output by the large language model.

2. The data center operation and maintenance method according to claim 1, characterized in that, After obtaining the data center operation and maintenance decision output by the large language model, the method includes: Determine and execute the target action sequence corresponding to the data center operation and maintenance decision; Collect feedback status data of the computer room after the target action sequence is completed; Obtain the predicted state data generated based on the current physical field twin model for the target action sequence; Calculate the state deviation between the feedback state data and the predicted state data; If the state deviation exceeds a preset state deviation threshold, a reconstruction operation for the target region is triggered to obtain an optimized current physical field twin model. Based on the state deviation, the large language model is fine-tuned to obtain an optimized large language model.

3. The data center operation and maintenance method according to claim 2, characterized in that, The process of determining and executing the target action sequence corresponding to the data center operation and maintenance decision includes: Extract the initial operation parameters corresponding to the data center operation and maintenance decision, and determine whether the initial operation parameters meet the preset data center security boundary specifications; Detect the execution timing relationship and resource mutual exclusion relationship among different operation parameters in the initial operation parameters; The target action sequence is obtained based on the initial operation parameters that meet the data center security boundary specifications and do not have execution timing and resource conflicts.

4. The method for operating and maintaining a computer room according to any one of claims 1 to 3, characterized in that, The target tensor data includes the three-dimensional temperature tensor, airflow vector field tensor, and power consumption tensor of the target region at the current moment; The process of integrating the target tensor data to generate a physical tensor summary of the target region includes: Spatial dimension reduction projection is performed on the three-dimensional temperature tensor, the airflow vector field tensor, and the power consumption tensor to extract the environmental features and equipment load features of the target area; Based on the environmental characteristics and the load characteristics, the physical tensor summary is constructed.

5. The method for operating and maintaining a computer room according to any one of claims 1 to 3, characterized in that, The step of inputting the multimodal sensing data into the neural radiation field model to reconstruct the previous physical field twin model and construct the current physical field twin model of the computer room includes: The sensor sequence data is extracted from the multimodal sensor data by a preset sliding window, and the sensor sequence data is input into the neural radiation field model. The current spatial density and current color parameters of the previous physical field twin model are updated based on incremental training. The high-resolution network parameters in the previous physical twin model are converted into a sparse voxel mesh; Based on the sparse voxel mesh, the current spatial density, and the current color parameters, a lookup table index for spatial location and rendering attributes is constructed, and the current physical field twin model is reconstructed.

6. The data center operation and maintenance method according to claim 5, characterized in that, The step of converting the high-resolution network parameters in the previous physics twin model into a sparse voxel mesh includes: Obtain the spatial location of the equipment in the computer room; Based on the spatial location of the equipment, the computer room space is divided into an adjacent area and a background area. The sparse voxel mesh is obtained by sampling the high-resolution network parameters according to the first sampling density corresponding to the device's adjacent area and the second sampling density corresponding to the device's background area. The first sampling density is higher than the second sampling density.

7. The method for operating and maintaining a computer room according to any one of claims 1 to 3, characterized in that, The decision reference information associated with the target area includes at least one of the following: topological connections associated with the target area, historical operation and maintenance orders, energy quotas, and security thresholds. The topology connection includes at least one of spatial location topology, power supply cascade topology, and cooling coverage topology.

8. The method for operating and maintaining a computer room according to any one of claims 1 to 3, characterized in that, The acquisition of multimodal sensing data from the computer room includes: Acquire initial multi-source sensor data for the computer room; Perform a rendering consistency check on the image data and point cloud data in the initial multi-source sensor data to obtain the rendering consistency check result; Based on the rendering consistency test results, the initial multi-source sensor data is filtered to obtain the multimodal sensing data.

9. A computer room operation and maintenance device, characterized in that, include: The acquisition unit acquires multimodal sensing data from the computer room; The reconstruction unit inputs the multimodal sensing data into the neural radiation field model to reconstruct the previous physical field twin model and build the current physical field twin model of the computer room. The tensor summary generation unit, in response to a detected data center operation and maintenance event, extracts target tensor data of the target region corresponding to the data center operation and maintenance event from the current physical field twin model, integrates the target tensor data, and generates a physical tensor summary of the target region. The decision-making unit retrieves decision reference information associated with the target area, inputs the decision reference information and the physical tensor summary into the large language model, and obtains the data center operation and maintenance decision output by the large language model.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the operation and maintenance method of the computer room as described in any one of claims 1 to 7.

Images