An AI and knowledge graph-based server demand intelligent configuration method

Abstract

The application discloses an AI and knowledge graph-based server demand intelligent configuration method, and belongs to the technical field of intelligent product development management.The application constructs an enhanced intelligent interaction template comprising a natural language processing input box and an intelligent recommendation and conflict detection module, and a product knowledge graph with a server product as a core, uses a pre-trained natural language processing model to identify key entities and intentions in demand text and automatically fills the template, completes configuration compatibility verification, intelligent recommendation and conflict resolution based on the knowledge graph, finally maps the configuration content to a graph node and traverses and aggregates demand baseline atomic tasks to generate a demand decomposition document.The application realizes automatic analysis and configuration of unstructured demands, solves configuration conflicts in the demand reporting stage, improves the efficiency and quality of server customized development, and forms reusable enterprise product knowledge assets.

Description

Technical Field

[0001] This invention relates to the field of intelligent product development and management technology, and in particular to an intelligent configuration method for server requirements based on AI and knowledge graphs. Background Technology

[0002] As a core component of information technology infrastructure, servers are widely used in various fields such as cloud computing, big data processing, artificial intelligence training, and enterprise information system construction. With the deepening of digital transformation across industries, different application scenarios have placed differentiated demands on server performance, functionality, hardware configuration, and customization attributes. Customized server development has become the mainstream product delivery model in the industry. In the customized server development process, the decomposition of requirements from the market to the R&D side is a crucial node connecting customer needs with product development. Currently, the industry generally uses standardized requirement templates as the carrier for requirement transmission. The R&D department pre-designs a structured template containing product models, hardware component configurations, software parameter settings, and common customization options. Frontline marketing personnel select and fill in the template according to customer needs, and then submit the completed template to the R&D department for subsequent requirement decomposition and development work.

[0003] Existing standardized requirement templates have several limitations in practical applications, failing to meet the increasingly complex needs of customized server development. First, standardized templates can only handle structured checkboxes and fill-in content, unable to automatically identify and parse unstructured free text requirements filled in by frontline staff in specific fields. This content still requires manual interpretation and conversion by developers, increasing information loss and communication costs during requirement transmission. Second, standardized requirement templates lack real-time configuration validation capabilities, failing to detect compatibility, dependencies, and mutual exclusion between different configuration items during the requirement submission phase. Incorrect configuration schemes can only be discovered during developer review, easily leading to rework and delays in subsequent development. Finally, standardized requirement templates lack deep semantic connections with the backend requirement baseline library. The matching and decomposition of requirements into development tasks heavily relies on the personal experience of developers, making it difficult to guarantee the consistency and completeness of requirement decomposition and limiting the improvement of requirement processing efficiency. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a server demand intelligent configuration method based on AI and knowledge graph.

[0005] The objective of this invention is achieved through the following technical solution: A method for intelligent configuration of server requirements based on AI and knowledge graphs is provided, which includes the following steps: S1. Construct an enhanced intelligent interaction template and a product knowledge graph. The enhanced intelligent interaction template includes a natural language processing input box and an intelligent recommendation and conflict detection module. The product knowledge graph is centered on server products and includes nodes and edges. Nodes include product models, hardware components, software configurations, customized attributes, and baseline atomic tasks. Edges represent the relationships between nodes. S2. Receive the request text input into the natural language processing input box of the enhanced intelligent interaction template, identify the key entities and intents in the request text through a pre-trained natural language processing model, and automatically fill in the enhanced intelligent interaction template based on the key entities and intents. S3. Real-time acquisition of configuration content in the enhanced intelligent interaction template, compatibility verification, intelligent recommendation and conflict resolution of configuration content in the enhanced intelligent interaction template based on product knowledge graph, and feedback of compatibility verification results, intelligent recommendation content and conflict resolution results to the enhanced intelligent interaction template; S4. Obtain the final configuration content in the enhanced intelligent interaction template, map the final configuration content to a set of nodes in the product knowledge graph, traverse the product knowledge graph to obtain the requirement baseline atomic tasks required to cover the set of nodes, aggregate the document content corresponding to the requirement baseline atomic tasks, and generate a requirement decomposition document.

[0006] Furthermore, step S1 includes the following sub-steps: S1.1. An enhanced intelligent interactive template is constructed by adding a natural language processing input box and an intelligent recommendation and conflict detection module to the traditional structured template. The traditional structured template includes a product model selection block, a hardware component configuration block, a software configuration block, and a customized attribute block. The natural language processing input box is set in the top area of ​​the traditional structured template, and the intelligent recommendation and conflict detection module is embedded below each block of the traditional structured template. S1.2. Construct a product knowledge graph with server products as the core. The product knowledge graph is stored using resource description framework triples. The nodes of the product knowledge graph include product model, hardware components, software configuration, customized attributes, and requirement baseline atomic tasks. The edges of the product knowledge graph include containment, dependency, mutual exclusion, substitutability, and implementation requirement relationships. Each edge contains a weight value.

[0007] Furthermore, step S2 includes the following sub-steps: S2.1. Receive the request text input into the natural language processing input box of the enhanced intelligent interactive template; S2.2. Named entity recognition is performed on the requirement text using a pre-trained natural language processing model to extract key entities from the requirement text. The pre-trained natural language processing model adopts the BERT-base model, which contains a set number of Transformer encoders. Each Transformer encoder contains a multi-head self-attention mechanism and a feedforward neural network. The number of heads in the multi-head self-attention mechanism is a set value, and the hidden layer dimension of the feedforward neural network is a set value. S2.3. Use a pre-trained natural language processing model to identify the intent of the request text and obtain the intent corresponding to the request text; S2.4. Automatically fill in the corresponding blocks of the enhanced smart interaction template based on key entities and intents.

[0008] Furthermore, step S3 includes the following sub-steps: S3.1. Real-time acquisition of the configuration content of each block in the enhanced intelligent interaction template, with an acquisition frequency of a set value; S3.2. Convert the obtained configuration content into a node query statement for the product knowledge graph, and query the nodes in the product knowledge graph that correspond to the configuration content and the edges between the nodes; S3.3. Based on the query results, perform compatibility checks, intelligent recommendations, and conflict resolution on the configuration content. The compatibility check detects mutual exclusion relationships in the configuration content, the intelligent recommendation generates a list of alternative components, and the conflict resolution generates alternative solutions for conflicting content. S3.4. Send the compatibility verification results, intelligent recommended content, and conflict resolution results to the intelligent recommendation and conflict detection module of the enhanced intelligent interaction template for display.

[0009] Furthermore, step S4 includes the following sub-steps: S4.1. Receive the confirmation signal submitted by the enhanced smart interaction template and obtain the final configuration content of all blocks in the enhanced smart interaction template; S4.2. Match each configuration item in the final configuration content with the nodes in the product knowledge graph to generate a node set consisting of successfully matched nodes; S4.3. Starting from each node in the node set, traverse along the implementation requirement relationships in the product knowledge graph and collect all the requirement baseline atomic task nodes traversed; S4.4. Extract the document content corresponding to each atomic task node of the requirement baseline from the requirement baseline library, and concatenate the extracted document content according to the preset document structure to generate the requirement decomposition document.

[0010] Furthermore, in step S1.2, the hardware components include a central processing unit, memory, hard disk, independent disk redundant array card, network card, power supply and graphics card; the software configuration includes baseboard management controller parameters, basic input / output system parameters, operating system version and driver version; the customized attributes include nameplate, label, packaging identification and service label; the hardware component node includes interface type, power consumption parameters and size parameters; the software configuration node includes version number and a list of supported hardware; and the customized attribute node includes manufacturing process, size parameters and material parameters.

[0011] Furthermore, in step S2.2, named entity recognition adopts a sequence labeling method, and the labeling system adopts the BIO labeling system, labeling each character as B-entity type, I-entity type, or O. The output layer of the pre-trained natural language processing model contains two branches. The first branch outputs the entity type probability distribution of each character, and the second branch outputs the intent type probability distribution of the demand text. The intent types include performance requirements, functional requirements, customization requirements, and compatibility requirements. Key entities include product model, hardware components, software configuration, and customization attributes.

[0012] Furthermore, in step S3.3, the compatibility verification includes detecting whether there are two nodes with a mutually exclusive relationship in the configuration content. If they exist, it is determined to be a configuration conflict. The intelligent recommendation includes querying the product knowledge graph for nodes with a substitutable relationship with the node in the configuration content, sorting them according to the edge weight values, and generating a recommendation list. The conflict resolution includes sending the two conflicting nodes and the corresponding list of substitutable nodes to the intelligent recommendation and conflict detection module of the enhanced intelligent interaction template when a configuration conflict is detected.

[0013] Furthermore, in step S4.3, the product knowledge graph is traversed using a depth-first traversal algorithm. The starting node of the depth-first traversal algorithm is all nodes in the node set. During the traversal, visited nodes are marked to avoid repeated visits. During the traversal, only the implementation requirement relationship is traversed, and other relationships are not traversed. After the traversal is completed, the collected requirement baseline atomic task nodes are deduplicated to generate the final requirement baseline atomic task list.

[0014] Furthermore, in step S4, when generating the requirement decomposition document, the notes in the requirement text parsed by the pre-trained natural language processing model are first extracted and added to the notes section of the requirement decomposition document. Then, the content of the requirement decomposition document is organized in the order of product model section, hardware configuration section, software configuration section, customization attribute section and notes section. Finally, the format of the requirement decomposition document is uniformly processed to generate the final requirement decomposition document.

[0015] The beneficial effects of this invention are: (1) By constructing intelligent interactive templates and product knowledge graphs, the automatic parsing and structural conversion of unstructured requirement texts can be realized, and the real-time verification and feedback of the configuration process can be completed, reducing the manual processing links; (2) Based on the relational reasoning ability of the product knowledge graph, the compatibility detection and conflict resolution of configuration items are automatically completed, and the accurate mapping and automatic aggregation of requirements to baseline tasks are achieved at the same time. (3) Systematically encode and manage the scattered product configuration rules and development tasks to form reusable enterprise knowledge assets and improve the overall efficiency and quality of server customization development. Attached Figure Description

[0016] Figure 1 A flowchart outlining the steps of an AI- and knowledge graph-based intelligent configuration method for server requirements. Figure 2 The following is a flowchart illustrating the specific steps of an AI- and knowledge graph-based intelligent configuration method for server requirements, provided as an example. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1 See Figure 1 This embodiment provides a method for intelligent configuration of server requirements based on AI and knowledge graphs. The method includes the following steps: S1. Construct an enhanced intelligent interaction template and a product knowledge graph. The enhanced intelligent interaction template includes a natural language processing input box and an intelligent recommendation and conflict detection module. The product knowledge graph is centered on server products and includes nodes and edges. Nodes include product models, hardware components, software configurations, customized attributes, and baseline atomic tasks. Edges represent the relationships between nodes. S2. Receive the request text input into the natural language processing input box of the enhanced intelligent interaction template, identify the key entities and intents in the request text through a pre-trained natural language processing model, and automatically fill in the enhanced intelligent interaction template based on the key entities and intents. S3. Real-time acquisition of configuration content in the enhanced intelligent interaction template, compatibility verification, intelligent recommendation and conflict resolution of configuration content in the enhanced intelligent interaction template based on product knowledge graph, and feedback of compatibility verification results, intelligent recommendation content and conflict resolution results to the enhanced intelligent interaction template; S4. Obtain the final configuration content in the enhanced intelligent interaction template, map the final configuration content to a set of nodes in the product knowledge graph, traverse the product knowledge graph to obtain the requirement baseline atomic tasks required to cover the set of nodes, aggregate the document content corresponding to the requirement baseline atomic tasks, and generate a requirement decomposition document.

[0019] In some embodiments, step S1 includes the following sub-steps: S1.1. An enhanced intelligent interactive template is constructed by adding a natural language processing input box and an intelligent recommendation and conflict detection module to the traditional structured template. The traditional structured template includes a product model selection block, a hardware component configuration block, a software configuration block, and a customized attribute block. The natural language processing input box is set in the top area of ​​the traditional structured template, and the intelligent recommendation and conflict detection module is embedded below each block of the traditional structured template. S1.2. Construct a product knowledge graph with server products as the core. The product knowledge graph is stored using resource description framework triples. The nodes of the product knowledge graph include product model, hardware components, software configuration, customized attributes, and requirement baseline atomic tasks. The edges of the product knowledge graph include containment, dependency, mutual exclusion, substitutability, and implementation requirement relationships. Each edge contains a weight value.

[0020] In some embodiments, step S2 includes the following sub-steps: S2.1. Receive the request text input into the natural language processing input box of the enhanced intelligent interactive template; S2.2. Named entity recognition is performed on the requirement text using a pre-trained natural language processing model to extract key entities from the requirement text. The pre-trained natural language processing model adopts the BERT-base model, which contains a set number of Transformer encoders. Each Transformer encoder contains a multi-head self-attention mechanism and a feedforward neural network. The number of heads in the multi-head self-attention mechanism is a set value, and the hidden layer dimension of the feedforward neural network is a set value. S2.3. Use a pre-trained natural language processing model to identify the intent of the request text and obtain the intent corresponding to the request text; S2.4. Automatically fill in the corresponding blocks of the enhanced smart interaction template based on key entities and intents.

[0021] In some embodiments, step S3 includes the following sub-steps: S3.1. Real-time acquisition of the configuration content of each block in the enhanced intelligent interaction template, with an acquisition frequency of a set value; S3.2. Convert the obtained configuration content into a node query statement for the product knowledge graph, and query the nodes in the product knowledge graph that correspond to the configuration content and the edges between the nodes; S3.3. Based on the query results, perform compatibility checks, intelligent recommendations, and conflict resolution on the configuration content. The compatibility check detects mutual exclusion relationships in the configuration content, the intelligent recommendation generates a list of alternative components, and the conflict resolution generates alternative solutions for conflicting content. S3.4. Send the compatibility verification results, intelligent recommended content, and conflict resolution results to the intelligent recommendation and conflict detection module of the enhanced intelligent interaction template for display.

[0022] In some embodiments, step S4 includes the following sub-steps: S4.1. Receive the confirmation signal submitted by the enhanced smart interaction template and obtain the final configuration content of all blocks in the enhanced smart interaction template; S4.2. Match each configuration item in the final configuration content with the nodes in the product knowledge graph to generate a node set consisting of successfully matched nodes; S4.3. Starting from each node in the node set, traverse along the implementation requirement relationships in the product knowledge graph and collect all the requirement baseline atomic task nodes traversed; S4.4. Extract the document content corresponding to each atomic task node of the requirement baseline from the requirement baseline library, and concatenate the extracted document content according to the preset document structure to generate the requirement decomposition document.

[0023] In some embodiments, in step S1.2, the hardware components include a central processing unit, memory, hard disk, independent disk redundant array card, network card, power supply and graphics card, the software configuration includes baseboard management controller parameters, basic input / output system parameters, operating system version and driver version, the customized attributes include nameplate, label, packaging identification and service label, the hardware component node includes interface type, power consumption parameters and size parameters, the software configuration node includes version number and a list of supported hardware, and the customized attribute node includes manufacturing process, size parameters and material parameters.

[0024] In some embodiments, in step S2.2, named entity recognition adopts a sequence labeling method, and the labeling system adopts the BIO labeling system, labeling each character as B-entity type, I-entity type, or O. The output layer of the pre-trained natural language processing model contains two branches. The first branch outputs the entity type probability distribution of each character, and the second branch outputs the intent type probability distribution of the request text. The intent types include performance requirements, functional requirements, customization requirements, and compatibility requirements. Key entities include product model, hardware components, software configuration, and customization attributes.

[0025] In some embodiments, in step S3.3, the compatibility verification includes detecting whether there are two nodes with a mutually exclusive relationship in the configuration content. If they exist, it is determined to be a configuration conflict. The intelligent recommendation includes querying nodes in the product knowledge graph that have a substitutable relationship with the node in the configuration content, sorting them according to the edge weight values, and generating a recommendation list. The conflict resolution includes sending the two conflicting nodes and the corresponding list of substitutable nodes to the intelligent recommendation and conflict detection module of the enhanced intelligent interaction template when a configuration conflict is detected.

[0026] In some embodiments, in step S4.3, the product knowledge graph is traversed using a depth-first traversal algorithm. The starting node of the depth-first traversal algorithm is all nodes in the node set. During the traversal, visited nodes are marked to avoid repeated visits. During the traversal, only the implementation requirement relationship is traversed, and other relationships are not traversed. After the traversal is completed, the collected requirement baseline atomic task nodes are deduplicated to generate the final requirement baseline atomic task list.

[0027] In some embodiments, in step S4, when generating the requirement decomposition document, the notes in the requirement text parsed by the pre-trained natural language processing model are first extracted and added to the notes section of the requirement decomposition document. Then, the content of the requirement decomposition document is organized in the order of product model section, hardware configuration section, software configuration section, customization attribute section and notes section. Finally, the format of the requirement decomposition document is uniformly processed to generate the final requirement decomposition document.

[0028] Example 2 This embodiment provides a specific implementation process for an AI- and knowledge graph-based intelligent configuration method for server requirements. By integrating artificial intelligence and knowledge graph technologies, it achieves automatic conversion of server requirements from unstructured text to structured configuration schemes, while simultaneously generating configuration verification and requirement decomposition documents. This solves the problems of low efficiency and error-proneness in traditional server customization development. Figure 2 As shown, the specific implementation process is as follows: Step 1. Infrastructure Construction: Step 1.1. Enhanced Intelligent Interaction Template Construction: The enhanced intelligent interaction template is a user input interface with added artificial intelligence interactive functions on the basis of the traditional structured template. It is used to receive the user's server requirement input and provide real-time configuration feedback. A natural language processing input box and an intelligent recommendation and conflict detection module are added to the traditional structured template to construct the enhanced intelligent interaction template. The traditional structured template includes a product model selection block, a hardware component configuration block, a software configuration block, and a customization attribute block. Each block contains multiple selectable or fillable configuration items. The natural language processing input box is located at the top area of ​​the traditional structured template to receive unstructured requirement text input by the user. The intelligent recommendation and conflict detection module is embedded below each block of the traditional structured template to display configuration verification results, intelligent recommendation content, and conflict resolution suggestions.

[0029] Step 1.2. Product Knowledge Graph Construction: A product knowledge graph is a database that stores product-related knowledge in a graphical structure, intuitively representing the relationships between the various components of a product. A product knowledge graph centered on the server product is constructed, stored using resource description framework triples. A resource description framework triple is a standardized format for describing resources, consisting of a subject, predicate, and object. In this embodiment, the subject and object correspond to nodes in the product knowledge graph, and the predicate corresponds to edges. Nodes in the product knowledge graph include product model, hardware components, software configuration, customized attributes, and baseline atomic tasks. Edges in the product knowledge graph include contain-in, depend-in, mutually exclusive, substitutable, and fulfillment-requirement relationships. Each edge contains a weight value, which represents the strength or priority of the relationship; a larger value indicates a tighter relationship.

[0030] In some specific implementations, the initial construction of the product knowledge graph included 120 product model nodes, 860 hardware component nodes, 320 software configuration nodes, 150 customized attribute nodes, and 2100 requirement baseline atomic task nodes, with a total of 18,700 edges between nodes and 18,700 corresponding resource description framework triples. The construction process employed a batch import method, first importing all node data, then importing edge data sequentially according to the relationships between nodes. A data consistency check was performed every 1000 imported triples to ensure the accuracy of the correspondence between nodes and edges. This scale of knowledge graph can cover all configuration options and corresponding development tasks of mainstream server products, solving the problem of limited coverage and inability to meet complex configuration requirements of traditional knowledge graphs. Furthermore, the batch import and segmented verification methods ensured the accuracy and stability of the knowledge graph construction process.

[0031] In some specific implementations, the product knowledge graph is maintained using an incremental update mechanism to address the inconsistency between the knowledge graph content and actual product information during product iteration. The specific implementation steps of the incremental update mechanism are as follows: First, a product information change database is established to record changes to all product models, hardware components, software configurations, and customized attributes, including added, modified, and deleted content. Then, at set time intervals, out-of-sync change information is extracted from the product information change database. Next, the extracted change information is converted into resource description framework triplet format. Finally, the converted triplet data is merged into the product knowledge graph, while simultaneously updating the edges and edge weights between relevant nodes. For deleted content, the corresponding nodes and edges are marked as invalid, and historical records are retained for traceability. For modified content, the original triplet data is deleted first, and then the new triplet data is inserted. For added content, the corresponding triplet data is directly inserted. This incremental update method reduces the workload of data processing, improves the efficiency of knowledge graph updates, and ensures that the knowledge graph always remains consistent with the latest product information.

[0032] In some embodiments, the product knowledge graph can also be stored using an attribute graph model. The attribute graph model can store attribute information on nodes and edges, which can more flexibly represent complex product knowledge relationships and is suitable for scenarios with complex product configuration rules.

[0033] Step 1.3. Product Knowledge Graph Node Attribute Definition: Hardware components include a central processing unit (CPU), memory, hard drive, redundant RAID card, network card, power supply, and graphics card. Hardware component nodes include interface type, power consumption parameters, and size parameters, which are used to determine compatibility between different hardware components. Software configuration includes baseboard management controller parameters, basic input / output system parameters, operating system version, and driver version. Software configuration nodes include version numbers and a list of supported hardware, used to determine software and hardware compatibility. Customization attributes include nameplates, labels, packaging markings, and service labels. Customization attribute nodes include manufacturing process, size parameters, and material parameters, used to guide the customized production process. Requirements baseline atomic task nodes include task descriptions, task responsible persons, and task document links, used to generate requirements decomposition documents.

[0034] Step 2. Requirements Analysis and Template Filling: Step 2.1. Receiving Required Text: Receive the required text input into the natural language processing input box of the enhanced intelligent interaction template. The required text is a natural language description of the user's requirements for server functions, performance, and customization.

[0035] Step 2.2. Named Entity Recognition: Named entity recognition is a fundamental task in natural language processing (NLP) used to extract entities with specific meanings from text. A pre-trained NLP model is used to perform named entity recognition on the required text to extract key entities. The pre-trained NLP model employs the BERT-base model, a pre-trained language model based on the Transformer architecture capable of capturing contextual semantic information in text. The BERT-base model contains a set number of Transformer encoder layers, each containing a multi-head self-attention mechanism and a feedforward neural network. The multi-head self-attention mechanism can simultaneously focus on information from different locations in the text, capturing long-distance dependencies between words; the number of heads in the multi-head self-attention mechanism is a set value. The feedforward neural network is used to perform a non-linear transformation on the attention output, extracting higher-level semantic features; the hidden layer dimension of the feedforward neural network is a set value.

[0036] In some specific implementations, the fine-tuning process of the pre-trained natural language processing model uses 50,000 annotated server requirement texts as the training set and 10,000 annotated data points as the validation set. During training, the initial learning rate is set to 2e-5, the batch size to 32, and the number of training epochs to 10. Evaluation is performed on the validation set after each training epoch, and training is stopped early when the validation set loss no longer decreases after three consecutive epochs. After training, the model achieves a named entity recognition accuracy of 92.3% and an intent recognition accuracy of 89.7% on the test set, meeting the accuracy requirements for parsing requirements in practical applications. By using large-scale annotated data for fine-tuning and employing an early stopping mechanism, the problems of overfitting and insufficient generalization ability are solved, ensuring that the model maintains stable recognition performance when faced with requirement texts of different styles and expressions.

[0037] The training process of the pre-trained natural language processing model is as follows: First, collect demand text data from the server domain, label the data, and mark the key entities and intent types in the text; then, use the labeled data to fine-tune the BERT-base model. During the fine-tuning process, the cross-entropy loss function is used as the loss function, and the adaptive moment estimation optimizer is used to update the parameters. The number of training epochs and the batch size are set values. Finally, save the fine-tuned model parameters for subsequent demand parsing.

[0038] Named entity recognition employs sequence labeling, using the BIO (Browser-Integer Identification) labeling system. BIO is a commonly used sequence labeling method that labels each character as B (entity type), I (entity type), or O, where B represents the start of an entity, I represents its internal position, and O represents a non-entity part. The output layer of the pre-trained natural language processing model contains two branches. The first branch outputs the entity type probability distribution for each character. Based on this probability distribution, the labeling result for each character is determined, thereby extracting key entities from the requirement text. These key entities include product model, hardware components, software configuration, and customization attributes.

[0039] In some specific implementations, the pre-trained natural language processing model employs a domain-adaptive pre-training method for optimization, addressing the issue of low accuracy in recognizing industry-specific terminology in server requirement texts from different industries. The specific implementation steps of the domain-adaptive pre-training method are as follows: First, server requirement text data from multiple industries are collected to construct an industry-specific corpus. Then, the BERT-base model is further pre-trained using this industry-specific corpus. During this pre-training process, a masked language model task is used as the training task, randomly masking a certain proportion of characters in the corpus, allowing the model to predict the masked characters. Next, the model is fine-tuned using labeled general server requirement text data. Finally, the optimized model parameters are saved for subsequent requirement parsing. In this way, the model can learn the server requirement terminology and expressions of different industries, improving its entity recognition and intent recognition capabilities in cross-industry requirement texts and reducing requirement parsing errors caused by differences in industry terminology.

[0040] In some embodiments, named entity recognition can also employ a combination of a conditional random field model and a BERT-base model. The conditional random field model is used to optimize the sequence labeling of the output of the BERT-base model, taking into account the labeling dependencies between adjacent characters, thereby improving the accuracy of entity recognition.

[0041] Step 2.3. Intent Recognition: Intent recognition is a task in Natural Language Processing (NLP) used to understand the purpose and needs of user input text. A pre-trained NLP model is used to recognize the intent of the request text, obtaining the intent corresponding to the text. The second branch of the pre-trained NLP model outputs the probability distribution of the intent type of the request text. Based on the probability distribution, the intent type of the request text is determined. Intent types include performance requirements, functional requirements, customization requirements, and compatibility requirements. Performance requirements refer to the user's requirements for server processing power, storage capacity, and network transmission capabilities; functional requirements refer to the user's requirements for specific server functions; customization requirements refer to the user's personalized requirements for server appearance, logos, etc.; compatibility requirements refer to the user's requirements for compatibility between the server and other devices or software.

[0042] Step 2.4. Template Auto-Fill: Based on the extracted key entities and identified intents, automatically fill in the corresponding blocks of the enhanced intelligent interaction template. For the extracted product model entity, automatically select the option corresponding to the product model selection block in the enhanced intelligent interaction template; for the extracted hardware components, software configurations, and customized attribute entities, automatically fill in or select the configuration items in the corresponding blocks of the enhanced intelligent interaction template; for the identified intents, recommend corresponding configuration options based on the intent type, for example, when performance requirements are identified, recommend high-performance hardware component configurations.

[0043] Step 3. Configure verification and conflict handling: Step 3.1. Configuration Content Acquisition: Acquire the configuration content of each block in the enhanced intelligent interaction template in real time. The acquisition frequency is set to ensure timely response to user configuration operations.

[0044] Step 3.2. Knowledge Graph Query: Convert the obtained configuration content into node query statements for the product knowledge graph, querying the nodes in the product knowledge graph that correspond to the configuration content and the edges between those nodes. The node query statements are used to retrieve all information related to the configuration item from the product knowledge graph, including node attributes and relationships between nodes.

[0045] Step 3.3. Compatibility Verification: Based on the query results, perform compatibility verification on the configuration content. Compatibility verification checks for mutual exclusion relationships within the configuration content. A mutual exclusion relationship means that two nodes cannot coexist in the same configuration scheme. For example, some server models do not support specific combinations of independent disk redundant array cards and network cards. The system checks whether two nodes with a mutual exclusion relationship exist in the configuration content; if they do, it is determined to be a configuration conflict.

[0046] In some specific implementations, compatibility verification also includes detecting dependencies in the configuration content to address the issue of missing necessary dependencies during configuration. A dependency relationship indicates that the existence of one node is contingent upon the existence of another node; for example, configuring a redundant RAID card requires simultaneously configuring the corresponding hard drive. The specific steps for dependency detection are as follows: First, extract all nodes corresponding to the configuration content from the query results; then, traverse each node and query all preceding nodes in the product knowledge graph that have dependencies on that node; next, check if these preceding nodes exist in the current configuration content; if a preceding node is not in the configuration content, it is determined as a missing configuration, and a corresponding missing information is generated; finally, the missing information is sent to the intelligent recommendation and conflict detection module of the enhanced intelligent interaction template for display, while simultaneously recommending corresponding preceding node configuration options to guide the user to complete the full configuration.

[0047] Step 3.4. Intelligent Recommendation: Based on the query results, intelligent recommendations are made for the configuration content, generating a list of replaceable components. A replaceable relationship indicates that two nodes are functionally interchangeable; for example, different models of CPUs can be substituted for each other if their performance is similar. Based on the nodes in the configuration content, the product knowledge graph is queried for nodes with replaceable relationships. These nodes are then sorted according to their edge weights to generate a recommendation list, with nodes having higher weights appearing earlier in the list.

[0048] Step 3.5. Conflict Resolution: Based on the query results, conflict resolution is performed on the configuration content, generating alternative solutions for the conflicting content. When a configuration conflict is detected, the two conflicting nodes and their corresponding list of alternative nodes are sent to the intelligent recommendation and conflict detection module of the enhanced intelligent interaction template, allowing the user to choose a suitable alternative.

[0049] Step 3.6. Result Feedback: The compatibility verification results, intelligent recommended content, and conflict resolution results are sent to the intelligent recommendation and conflict detection module of the enhanced intelligent interaction template for display. The compatibility verification results are displayed as a prompt message, informing the user whether there is a configuration conflict; the intelligent recommended content is displayed in a list format for the user to choose from; the conflict resolution results are displayed as alternative solutions to help the user resolve configuration conflicts.

[0050] In some implementations, the system supports up to 200 users simultaneously performing online configuration operations. The knowledge graph query and compatibility verification process for a single configuration operation takes no more than 200 milliseconds, and the generation of intelligent recommended content takes no more than 150 milliseconds. The total latency from when the user completes the configuration selection to when the result is displayed on the interface does not exceed 500 milliseconds. To achieve this performance, the system uses a Redis caching mechanism to cache frequently used product knowledge graph nodes and relationship data, achieving a cache hit rate of over 85%. It also employs multi-threaded concurrent processing to handle configuration verification requests from multiple users simultaneously. This performance ensures a smooth user experience during configuration, solving the problems of slow response and long user waiting times in traditional configuration systems, and improving the efficiency of the configuration process.

[0051] In some embodiments, intelligent recommendation can also combine users' historical configuration data to generate personalized recommendation lists through collaborative filtering algorithms, thereby improving the relevance of the recommendations. Collaborative filtering algorithms analyze users' historical configuration behavior to identify other users with similar configuration preferences to the current user, and then recommend these users' configuration options to the current user.

[0052] Step 4. Baseline Aggregation and Document Generation: Step 4.1. Final Configuration Content Acquisition: Receive the confirmation signal from the enhanced smart interaction template submission and obtain the final configuration content of all blocks in the enhanced smart interaction template. The confirmation signal is the signal submitted after the user completes all configuration operations, indicating that the configuration content has been finalized.

[0053] Step 4.2. Node Matching and Set Generation: Each configuration item in the final configuration content is matched with nodes in the product knowledge graph to generate a node set consisting of successfully matched nodes. During the matching process, the name and attributes of the configuration item are compared with the nodes in the product knowledge graph to find the most matching node.

[0054] Step 4.3. Knowledge Graph Traversal and Task Collection: Starting from each node in the node set, traverse along the implementation requirement relationships in the product knowledge graph, collecting all traversed requirement baseline atomic task nodes. Implementation requirement relationships indicate that the implementation of one node depends on the task represented by another node; for example, configuring a specific model of CPU requires executing the corresponding BIOS parameter setting task.

[0055] The product knowledge graph is traversed using a depth-first search (DFS) algorithm. DFS is a graph traversal algorithm that starts from a starting node and visits nodes as deeply as possible in the graph until no further access is possible, then backtracks to the last unvisited node. The specific implementation steps of the DFS algorithm are as follows: First, initialize a set of visited nodes and a set of task nodes; then, for each node in the node set, if the node is not in the visited node set, add it to the visited node set; next, query all child nodes of the node that are connected by the required relationships; for each child node, if the child node is not in the visited node set, recursively call the DFS algorithm to process the child node; when the recursion returns, add the child node to the task node set; finally, deduplicate the nodes in the task node set to generate the final baseline atomic task list.

[0056] In some specific implementations, after collecting all baseline requirement atomic task nodes, the baseline requirement atomic tasks are prioritized to resolve the issue of disordered task execution order after requirement decomposition. The specific implementation steps of priority prioritization are as follows: First, based on the dependency relationships between nodes in the product knowledge graph, a dependency graph of the baseline requirement atomic tasks is constructed. Nodes in the dependency graph represent baseline requirement atomic tasks, and edges represent dependencies between tasks. Then, the in-degree of each task is calculated based on the dependency graph, where the in-degree represents the number of predecessor tasks that the task depends on. Next, tasks with an in-degree of zero are added to the ready task queue. Then, tasks are retrieved from the ready task queue, assigned the highest priority, and the in-degree of all subsequent tasks of that task is decremented by one. If the in-degree of a subsequent task becomes zero, it is added to the ready task queue. This process is repeated until all tasks are assigned priority. In this way, it is ensured that tasks are executed in the correct order, avoiding development delays caused by incorrect task execution order.

[0057] In some embodiments, knowledge graph traversal can also employ a breadth-first traversal algorithm. The breadth-first traversal algorithm starts from the starting node, visits all adjacent nodes first, and then visits the adjacent nodes of these adjacent nodes in turn. This can more comprehensively cover the nodes in the graph and is suitable for scenarios that require obtaining all relevant tasks.

[0058] Step 4.4. Document Content Extraction and Assembly: Extract the document content corresponding to each requirement baseline atomic task node from the requirement baseline library. Assemble the extracted document content according to the preset document structure to generate the requirement decomposition document. The requirement baseline library is a database storing the document content corresponding to all requirement baseline atomic tasks. The document content includes task descriptions, execution steps, acceptance criteria, and other information. The preset document structure includes a product model section, a hardware configuration section, a software configuration section, a customization attribute section, and a remarks section. Each section contains the corresponding requirement baseline atomic task document content.

[0059] Step 4.5. Requirements Decomposition Document Generation: When generating the requirements decomposition document, firstly, extract the notes from the requirements text parsed by the pre-trained natural language processing model and add them to the notes section of the requirements decomposition document. These notes are special requirements or explanatory information entered by the user in the requirements text. Then, organize the content of the requirements decomposition document according to the following order: product model section, hardware configuration section, software configuration section, customization attribute section, and notes section, ensuring a clear document structure and logical coherence. Finally, standardize the format of the requirements decomposition document, including font, font size, paragraph spacing, and heading levels, to generate the final requirements decomposition document.

[0060] In some specific implementations, for complex server configuration schemes containing more than 50 configuration items, the knowledge graph traversal and requirement baseline atomic task aggregation process takes no more than 3 seconds. The generated requirement decomposition document contains more than 8 chapters and more than 120 task descriptions, with a total word count exceeding 8,000 words. During the generation process, the system automatically merges duplicate task descriptions and standardizes the task execution steps to ensure the consistency and standardization of the document content. Compared with traditional manual writing of requirement decomposition documents, this automated generation method can reduce document writing time from hours to seconds, solving the problems of excessively long writing time, omissions, and errors in requirement decomposition documents under complex configurations, and significantly improving the efficiency and quality of requirement decomposition.

[0061] In some implementations, after generating the final requirement decomposition document (DFD), version management is implemented to address the issues of confusion and difficulty in tracing multiple versions of requirement documents. The specific steps for version management are as follows: First, a unique version identifier is assigned to each generated DFD, containing both the document number and the version number. Then, the generation time, the person who generated the document, and the corresponding configuration information are recorded. Next, the document and its version information are stored in a document management database. When the DFD needs to be modified, a new version identifier is generated, recording the changes and the modification time, while retaining the old version's document and information. Finally, a document version query and comparison function is provided, allowing users to query any version of the document based on its version identifier and compare the differences between different versions. This version management method clearly records the change history of the requirement documents, facilitating subsequent tracing and auditing, and preventing development errors caused by version confusion.

[0062] In some embodiments, the requirements decomposition document can also be exported in multiple formats to facilitate transmission and use between different systems and meet the needs of different users.

[0063] This embodiment achieves intelligent parsing and configuration of server requirements by constructing an enhanced intelligent interactive template and a product knowledge graph. It can automatically process unstructured requirement text, converting it into a structured configuration scheme, reducing manual processing workload. Simultaneously, through real-time configuration verification and conflict resolution based on the product knowledge graph, configuration conflicts can be detected and resolved promptly during the requirement configuration process, reducing the possibility of subsequent rework. Furthermore, by automatically aggregating requirement baseline atomic tasks through knowledge graph traversal, requirement decomposition documents can be quickly generated, improving the efficiency of requirement decomposition. This embodiment combines artificial intelligence technology with knowledge graph technology to achieve intelligent and automated server requirement configuration, effectively improving the efficiency and quality of customized server development. It also systematically stores and manages scattered product knowledge, forming reusable enterprise knowledge assets that support subsequent product development and optimization. The technical solution of this embodiment also has good scalability, allowing for timely updates to the product knowledge graph based on product iterations, adapting to ever-changing market demands.

[0064] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for intelligent configuration of server requirements based on AI and knowledge graphs, characterized in that, Includes the following steps: S1. Construct an enhanced intelligent interaction template and a product knowledge graph. The enhanced intelligent interaction template includes a natural language processing input box and an intelligent recommendation and conflict detection module. The product knowledge graph is centered on server products and includes nodes and edges. Nodes include product models, hardware components, software configurations, customized attributes, and baseline atomic tasks. Edges represent the relationships between nodes. S2. Receive the request text input into the natural language processing input box of the enhanced intelligent interaction template, identify the key entities and intents in the request text through a pre-trained natural language processing model, and automatically fill in the enhanced intelligent interaction template based on the key entities and intents. S3. Real-time acquisition of configuration content in the enhanced intelligent interaction template, compatibility verification, intelligent recommendation and conflict resolution of configuration content in the enhanced intelligent interaction template based on product knowledge graph, and feedback of compatibility verification results, intelligent recommendation content and conflict resolution results to the enhanced intelligent interaction template; S4. Obtain the final configuration content in the enhanced intelligent interaction template, map the final configuration content to a set of nodes in the product knowledge graph, traverse the product knowledge graph to obtain the requirement baseline atomic tasks required to cover the set of nodes, aggregate the document content corresponding to the requirement baseline atomic tasks, and generate a requirement decomposition document.

2. The method according to claim 1, characterized in that, Step S1 includes the following sub-steps: S1.

1. An enhanced intelligent interactive template is constructed by adding a natural language processing input box and an intelligent recommendation and conflict detection module to the traditional structured template. The traditional structured template includes a product model selection block, a hardware component configuration block, a software configuration block, and a customized attribute block. The natural language processing input box is set in the top area of ​​the traditional structured template, and the intelligent recommendation and conflict detection module is embedded below each block of the traditional structured template. S1.

2. Construct a product knowledge graph with server products as the core. The product knowledge graph is stored using resource description framework triples. The nodes of the product knowledge graph include product model, hardware components, software configuration, customized attributes, and requirement baseline atomic tasks. The edges of the product knowledge graph include containment, dependency, mutual exclusion, substitutability, and implementation requirement relationships. Each edge contains a weight value.

3. The method according to claim 1, characterized in that, Step S2 includes the following sub-steps: S2.

1. Receive the request text input into the natural language processing input box of the enhanced intelligent interactive template; S2.

2. Named entity recognition is performed on the requirement text using a pre-trained natural language processing model to extract key entities from the requirement text. The pre-trained natural language processing model adopts the BERT-base model, which contains a set number of Transformer encoders. Each Transformer encoder contains a multi-head self-attention mechanism and a feedforward neural network. The number of heads in the multi-head self-attention mechanism is a set value, and the hidden layer dimension of the feedforward neural network is a set value. S2.

3. Use a pre-trained natural language processing model to identify the intent of the request text and obtain the intent corresponding to the request text; S2.

4. Automatically populate the corresponding blocks of the enhanced smart interaction template based on key entities and intents.

4. The method according to claim 1, characterized in that, Step S3 includes the following sub-steps: S3.

1. Real-time acquisition of the configuration content of each block in the enhanced intelligent interaction template, with an acquisition frequency of a set value; S3.

2. Convert the obtained configuration content into a node query statement for the product knowledge graph, and query the nodes in the product knowledge graph that correspond to the configuration content and the edges between the nodes; S3.

3. Based on the query results, perform compatibility checks, intelligent recommendations, and conflict resolution on the configuration content. The compatibility check detects mutual exclusion relationships in the configuration content, the intelligent recommendation generates a list of alternative components, and the conflict resolution generates alternative solutions for conflicting content. S3.

4. Send the compatibility verification results, intelligent recommended content, and conflict resolution results to the intelligent recommendation and conflict detection module of the enhanced intelligent interaction template for display.

5. The method according to claim 1, characterized in that, Step S4 includes the following sub-steps: S4.

1. Receive the confirmation signal submitted by the enhanced smart interaction template and obtain the final configuration content of all blocks in the enhanced smart interaction template; S4.

2. Match each configuration item in the final configuration content with the nodes in the product knowledge graph to generate a node set consisting of successfully matched nodes; S4.

3. Starting from each node in the node set, traverse along the implementation requirement relationships in the product knowledge graph and collect all the requirement baseline atomic task nodes traversed; S4.

4. Extract the document content corresponding to each atomic task node of the requirement baseline from the requirement baseline library, and concatenate the extracted document content according to the preset document structure to generate the requirement decomposition document.

6. The method according to claim 2, characterized in that, In step S1.2, the hardware components include a central processing unit, memory, hard disk, independent disk redundant array card, network card, power supply and graphics card, the software configuration includes baseboard management controller parameters, basic input / output system parameters, operating system version and driver version, the customized attributes include nameplate, label, packaging identification and service label, the hardware component node includes interface type, power consumption parameters and size parameters, the software configuration node includes version number and a list of supported hardware, and the customized attribute node includes manufacturing process, size parameters and material parameters.

7. The method according to claim 3, characterized in that, In step S2.2, named entity recognition uses a sequence labeling method, and the labeling system adopts the BIO labeling system, labeling each character as B-entity type, I-entity type, or O. The output layer of the pre-trained natural language processing model contains two branches. The first branch outputs the entity type probability distribution of each character, and the second branch outputs the intent type probability distribution of the request text. The intent types include performance requirements, functional requirements, customization requirements, and compatibility requirements. Key entities include product model, hardware components, software configuration, and customization attributes.

8. The method according to claim 4, characterized in that, In step S3.3, compatibility verification includes detecting whether there are two nodes with a mutually exclusive relationship in the configuration content. If they exist, it is determined to be a configuration conflict. Intelligent recommendation includes querying nodes with a substitutable relationship in the product knowledge graph based on the nodes in the configuration content, sorting them according to the edge weight values, and generating a recommendation list. Conflict resolution includes sending the two conflicting nodes and the corresponding list of substitutable nodes to the intelligent recommendation and conflict detection module of the enhanced intelligent interaction template when a configuration conflict is detected.

9. The method according to claim 5, characterized in that, In step S4.3, the product knowledge graph is traversed using a depth-first traversal algorithm. The starting node of the depth-first traversal algorithm is all nodes in the node set. During the traversal, visited nodes are marked to avoid repeated visits. The traversal is only performed along the implementation requirement relationship and not along other relationships. After the traversal is completed, the collected requirement baseline atomic task nodes are deduplicated to generate the final requirement baseline atomic task list.

10. The method according to claim 1, characterized in that, In step S4, when generating the requirement decomposition document, the notes in the requirement text parsed by the pre-trained natural language processing model are first extracted and added to the notes section of the requirement decomposition document. Then, the content of the requirement decomposition document is organized in the order of product model section, hardware configuration section, software configuration section, customization attribute section and notes section. Finally, the format of the requirement decomposition document is uniformly processed to generate the final requirement decomposition document.

Images