Machine room operation and maintenance management method, device and electronic equipment

Abstract

The application provides a kind of machine room operation and maintenance management method, device and electronic equipment, method includes: obtaining the operation and maintenance instruction data of target machine room input by user, extracts the instruction feature vector of operation and maintenance instruction data;In the semantic vector space constructed in advance, the fusion feature vector of target equipment matched with instruction feature vector is retrieved;Semantic vector space is constructed based on the text feature vector, point cloud feature vector and topological feature vector of each equipment of target machine room;Instruction feature vector and the fusion feature vector of target equipment are input into large language model, and the operation and maintenance scheme of target machine room output by large language model is obtained.The application extracts instruction feature vector and retrieves in unified semantic vector space that fuses text, point cloud and topological feature, inputs instruction feature vector and the fusion feature vector of target equipment retrieved into large language model, generates operation and maintenance scheme, improves the efficiency and accuracy of operation and maintenance, and is suitable for complex application scenarios.

Description

Technical Field

[0001] This invention relates to the field of data center operation and maintenance technology, and in particular to a data center operation and maintenance management method, device and electronic equipment. Background Technology

[0002] As data centers expand in scale, the complexity of data center operations and maintenance (O&M) is increasing. Existing O&M management methods rely on static models and manual labels, which cannot quickly identify new or mobile devices, leading to a disconnect between digital twins and the physical environment. Furthermore, traditional methods depend on the experience of O&M personnel, who must locate devices layer by layer through complex graphical interfaces or manually write scripts to call underlying interfaces, resulting in low efficiency and a high risk of errors. While some solutions attempt to introduce language models to assist question answering, the different modalities of data reside in different vector spaces, making accurate semantic alignment difficult. This prevents the system from accurately parsing ambiguous commands and generating executable O&M solutions. Summary of the Invention

[0003] This invention provides a data center operation and maintenance management method, device, and electronic device to address the shortcomings of existing data center operation and maintenance management methods, such as low efficiency, poor accuracy, and poor adaptability.

[0004] This invention provides a data center operation and maintenance management method, comprising: Obtain the operation and maintenance instruction data for the target data center input by the user, and extract the instruction feature vector of the operation and maintenance instruction data; The fused feature vector of the target device that matches the instruction feature vector is retrieved from the pre-constructed semantic vector space; the semantic vector space is constructed based on the text feature vectors, point cloud feature vectors, and topological feature vectors of each device in the target computer room. The fused feature vector of the instruction and the feature vector of the target device are input into the large language model to obtain the operation and maintenance plan of the target data center output by the large language model.

[0005] In some embodiments, the process of constructing the semantic vector space includes: The text data and point cloud data of each device in the target data center are obtained, and the scene data of the target data center is also obtained; the text data includes operation log data and attribute data. Extract the text feature vector from the text data, and extract the point cloud feature vector from the point cloud data; Based on the scene data, the connection relationships between the devices are determined. A scene graph is constructed with each device as a node and the connection relationships between the devices as edges. The topological feature vector of the scene graph is then extracted. Align the text feature vector, the point cloud feature vector, and the topological feature vector to establish a mapping relationship between the text feature vector, the point cloud feature vector, and the topological feature vector.

[0006] In some embodiments, retrieving the fused feature vector of the target device that matches the instruction feature vector in a pre-constructed semantic vector space includes: The instruction feature vector is matched with the text feature vector, point cloud feature vector, and topological feature vector of each device to obtain the retrieval results; Based on the search results, the target text feature vector, target point cloud feature vector, and target topology feature vector of the target device that match the instruction feature vector are determined. The target text feature vector, the target point cloud feature vector, and the target topology feature vector are fused to obtain the fused feature vector of the target device.

[0007] In some embodiments, fusing the target text feature vector, the target point cloud feature vector, and the target topological feature vector includes: Based on the cross-spatial attention mechanism, the weights of the target text feature vector, the target point cloud feature vector, and the target topological feature vector are determined. Based on the weights of the target text feature vector, the target point cloud feature vector, and the target topology feature vector, the target text feature vector, the target point cloud feature vector, and the target topology feature vector are fused.

[0008] In some embodiments, after obtaining the operation and maintenance plan for the target data center output by the large language model, the method further includes: The operation and maintenance plan shall be subject to security verification. If the security verification is successful, control commands for the target device are generated based on the operation and maintenance plan, and the control commands are sent to the target device. After the target device executes the control command, the status data of the target device is acquired; Based on the state data of the target device, optimize the parameters of the semantic vector space and / or the large language model.

[0009] In some embodiments, the operation and maintenance solution includes a semantic operation protocol, which includes the following four elements: the user's subject identifier, the action for the target device, the identifier of the target device, and the parameters required to perform the action.

[0010] In some embodiments, the method further includes: Acquire new text data and new point cloud data of new devices in the target data center, and acquire new scene data of the target data center; the new text data includes new operation log data and new attribute data; The semantic vector space is updated based on the new text data, the new point cloud data, and the new scene data.

[0011] In some embodiments, the large language model is trained based on instruction feature vector samples corresponding to data center samples, fused feature vector samples corresponding to target device samples, and operation and maintenance scheme labels of data center samples.

[0012] The present invention also provides a data center operation and maintenance management device, comprising: The first acquisition unit is used to acquire operation and maintenance instruction data for the target data center input by the user, and extract the instruction feature vector of the operation and maintenance instruction data; The retrieval unit is used to retrieve the fused feature vector of the target device that matches the instruction feature vector in a pre-constructed semantic vector space; the semantic vector space is constructed based on the text feature vectors, point cloud feature vectors, and topological feature vectors of each device in the target computer room; The prediction unit is used to input the fused feature vector of the instruction and the feature vector of the target device into the large language model to obtain the operation and maintenance plan of the target data center output by the large language model.

[0013] 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 data center operation and maintenance management method described above.

[0014] The data center operation and maintenance management method, device, and electronic device provided by this invention acquire user-input operation and maintenance instruction data for a target data center and extract instruction feature vectors from the operation and maintenance instruction data. Based on the instruction feature vectors, a retrieval is performed in a unified semantic vector space containing text feature vectors, point cloud feature vectors, and topological feature vectors to obtain the fused feature vector of the target device, thus achieving accurate positioning of the target device. The instruction feature vectors and the fused feature vector of the target device are input into a large language model to obtain the operation and maintenance solution for the target data center output by the large language model, improving the efficiency and accuracy of operation and maintenance and making it suitable for complex application scenarios. Attached Figure Description

[0015] 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.

[0016] Figure 1 This is a flowchart illustrating the data center operation and maintenance management method provided in this embodiment of the invention.

[0017] Figure 2 This is a flowchart illustrating the construction process of the semantic vector space provided in this embodiment of the invention.

[0018] Figure 3 This is a schematic diagram of the structure of the data center operation and maintenance management device provided in an embodiment of the present invention.

[0019] Figure 4 This is a schematic diagram of the structure of the electronic device provided in an embodiment of 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] The terms "first," "second," etc., used in this invention are used to distinguish similar objects, not to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that embodiments of the invention can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first" and "second" are generally of the same class, not limited in number; for example, the first object can be one or more. Furthermore, in this invention, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0022] Figure 1 This is a flowchart illustrating the data center operation and maintenance management method provided in an embodiment of the present invention. Figure 1 As shown, a data center operation and maintenance management method is provided, including the following steps: step 110, step 120, and step 130. This method's steps are merely one possible implementation of the present invention.

[0023] Step 110: Obtain the operation and maintenance instruction data for the target data center input by the user, and extract the instruction feature vector of the operation and maintenance instruction data.

[0024] Optionally, the target data center can be a standalone data center building, one or more specific areas within a large data center, such as Area A, Cold Aisle Zone 3, or a cluster of multiple data centers deployed across different regions.

[0025] Operation and maintenance command data refers to the raw information entered by users to query, control, or manage equipment within the target data center. There are various ways to obtain this operation and maintenance command data. For example, users can input voice commands via microphone, in which case the operation and maintenance command data is audio stream data; users can also input text commands through input boxes in command-line interfaces, instant messaging software, or graphical user interfaces, in which case the operation and maintenance command data is text string data.

[0026] After obtaining the user-inputted operation and maintenance instructions, it needs to be processed to extract its semantic information, i.e., to extract the instruction feature vector. This extraction is achieved using a pre-trained language encoding model, such as the BERT model based on the Transformer architecture or other text or model known to those skilled in the art, to encode the operation and maintenance instructions. This process maps the input text string or converted text to a high-dimensional real-number vector, which is the instruction feature vector. This instruction feature vector can capture the core intent in the operation and maintenance instructions, including semantic information such as the object of operation, actions, and related constraints.

[0027] Step 120: Retrieve the fused feature vector of the target device that matches the instruction feature vector in the pre-constructed semantic vector space; the semantic vector space is constructed based on the text feature vectors, point cloud feature vectors and topological feature vectors of each device in the target computer room.

[0028] The semantic vector space is a pre-constructed multidimensional mathematical space. Its key feature is its ability to uniformly represent and organize data from different modalities, ensuring that semantically related entities are located close to each other. Specifically, this semantic vector space is constructed based on various heterogeneous information from devices within the target data center. This information includes at least: Text feature vector: Represents the textual description information of the device. This information can come from the device's static nameplate data, such as the device name, model, and manufacturer, or from the device's dynamic operation log data, such as alarm information and performance indicator records. The text feature vector is obtained by extracting the feature vectors of this textual information using a text encoding model.

[0029] Point cloud feature vectors: These represent the physical geometric information of a device. Three-dimensional point cloud data of the device is acquired through LiDAR scanning or depth camera imaging. A point cloud feature extraction network is then used to extract feature vectors from the point cloud data that characterize the device's shape, size, and spatial location; these are the point cloud feature vectors.

[0030] Topology feature vectors represent the relationships between devices. These relationships include power supply links, network links, and environmental links. By encoding the graph structure formed by these relationships using techniques such as graph neural networks, topology feature vectors that characterize the position and role of each device within the entire data center system are obtained.

[0031] The retrieval process involves finding the device that best matches the instruction feature vector within a pre-constructed semantic vector space. Specific retrieval methods can employ vector similarity calculations, such as calculating the cosine similarity or Euclidean distance between the instruction feature vector and the fused feature vector of each device in the semantic vector space. The fused feature vector is the final representation of each device in a unified semantic vector space, integrating feature information from the device's text, point cloud, and topological modalities. The goal of the retrieval is to find the fused feature vectors of one or more devices with the highest similarity or closest proximity.

[0032] Step 130: Input the fused feature vector of the instruction and the feature vector of the target device into the large language model to obtain the operation and maintenance plan of the target data center output by the large language model.

[0033] This step leverages the powerful reasoning and generation capabilities of large language models to formulate specific operation and maintenance plans. Large language models refer to deep learning models with a huge number of parameters that have been pre-trained on massive amounts of data, such as the GPT series models or specialized large language models fine-tuned with industry data, which possess excellent contextual understanding, logical reasoning, and text generation capabilities.

[0034] After receiving the fused feature vectors of the instruction and the target device, the large language model comprehensively understands the user's instruction and the specific situation of the target device, performs inference analysis, and ultimately outputs an operation and maintenance plan for the target data center. This plan includes specific steps, suggestions, or executable plans to address the user's instruction. For example, for the instruction: "Lower the temperature of the air conditioner in the cabinet with the highest temperature in Area A by 2 degrees Celsius," after retrieving the corresponding cabinet and its associated air conditioner, the large language model outputs an operation and maintenance plan that suggests lowering the temperature of air conditioner AC-05 by 2 degrees Celsius. This plan can be a suggestion described in natural language or a more structured data format, providing a foundation for subsequent automated execution.

[0035] In some embodiments, the operation and maintenance solution includes a semantic operation protocol, which includes the following four elements: the user's subject identifier, the action for the target device, the identifier of the target device, and the parameters required to perform the action.

[0036] In this embodiment of the invention, the instruction feature vector of the operation and maintenance instruction data is extracted; based on the instruction feature vector, a retrieval is performed in a unified semantic vector space containing text feature vectors, point cloud feature vectors, and topological feature vectors to obtain the fused feature vector of the target device. This accurately semantically associates the user's natural language instructions with the complex three-dimensional physical entities and their system relationships in the data center, achieving precise positioning of the target device; by inputting the instruction feature vector and the fused feature vector of the target device into a large language model, an operation and maintenance solution is obtained, thereby realizing the intelligent conversion from fuzzy natural language input to clear operation and maintenance intentions, greatly improving the intelligence level of data center operation and maintenance and the efficiency of human-computer interaction, improving the accuracy of operation and maintenance, and making it suitable for complex application scenarios.

[0037] Figure 2 This is a flowchart illustrating the construction process of the semantic vector space provided in an embodiment of the present invention. Figure 2 As shown, in some embodiments, the process of constructing the semantic vector space includes: Step 210: Obtain text data and point cloud data of each device in the target data center, and obtain scene data of the target data center; text data includes operation log data and attribute data.

[0038] Text data refers to textual information related to various devices within the target data center. This text data can be further divided into two categories: first, attribute data, which is typically static information about the devices, such as device name, model, manufacturer, purchase date, and rack number obtained from asset management databases or equipment design drawings; second, operational log data, which is typically dynamic information about the devices, such as real-time device operation status logs, performance indicator records, and alarm information obtained from data center infrastructure management systems or building management systems. This data can be structured key-value pairs or unformatted text logs.

[0039] Point cloud data refers to geometric data that characterizes the three-dimensional physical form of equipment. This can be achieved by deploying high-resolution LiDAR or depth cameras within the data center. The scanning equipment can periodically perform a global scan of the data center, or trigger a scan when environmental changes are detected, generating a raw point cloud covering the entire data center. The acquired raw point cloud data may contain noise caused by dust reflections from the air or surface reflections from the equipment; this noise can be denoised and registered before subsequent processing.

[0040] Scenario data refers to global information that reflects the relationships between various devices within the target data center. This data can come from various sources, such as data center design drawings (CAD floor plans, power supply topology diagrams, network cabling diagrams), or existing device association configuration information in the data center operation and maintenance management platform. Scenario data forms the foundation for subsequently constructing the topology relationships between devices.

[0041] Step 220: Extract the text feature vector from the text data and extract the point cloud feature vector from the point cloud data.

[0042] Alternatively, a natural language processing model known in the field can be used, such as a pre-trained language model based on the Transformer architecture. The device's attribute data and operation log data are input into the model, and the output vector is the text feature vector, which contains the device's textual semantic information.

[0043] Optionally, the overall point cloud data of the target computer room can be voxelized and downsampled to reduce the data volume. Then, the point cloud can be segmented into local point cloud blocks corresponding to a single device, such as a server rack or an air conditioner, through methods such as spatial clustering or connectivity analysis. Next, these point cloud blocks are input into a pre-trained point cloud feature extraction network, such as a point cloud model using a mask autoencoder mechanism. The network learns to reconstruct the occluded geometric structure in a self-supervised task, thereby outputting a robust point cloud feature vector containing rich geometric semantic information.

[0044] Step 230: Based on the scene data, determine the connection relationship between each device, construct a scene graph with each device as a node and the connection relationship between each device as an edge, and extract the topological feature vector of the scene graph.

[0045] This step aims to capture the role and relationships of devices within the entire data center system from a macro perspective. First, based on scenario data, the connection relationships between various devices are analyzed and determined. These connections are not limited to physical connections but can also include logical influence relationships.

[0046] Once the connections are established, a scene graph can be constructed. In this scene graph, each individual device in the data center, such as a server, switch, air conditioner, or power distribution unit, is abstracted as a node; and the connections between devices are abstracted as edges connecting the corresponding nodes. These edges can have different types and attributes, such as power supply relationships, network relationships, and cooling relationships.

[0047] Finally, the topological feature vector of the scene graph is extracted. This can be achieved using a graph neural network. The constructed scene graph is input into a pre-trained graph neural network model, such as a graph autoregressive network or a graph attention network. The model aggregates the neighbor node information of each node, learns the contextual representation of each device in the entire graph structure, and finally generates a fixed-dimensional topological feature vector for each device node. This vector encodes the importance and relevance of the device in the data center system topology.

[0048] Step 240: Align the text feature vector, point cloud feature vector, and topological feature vector to establish a mapping relationship between them.

[0049] This step is crucial for constructing a unified semantic vector space. Its goal is to align the feature vectors extracted from the three different modalities in the preceding steps. Alignment means that, by learning one or more mapping functions, the text feature vectors, point cloud feature vectors, and topological feature vectors, which were originally in their own independent feature spaces, are projected into the same shared, unified semantic vector space.

[0050] The goal of establishing a mapping relationship is to ensure that, within the aligned unified semantic vector space, feature vectors from three different modalities belonging to the same physical device are as close as possible in spatial location, while feature vectors from different physical devices are as far apart as possible. This mapping relationship can be established in various ways. For example, a deep neural network model with three input branches can be designed and trained using a contrastive learning strategy. During training, text, point cloud, and topological feature vectors from the same device are used as positive sample pairs, and feature vectors from different devices are used as negative sample pairs. The loss function is optimized to bring positive sample pairs closer together and distance negative sample pairs further apart. After sufficient training, the model learns how to map feature vectors from different modalities to a unified semantic vector space, thus completing the establishment of the mapping relationship.

[0051] Through the steps described in this embodiment, the three heterogeneous pieces of information—textual description, physical form, and system relationships—of each device in the data center are successfully encoded and aligned into a unified semantic vector space. This not only allows for the construction of a comprehensive and rich multimodal digital profile for each device but also lays a solid data foundation for subsequent cross-modal, natural language-based accurate retrieval and intelligent understanding, thereby significantly improving the accuracy and intelligence of the entire operation and maintenance management method.

[0052] In some embodiments, step 120 retrieves the fused feature vector of the target device that matches the instruction feature vector in a pre-constructed semantic vector space, including: Step 121: Match the instruction feature vector with the text feature vector, point cloud feature vector, and topological feature vector of each device to obtain the retrieval results.

[0053] Optionally, the similarity between the instruction feature vector and the text feature vector, point cloud feature vector, and topological feature vector of each device stored in the semantic vector space can be calculated separately.

[0054] The final retrieval result obtained from this step is a list of one or more candidate devices, each accompanied by its matching score or similarity with the instruction feature vector in different modalities.

[0055] Step 122: Based on the search results, determine the target text feature vector, target point cloud feature vector, and target topology feature vector of the target device that match the instruction feature vector.

[0056] This step builds upon the initial matching in the previous step, making decisions and filtering to ultimately determine one or more target devices that best match the user's intent.

[0057] Optionally, a total score is calculated by combining the matching scores of each candidate device across all modalities, and the device with the highest score is identified as the target device. For example, the matching scores of text, point cloud, and topology modalities can be weighted and summed, with the weights either preset or dynamically adjusted based on the nature of the instruction. Alternatively, a ranking learning model can be used to reorder the candidate device list to more accurately determine the final target device.

[0058] Once the target device is identified, the complete set of feature vectors corresponding to that target device can be obtained, namely the target text feature vector, target point cloud feature vector, and target topology feature vector. If multiple target devices are identified, the set of feature vectors corresponding to all of these target devices is obtained.

[0059] Step 123: Fuse the target text feature vector, target point cloud feature vector, and target topology feature vector to obtain the fused feature vector of the target device.

[0060] The purpose of this step is to integrate the multiple single-modal feature vectors of the target device identified in the previous step into a single, more comprehensive fused feature vector. This fused feature vector will serve as the core contextual information input to the large language model.

[0061] There are several ways to fuse features. A simple approach is vector concatenation, which involves linking the target text feature vector, target point cloud feature vector, and target topological feature vector in a predetermined order into a longer vector. A more advanced approach uses a small neural network, such as a multilayer perceptron, taking the three single-modal feature vectors as input and outputting the fused feature vector. This method can learn better fusion strategies through training. Another approach utilizes attention mechanisms, calculating the importance of different modal features to the current instruction and then weighting and fusing them.

[0062] Through the retrieval and fusion process described in this embodiment, the present invention can respond to complex user commands more flexibly and accurately. By matching features of different modalities in parallel, the system can fully utilize the advantages of each data source to understand different aspects of the user's intent. Then, through decision-making and fusion, a highly condensed and comprehensive representation of the target device is finally formed. Compared with direct retrieval on a single fusion vector, this method has better interpretability and the ability to parse complex and fuzzy commands, thus providing higher-quality input for generating accurate operation and maintenance solutions for large language models, further improving the reliability and accuracy of the entire operation and maintenance management process.

[0063] In some embodiments, step 123, fusing the target text feature vector, the target point cloud feature vector, and the target topological feature vector, includes: Step 1231: Based on the cross-space attention mechanism, determine the weights of the target text feature vector, the target point cloud feature vector, and the target topological feature vector.

[0064] The core idea of ​​the cross-space attention mechanism is to dynamically determine the importance of information from the three modalities of text, point cloud, and topology based on the current input operation and maintenance instructions, and assign different weights to the three accordingly.

[0065] Step 1232: Based on the weights of the target text feature vector, the target point cloud feature vector, and the target topological feature vector, fuse the target text feature vector, the target point cloud feature vector, and the target topological feature vector.

[0066] Optionally, the target text feature vector is multiplied by its corresponding weight, the target point cloud feature vector is multiplied by its corresponding weight, and the target topological feature vector is multiplied by its corresponding weight. Then, these three weighted vectors are added together to obtain the final fused feature vector.

[0067] The dynamic fusion method based on cross-space attention mechanism described in this embodiment transforms the feature fusion process from a static or simple concatenation into an adaptive process closely related to user intent. This significantly improves the representational ability of the fused feature vectors and their relevance to the current task, providing higher-quality input information for generating more accurate and targeted operation and maintenance solutions for subsequent large language models. Consequently, it enhances the intelligence level and decision-making quality of the entire operation and maintenance management method.

[0068] In some embodiments, after obtaining the operation and maintenance plan for the target data center output by the large language model, the system further includes: Perform security verification on the operation and maintenance plan; Once the security verification is confirmed to be successful, control commands for the target device are generated based on the operation and maintenance plan, and the control commands are sent to the target device. After the target device completes the control command, acquire the target device's status data; Based on the state data of the target device, optimize the parameters of the semantic vector space and / or the large language model.

[0069] First, performing security verification on the operation and maintenance plan is a key protection step after the large language model outputs the operation and maintenance plan and before its actual execution. Its purpose is to prevent potential misoperations or execution of high-risk instructions, and to ensure the stability and security of the data center operation.

[0070] Security checks may include, but are not limited to, the following two aspects: First, threshold checks, which determine whether the operational parameters included in the maintenance plan are within preset safety ranges. For example, if the plan involves temperature regulation, it checks whether the target temperature exceeds the safe operating temperature threshold; if the plan involves power adjustment, it determines whether the adjusted current will exceed the rated load of the power supply line or power distribution unit. Second, mutual exclusion analysis, which determines whether the actions in the maintenance plan logically conflict with other ongoing operations or equipment statuses in the current data center. For example, it checks whether the power switching operation to be performed will affect servers performing critical business operations, or whether a running command is being issued to equipment marked for maintenance. If the security check fails, the maintenance plan is intercepted, the execution process is terminated, and a risk alert is issued to the user, requesting manual confirmation.

[0071] Specifically, generating control commands for target devices based on the operation and maintenance plan involves transforming the high-level semantic plan generated by the large language model into machine instructions that the underlying hardware can recognize and execute. This transformation can be accomplished by a protocol conversion module or a protocol conversion gateway. Based on the operation and maintenance plan, this module queries a mapping table to generate low-level control commands conforming to standard industrial bus protocols. Then, through appropriate communication interfaces, such as serial ports or Ethernet ports, it sends these control commands to the target devices retrieved in the semantic vector space, such as air conditioners or power distribution units.

[0072] Optionally, the acquired target device status data is compared with the expected results in the operation and maintenance plan. If there is a significant deviation, such as a command requiring a 2-degree temperature reduction but only a 0.5-degree reduction actually occurring, or a command being sent to the wrong device, the system will trigger an optimization process. Optimization can proceed in two directions: First, optimize the semantic vector space. If the deviation is due to an initial target device retrieval error, it indicates a deficiency in the alignment relationships within the semantic vector space. The system can treat this failed interaction—the matching pair from the maintenance command to the incorrect target device—as a negative sample and fine-tune the network model parameters used for feature extraction and alignment through online incremental learning. This approach adjusts the positions of relevant vectors in the semantic vector space, improving the accuracy of future retrievals.

[0073] Second, optimize the parameters of the large language model. If the target device is selected correctly, but the operation and maintenance plan generated by the large language model itself has logical flaws or is not optimized enough, the system can use the complete link data of this interaction, including instructions, context, incorrect solutions, and correct feedback results or manually corrected solutions, as a new training sample to fine-tune the large language model, thereby improving its logical reasoning ability and solution quality when generating operation and maintenance plans in the future.

[0074] The security verification, instruction execution, effect feedback, and closed-loop optimization processes added in this embodiment greatly ensure the security and reliability of the automated operation and maintenance process and improve the system's adaptability.

[0075] In some embodiments, the above method further includes: Acquire new text data and new point cloud data of new devices in the target data center, and acquire new scene data of the target data center; the new text data includes new operation log data and new attribute data; The semantic vector space is updated based on new text data, new point cloud data, and new scene data.

[0076] First, acquire new text data and point cloud data of new devices in the target data center, as well as new scene data of the target data center. This step aims to detect physical and logical changes within the data center, such as the installation of new servers, replacement of old equipment, or adjustments to network cabling. These new devices or connections can be discovered by periodically scanning or comparing data collected before and after.

[0077] Secondly, the semantic vector space is updated based on new text data, new point cloud data, and new scene data. This step integrates newly introduced devices or changed relationships into the existing model. Feature extraction can be performed on the new data to generate text feature vectors, point cloud feature vectors, and topological feature vectors for the new devices. Then, the network model used for modality alignment is updated through online incremental learning or lightweight fine-tuning, mapping and adding these newly generated feature vectors to the unified semantic vector space. In this way, zero-shot recognition of new devices can be achieved without retraining the entire model, ensuring continuous synchronization between the digital twin and the physical environment, and significantly improving the system's dynamic adaptability and scalability.

[0078] In some embodiments, the large language model is trained based on instruction feature vector samples corresponding to data center samples, fused feature vector samples corresponding to target device samples, and operation and maintenance scheme labels of data center samples.

[0079] The training process in this step can be understood as a form of supervised fine-tuning. Its purpose is to enable the large language model to learn the mapping relationship from user intent and device context to specific operation and maintenance solutions.

[0080] Among them, the instruction feature vector samples cover various query, control and management requests that may occur in the operation and maintenance scenario; the fused feature vector samples represent the multimodal feature representation of the target device pointed to by the corresponding instruction.

[0081] Optionally, obtain operation and maintenance instruction data samples for the data center sample, and perform feature extraction on the operation and maintenance instruction data samples to obtain instruction feature vectors; Optionally, based on the instruction feature vector samples, a search is performed in the pre-constructed semantic vector space samples to obtain the target text feature vector samples, target point cloud feature vector samples, and target topology feature vector samples corresponding to the matching target device samples; the target text feature vector samples, target point cloud feature vector samples, and target topology feature vector samples are fused to obtain the fused feature vector samples corresponding to the target device samples.

[0082] Among them, the semantic vector space samples are constructed based on the text feature vector samples, point cloud feature vector samples, and topological feature vector samples corresponding to each device sample in the data center sample.

[0083] Optionally, the process of constructing semantic vector space samples includes: Obtain text data samples and point cloud data samples corresponding to each device sample in the data center sample, and obtain scene data samples corresponding to the data center sample; the text data samples include operation log data samples and attribute data samples; Extract text feature vector samples corresponding to text data samples, and extract point cloud feature vector samples corresponding to point cloud data samples; Based on scene data samples, the connection relationships between each device sample are determined. Using each device sample as a node and the connection relationships between each device sample as edges, a scene graph sample is constructed, and the topological feature vector samples corresponding to the scene graph sample are extracted. Align text feature vector samples, point cloud feature vector samples, and topological feature vector samples to establish a mapping relationship between them.

[0084] During training, the system concatenates instruction feature vector samples and fused feature vector samples as input to the large language model, which generates a predicted operation and maintenance (O&M) plan. Then, by comparing the predicted plan with the actual O&M plan labels, the loss function is calculated, and the internal parameters of the large language model are updated using the backpropagation algorithm. By repeating this process on a large number of data center samples, the large language model's capabilities are specialized from a general model into an expert model proficient in data center O&M.

[0085] Through the targeted training described in this embodiment, the large language model can deeply learn the specific knowledge, logic, and constraints of data center operation and maintenance. As a result, when faced with actual input, it can generate more accurate, secure, and industry-compliant operation and maintenance solutions, significantly improving the reliability and professionalism of the method described in this invention.

[0086] The data center operation and maintenance management device provided in the embodiments of the present invention is described below. The data center operation and maintenance management device described below can be referred to in correspondence with the data center operation and maintenance management method described above.

[0087] Figure 3 This is a schematic diagram of the structure of the data center operation and maintenance management device provided in an embodiment of the present invention, as shown below. Figure 3 As shown, the data center operation and maintenance management device 300 includes: The first acquisition unit 310 is used to acquire the operation and maintenance instruction data for the target data center input by the user, and extract the instruction feature vector of the operation and maintenance instruction data; The retrieval unit 320 is used to retrieve the fused feature vector of the target device that matches the instruction feature vector in a pre-constructed semantic vector space; the semantic vector space is constructed based on the text feature vectors, point cloud feature vectors, and topological feature vectors of each device in the target computer room; The prediction unit 330 is used to input the fused feature vector of the instruction and the feature vector of the target device into the large language model to obtain the operation and maintenance plan of the target data center output by the large language model.

[0088] Optionally, the process of constructing the semantic vector space includes: Acquire text data and point cloud data of each device in the target data center, and acquire scene data of the target data center; text data includes operation log data and attribute data; Extract text feature vectors from text data and extract point cloud feature vectors from point cloud data; Based on scene data, the connection relationships between devices are determined. Using each device as a node and the connection relationships between devices as edges, a scene graph is constructed, and the topological feature vector of the scene graph is extracted. Align text feature vectors, point cloud feature vectors, and topological feature vectors to establish a mapping relationship between them.

[0089] Optionally, a fused feature vector of the target device matching the instruction feature vector is retrieved from a pre-constructed semantic vector space, including: The command feature vector is matched with the text feature vector, point cloud feature vector, and topological feature vector of each device to obtain the retrieval results; Based on the search results, the target text feature vector, target point cloud feature vector, and target topology feature vector of the target device that match the instruction feature vector are determined. The target text feature vector, target point cloud feature vector, and target topology feature vector are fused to obtain the target device's fused feature vector.

[0090] Optionally, the target text feature vector, target point cloud feature vector, and target topological feature vector are fused, including: Based on the cross-spatial attention mechanism, the weights of the target text feature vector, the target point cloud feature vector, and the target topological feature vector are determined. Based on the weights of the target text feature vector, the target point cloud feature vector, and the target topological feature vector, the target text feature vector, the target point cloud feature vector, and the target topological feature vector are fused.

[0091] Optionally, the data center operation and maintenance management device also includes: The verification unit is used to perform security verification on the operation and maintenance plan. The generation unit is used to generate control commands for the target device based on the operation and maintenance plan after confirming that the security verification has passed, and then send the control commands to the target device. The second acquisition unit is used to acquire the status data of the target device after the target device has executed the control command; The optimization unit is used to optimize the parameters of the semantic vector space and / or the large language model based on the state data of the target device.

[0092] Optionally, the operation and maintenance solution includes a semantic operation protocol, which includes the following four elements: the user's subject identifier, the action for the target device, the identifier of the target device, and the parameters required to perform the action.

[0093] Optionally, the data center operation and maintenance management device also includes: The third acquisition unit is used to acquire new text data and new point cloud data of new devices in the target data center, and to acquire new scene data of the target data center; the new text data includes new operation log data and new attribute data; The update unit is used to update the semantic vector space based on new text data, new point cloud data, and new scene data.

[0094] Optionally, the large language model is trained based on the instruction feature vector samples corresponding to the data center samples, the fused feature vector samples corresponding to the target device samples, and the operation and maintenance scheme labels of the data center samples.

[0095] It should be noted that the data center operation and maintenance management device provided in this embodiment of the invention can implement all the method steps implemented in the above-mentioned data center operation and maintenance management method embodiment, and can achieve the same technical effect. Here, the parts that are the same as those in the method embodiment and the beneficial effects will not be described in detail.

[0096] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention, such as... Figure 4 As shown, the electronic device may include a processor 410, a communications interface 420, a memory 430, and a communication bus 440. The processor 410, communications interface 420, and memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions stored in the memory 430 to execute a data center operation and maintenance management method. This method includes: acquiring operation and maintenance instruction data for the target data center input by the user; extracting instruction feature vectors from the operation and maintenance instruction data; retrieving the fused feature vector of the target device matching the instruction feature vector in a pre-constructed semantic vector space; the semantic vector space is constructed based on the text feature vectors, point cloud feature vectors, and topological feature vectors of each device in the target data center; and inputting the instruction feature vector and the fused feature vector of the target device into a large language model to obtain the operation and maintenance plan for the target data center output by the large language model.

[0097] 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.

[0098] 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.

[0099] 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.

[0100] 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 data center operation and maintenance management method, characterized in that, include: Obtain the operation and maintenance instruction data for the target data center input by the user, and extract the instruction feature vector of the operation and maintenance instruction data; The fused feature vector of the target device that matches the instruction feature vector is retrieved from the pre-constructed semantic vector space; the semantic vector space is constructed based on the text feature vectors, point cloud feature vectors, and topological feature vectors of each device in the target computer room. The fused feature vector of the instruction and the feature vector of the target device are input into the large language model to obtain the operation and maintenance plan of the target data center output by the large language model.

2. The data center operation and maintenance management method according to claim 1, characterized in that, The process of constructing the semantic vector space includes: The text data and point cloud data of each device in the target data center are obtained, and the scene data of the target data center is also obtained; the text data includes operation log data and attribute data. Extract the text feature vector from the text data, and extract the point cloud feature vector from the point cloud data; Based on the scene data, the connection relationships between the devices are determined. A scene graph is constructed with each device as a node and the connection relationships between the devices as edges. The topological feature vector of the scene graph is then extracted. Align the text feature vector, the point cloud feature vector, and the topological feature vector to establish a mapping relationship between the text feature vector, the point cloud feature vector, and the topological feature vector.

3. The data center operation and maintenance management method according to claim 1, characterized in that, The step of retrieving the fused feature vector of the target device that matches the instruction feature vector from the pre-constructed semantic vector space includes: The instruction feature vector is matched with the text feature vector, point cloud feature vector, and topological feature vector of each device to obtain the retrieval results; Based on the search results, the target text feature vector, target point cloud feature vector, and target topology feature vector of the target device that match the instruction feature vector are determined. The target text feature vector, the target point cloud feature vector, and the target topology feature vector are fused to obtain the fused feature vector of the target device.

4. The data center operation and maintenance management method according to claim 3, characterized in that, The fusion of the target text feature vector, the target point cloud feature vector, and the target topological feature vector includes: Based on the cross-spatial attention mechanism, the weights of the target text feature vector, the target point cloud feature vector, and the target topological feature vector are determined. Based on the weights of the target text feature vector, the target point cloud feature vector, and the target topology feature vector, the target text feature vector, the target point cloud feature vector, and the target topology feature vector are fused.

5. The data center operation and maintenance management method according to claim 1, characterized in that, After obtaining the operation and maintenance plan for the target data center output by the large language model, the method further includes: The aforementioned operation and maintenance plan shall be subject to security verification; If the security verification is successful, control commands for the target device are generated based on the operation and maintenance plan, and the control commands are sent to the target device. After the target device executes the control command, the status data of the target device is acquired; Based on the state data of the target device, optimize the parameters of the semantic vector space and / or the large language model.

6. The data center operation and maintenance management method according to claim 1, characterized in that, The operation and maintenance solution includes a semantic operation protocol, which includes the following four elements: the user's subject identifier, the action for the target device, the identifier of the target device, and the parameters required to execute the action.

7. The data center operation and maintenance management method according to claim 1, characterized in that, The method further includes: Acquire new text data and new point cloud data of new devices in the target data center, and acquire new scene data of the target data center; the new text data includes new operation log data and new attribute data; The semantic vector space is updated based on the new text data, the new point cloud data, and the new scene data.

8. The data center operation and maintenance management method according to claim 1, characterized in that, The large language model is trained based on instruction feature vector samples corresponding to data center samples, fused feature vector samples corresponding to target device samples, and operation and maintenance scheme labels of data center samples.

9. A data center operation and maintenance management device, characterized in that, include: The first acquisition unit is used to acquire operation and maintenance instruction data for the target data center input by the user, and extract the instruction feature vector of the operation and maintenance instruction data; The retrieval unit is used to retrieve the fused feature vector of the target device that matches the instruction feature vector in a pre-constructed semantic vector space; the semantic vector space is constructed based on the text feature vectors, point cloud feature vectors, and topological feature vectors of each device in the target computer room; The prediction unit is used to input the fused feature vector of the instruction and the feature vector of the target device into the large language model to obtain the operation and maintenance plan of the target data center output by the large language model.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the data center operation and maintenance management method as described in any one of claims 1 to 8.

Images