Network operation and maintenance analysis method and device and storage medium
Network operation and maintenance analysis is carried out through a large model of Transformer network architecture, which solves the problem of relying on experience and intuition in the existing technology, improves the accuracy of analysis and the generalization ability of the model, and simplifies the data processing process.
Patent Information
- Application Number
- CN202411404576.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-10
- Publication Date
- 2024-11-08
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing network operation and maintenance analysis methods rely on experience and intuition, lack stability and repeatability. Traditional machine learning algorithms require a large amount of labeled data and cumbersome feature engineering, which reduces the generalization ability of the model.
Using a large model of Transformer network architecture for network operation and maintenance analysis, pre-processing and splitting by obtaining multimodal problem information, using intent classification to generate SQL query statements, vector database query or API interface query, and combining multimodal feature vectors for Query and generate network operation and maintenance analysis results.
It improves the accuracy of retrieval and analysis accuracy, reduces the dependence of manual experience, enhances the generalization ability of the model, and simplifies the data processing process.
Smart Images

Figure CN118916618A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a network operation and maintenance analysis method, device and storage medium. Background Art
[0002] In the existing network operation and maintenance analysis field, the guidance and suggestions for network operation and maintenance analysis mainly rely on traditional expert experience and symbolic rules. This method usually requires operation and maintenance personnel to use their own experience and intuition, or to consult a large amount of technical documents, historical data and solutions to diagnose and deal with various operation and maintenance problems. However, this method that relies on experience and intuition often has a lot of randomness and lacks stability and repeatability.
[0003] In addition, traditional machine learning algorithms can be used for network operation and maintenance analysis, but they also face a series of challenges in practical applications. First, these algorithms often require a large amount of labeled data for training, and the acquisition and processing of labeled data is often a time-consuming and labor-intensive task. Second, traditional machine learning models usually require tedious feature engineering, which not only increases the complexity of the model, but also reduces the generalization ability of the model. Summary of the invention
[0004] The present invention provides a network operation and maintenance analysis method, device and storage medium, aiming to solve at least one of the above-mentioned technical problems.
[0005] The present invention provides a network operation and maintenance analysis method, comprising: Obtain multimodal question information from target users; Preprocessing and splitting the multimodal question information to obtain at least one target question information; Using the target large model to perform intent classification on at least one target question information, and obtaining an intent recognition result corresponding to the at least one target question information; According to the intention recognition results, query the target question information to obtain the query results; Based on the query results corresponding to different intent recognition results, the network operation and maintenance analysis results are obtained by using the target large model reasoning; Among them, the target large model is the large model corresponding to the Transformer network architecture.
[0006] According to a network operation and maintenance analysis method provided by the present invention, according to the intention recognition result, the target problem information is queried to obtain the query result, including: If the intent recognition result is an SQL database query intent, then a SQL query statement is generated based on the target question information, and a query result is obtained by querying a preset SQL database based on the SQL query statement; If the intent recognition result is a vector database query intent, a multimodal feature vector is extracted based on the target question information, and a query result is obtained by querying in a preset vector database based on the multimodal feature vector; If the intent recognition result is an API interface query intent, an API interface query request is generated based on the target question information, and a query result is obtained based on the API interface query request.
[0007] According to a network operation and maintenance analysis method provided by the present invention, based on target problem information, an SQL query statement is generated, including: Determine the similarity between the target question information and each historical question in the preset knowledge base; Use the historical questions with the highest similarity as question samples; The target problem information, problem samples and SQL statements associated with the problem samples are input into the target large model to obtain SQL query statements.
[0008] According to a network operation and maintenance analysis method provided by the present invention, the multimodal feature vector includes a text feature vector, an image feature vector and a speech feature vector; Based on the target question information, a multimodal feature vector is extracted, and based on the multimodal feature vector, a query result is obtained by querying in a preset vector database, including: When the target question information includes text information, the text information in the target question information is segmented; based on the text information after the segmentation, a word vector, a text vector and a position vector are extracted; the word vector, the text vector and the position vector are added together to perform feature extraction to obtain a text feature vector; When the target problem information includes image information, feature extraction is performed on the image information to obtain image feature information; the image feature information is transformed to obtain an image feature vector; When the target question information includes voice information, voice recognition is performed on the voice information to obtain voice feature information; the voice feature information is transformed to obtain a voice feature vector; wherein the vector dimensions of the voice feature vector, the image feature vector and the text feature vector are the same; Fusing multi-modal feature vectors to obtain a fused feature vector; Based on the fused feature vector, the query result is obtained by querying the vector database.
[0009] According to a network operation and maintenance analysis method provided by the present invention, an API interface query request is generated based on target problem information, and query results are obtained based on the API interface query request, including: Based on the target question information, analyze and obtain entity information and keyword information; Generate parameter information based on entity information and keyword information; Based on the parameter information, an API interface query request is generated, and based on the API interface query request, an external API interface query is called to obtain the query result.
[0010] According to a network operation and maintenance analysis method provided by the present invention, the target big model is used to perform intent classification on at least one target problem information to obtain an intent recognition result corresponding to the at least one target problem information, including: For any target question information: Use the target large model to perform entity recognition and semantic analysis on the target question information to obtain entity recognition results and semantic analysis results; Based on the entity recognition results and semantic analysis results, determine the intent recognition result corresponding to the target question information, where the intent recognition result includes SQL database query intent, vector database query intent and / or API interface query intent.
[0011] According to a network operation and maintenance analysis method provided by the present invention, multimodal problem information is preprocessed and split to obtain at least one target problem information, including: Performing a grammar check on the multimodal question information, and correcting the multimodal question information according to the grammar check result; The corrected multimodal question information is split into question pieces to obtain at least one target question information.
[0012] According to a network operation and maintenance analysis method provided by the present invention, based on query results corresponding to different intent recognition results, a network operation and maintenance analysis result is obtained by using a target large model inference, including: Integrate the query results corresponding to different intent recognition results to obtain an integrated result; The integration results are packaged into a preset prompt word format, and the prompt word is input into the target large model to infer the network operation and maintenance analysis results. The present invention also provides a network operation and maintenance analysis device, comprising: The acquisition module is used to obtain the multimodal question information of the target user; A processing module, used for preprocessing and splitting the multimodal question information to obtain at least one target question information; A classification module, used to classify the intent of at least one target question information using the target large model, and obtain an intent recognition result corresponding to the at least one target question information; The query module is used to query the target question information according to the intention recognition result and obtain the query result; The reasoning module is used to integrate the query results corresponding to different intent recognition results, and to obtain the network operation and maintenance analysis results based on the integration results and the target large model reasoning; Among them, the target large model is the large model corresponding to the Transformer network architecture.
[0013] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, any of the above-mentioned network operation and maintenance analysis methods is implemented.
[0014] The present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, any of the network operation and maintenance analysis methods described above is implemented.
[0015] The network operation and maintenance analysis method, device and storage medium provided by the present invention include: obtaining multimodal problem information of the target user; preprocessing and splitting the multimodal problem information to obtain at least one target problem information; using the target big model to classify the intent of at least one target problem information to obtain the intent recognition result corresponding to at least one target problem information; querying the target problem information according to the intent recognition result to obtain the query result; based on the query results corresponding to different intent recognition results, using the target big model to infer the network operation and maintenance analysis result. The present invention effectively corrects the ambiguity and ambiguity of user input by adding steps such as problem preprocessing and problem splitting in the retrieval process, thereby improving the accuracy of the retrieval; and using the big model to quickly and accurately understand the user's intent, querying the target problem information based on the intent recognition result to obtain the query result, and then combining the query results corresponding to the different intent recognition results to infer the final analysis result, without the need for manual analysis based on experience, thereby improving the accuracy of the big model analysis. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced one by one below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0017] Figure 1 It is a flow chart of the network operation and maintenance analysis method provided by the present invention; Figure 2 This is a schematic diagram of SQL statement generation provided by an embodiment of the present invention; Figure 3 It is a structural schematic diagram of the network operation and maintenance analysis device provided by the present invention; Figure 4 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION
[0018] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0019] The terms used in one or more embodiments of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit one or more embodiments of the present invention. The singular forms of "a", "" and "the" used in one or more embodiments of the present invention are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used in one or more embodiments of the present invention refers to and includes any or all possible combinations of one or more associated listed items.
[0020] Figure 1 FIG. 1 is a flow chart of the network operation and maintenance analysis method provided by the present invention. Figure 1 As shown, the network operation and maintenance analysis method includes: Step S11, obtaining multimodal question information of the target user; It should be noted that the multimodal problem information includes voice information, text information, image information, file information, etc., and the user can input the multimodal problem information at the system front end. In this embodiment, the multimodal problem information is for problems in the field of network operation and maintenance analysis, for example, the error log of a service shows that a specific error code appears frequently; the system performance data shows that the CPU and memory usage are abnormally high, and other related information.
[0021] Step S12, preprocessing and splitting the multimodal question information to obtain at least one target question information; Specifically, after the target user inputs the question, the target large model first rewrites the multimodal question information and performs other preprocessing, such as grammar checking on the multimodal question information, and correcting the multimodal question information based on the grammar checking results, thereby correcting grammatical errors, eliminating ambiguity, and ensuring the clarity and accuracy of the question. This process uses the model's contextual understanding ability to simulate the effect of manual refinement and achieve preliminary optimization of the problem.
[0022] Furthermore, for complex questions, the large model can split the corrected multimodal question information into multiple sub-questions and obtain at least one target question information, that is, split the multimodal question information into multiple sub-questions and send them to different retrieval channels for recall. For example, if the user's question is "Who has more mobile base stations, Guangzhou or Shenzhen?", this question can be broken down into two more specific and simpler sub-questions: How many mobile base stations are there in Guangzhou, and how many mobile base stations are there in Shenzhen. These sub-queries will be executed in parallel, and then the retrieved context will be merged into a whole, and finally input into the large model to generate the answer. By performing split query and other processing on the questions input by the user, the purpose of improving the retrieval effect can be achieved.
[0023] Step S13, using the target large model to perform intent classification on at least one target question information, and obtain an intent recognition result corresponding to the at least one target question information; It should be noted that the intent recognition results include SQL database query intent, vector database query intent and / or API interface query intent. SQL database query intent refers to the task intent when the problem needs to be queried in the SQL database, vector database query intent refers to the task intent when the problem needs to be queried in the vector database, and API interface query intent refers to the API interface query intent when the problem needs to call an external interface for query. The target large model is the large model corresponding to the Transformer network architecture. The Transformer network architecture includes an encoder. In addition, according to the specific scenario tasks (such as classification, retrieval, etc.), the corresponding task-specific layers are added after the Transformer encoder to obtain the corresponding result output. For example, in the intent classification scenario, the encoder can connect the fully connected layer and the softmax activation function to predict the intent category.
[0024] Specifically, for any target question information: use the target big model to perform entity recognition and semantic analysis on the target question information to obtain entity recognition results and semantic analysis results; then determine the intent recognition results corresponding to the target question information based on the entity recognition results and semantic analysis results. Thus, through the accurate semantic understanding and intent recognition of multimodal questions by the big model, the system can accurately determine the task intent type and automatically select the most appropriate retrieval and processing strategy.
[0025] Step S14, querying the target question information according to the intention recognition result to obtain the query result; In one embodiment, if the intent recognition result is an SQL database query intent, the keywords and entity information in the target question information are extracted, and based on the keywords and entity information, query prompt words that meet the requirements of the target large model are generated, so that the target question information is converted into an SQL query statement using the target large model according to the query prompt words. It should be noted that in the SQL statement generation scenario, the encoder in the target large model can be connected to the decoder to use the decoder structure to generate an output SQL query statement. Further, the SQL query statement is sent to the SQL database to perform the query operation and obtain the query result. In more detail: the similarity between the target question information and each historical question in the preset knowledge base is calculated, and then the question sample is selected from each historical question according to the similarity of the question; then based on the target question information, the question sample and the SQL statement associated with the question sample, the Prompt query information is constructed, and then the Prompt query information is input into the target large model, so that the target large model can accurately generate the SQL query statement of the target question information according to the learned historical questions and SQL statements.
[0026] In one embodiment, if the intent recognition result is a vector database query intent, a multimodal feature vector is extracted using multimodal information such as text, images, and voice in the question, and then based on the multimodal feature vector, a search is performed in a preset vector database to obtain a query result. The specific query process corresponding to the vector database query intent is specifically described in the following embodiments and will not be repeated here.
[0027] In one embodiment, if the intent recognition result is an API interface query intent, the keywords and entity information in the question are analyzed to determine the API to be accessed and its query parameter information. Based on the parameter information, an API interface query request is generated, and based on the API interface query request, an external API interface query is called to obtain the query result.
[0028] Step S15, based on the query results corresponding to the different intent recognition results, the network operation and maintenance analysis results are obtained by using the target large model reasoning; Specifically, the query results corresponding to different intent recognition results are integrated to obtain an integrated result; the integrated result is then encapsulated into a prompt word in a preset format, and the prompt word is input into the target large model. The target large model performs reasoning analysis based on the prompt word to generate comprehensive and targeted network operation and maintenance analysis results. In addition, if a matching answer cannot be found directly, the target large model will generate new answers or suggestions based on existing knowledge and contextual information. Furthermore, the network operation and maintenance analysis results are displayed to the user. The analysis results integrate the results obtained from the SQL database, vector knowledge base, and other tool platform API call queries to ensure the comprehensiveness and accuracy of the information.
[0029] The embodiment of the present invention, through the above scheme, includes: obtaining multimodal question information of the target user; preprocessing and splitting the multimodal question information to obtain at least one target question information; using the target big model to classify the intent of at least one target question information to obtain the intent recognition result corresponding to at least one target question information; querying the target question information according to the intent recognition result to obtain the query result; based on the query results corresponding to different intent recognition results, using the target big model to infer the network operation and maintenance analysis result. The embodiment of the present invention effectively corrects the ambiguity and ambiguity of user input by adding steps such as question rewriting and sub-question splitting in the retrieval process, thereby improving the accuracy of the retrieval; in addition, the big model is used to quickly and accurately understand the user's intent, and based on the intent recognition result, the target question information is queried to obtain the query result, and the query results corresponding to the different intent recognition results are combined to infer the final analysis result, without the need for manual analysis based on experience, thereby improving the accuracy of the big model analysis.
[0030] In one embodiment of the present invention, according to the intention recognition result, the target question information is queried to obtain the query result, including: Step S21, if the intention recognition result is an SQL database query intention, then based on the target question information, an SQL query statement is generated, and based on the SQL query statement, a query result is obtained by querying a preset SQL database; Specifically, if the intent recognition result is an SQL database query intention, the keywords and entity information required for the SQL statement query are extracted based on the target question information, and then an SQL query statement is generated based on the target question information. That is, the keywords and entity information are converted into SQL query statements, the SQL query statements are sent to the SQL database to execute the query operation, and the query results are returned.
[0031] In one embodiment, based on the target question information, an SQL query statement is generated, including: Determine the similarity between the target question information and each historical question in the preset knowledge base; take the historical question with the highest similarity as the question sample; input the target question information, the question sample and the SQL statement associated with the question sample into the target big model to obtain the SQL query statement.
[0032] Specifically, refer to Figure 2 , Figure 2 is a schematic diagram of SQL statement generation provided by an embodiment of the present invention. Figure 2 The LLM in the figure represents the target big model, which calculates the similarity between the target question information and each historical question in the preset knowledge base, and then uses the historical question with the highest similarity as the question sample; further, the target question information, the question sample and the SQL statement associated with the question sample are constructed to obtain the Prompt query information, and then the Prompt query information is input into the target big model to obtain the SQL query statement.
[0033] It should be noted that the SQL query database-related questions and SQL information associated with the preset knowledge base can be stored, and according to any question in the preset knowledge base, various target questions with a high similarity to the question and the SQL query statements associated with the target questions can be input into the target big model, and then the target big model can be fine-tuned and optimized in combination with the output of the target big model and the SQL query statement corresponding to the question.
[0034] Step S22: if the intention recognition result is a vector database query intention, a multimodal feature vector is extracted based on the target question information, and a query result is obtained by querying in a preset vector database based on the multimodal feature vector; It should be noted that the multimodal feature vector includes text feature vector, image feature vector and speech feature vector. Specifically, when the intent recognition result is the vector database query intent: in order to process non-text modal data such as images and speech, it is necessary to introduce corresponding modality embedding function modules (such as image feature embedding, speech feature embedding, etc.) in the input layer.
[0035] In a specific embodiment, when the target question information includes text information, the text information in the target question information is segmented; then based on the text information after segmentation, a word vector, a text vector and a position vector are extracted; further, the word vector, the text vector and the position vector are added together as the input of the target large model, so as to use the encoder of the target large model to extract features and obtain a text feature vector. It should be noted that the word vector is a vector representing a single word, which can capture the semantic meaning of the word. The text vector is a vector representing an entire text segment (such as a sentence or paragraph). In one embodiment, the text vector can be obtained by weighted summing each word vector in the text, and the weight is usually assigned according to the importance or frequency of occurrence of the word in the text. The position vector is a vector representing the position of each word in the text in a sentence or paragraph. In a sequence model, the position vector can be used to capture the position information of elements in a sequence.
[0036] In one embodiment, when the target question information includes image information, feature extraction is performed on the image information to obtain image feature information; and then the image feature information is transformed to obtain an image feature vector. It can be understood that the input image X is extracted by a trained visual encoder (such as ResNet, ViT, etc.) to obtain image feature information Z, and then the image feature information Z is converted into a visual token H through a preset mapping matrix W. These feature vectors can be regarded as "tokens" of the image modality and converted into vectors of the same dimension as the text feature vector embeddings. Connected through this mapping matrix, this matrix is responsible for aligning or transforming the visual features so that they can be operated in a unified space.
[0037] When the target question information includes speech information, speech recognition is performed on the speech information to obtain speech feature information; the speech feature information is converted to obtain a speech feature vector; wherein the vector dimensions of the speech feature vector, the image feature vector, and the text feature vector are the same. It can be understood that the pre-trained speech model (such as WaveNet, MelGAN, etc.) extracts the features of the speech information and converts them into vector form, similar to the image processing mechanism, and converts the speech feature Z' into a speech token H' through a preset mapping matrix W'. These feature vectors can be regarded as "tokens" of the speech modality and converted into vectors of the same dimension as the text feature vector embeddings.
[0038] Furthermore, the multimodal feature vectors are fused to obtain a fused feature vector. It can be understood that a multimodal fusion layer is added after the encoder. The layer can adopt a cross-modal attention mechanism. The cross-modal attention mechanism allows vectors between different modalities to interact directly, thereby capturing cross-modal semantic associations. For text, images, and speech, text feature vector embeddings, image feature vectors, speech feature vectors, etc. can be fused through a cross-modal attention layer and processed in the same feature space. This allows the model to enhance its ability to understand and fuse multimodal information, and then query the vector database based on the fused feature vector to obtain the query result; optionally, during the query process, the similarity score between the fused feature vector and each candidate vector data queried from the vector database is calculated, for example, the similarity score is calculated using the cosine similarity calculation method. According to the similarity score, the candidate vector data are sorted to select the candidate vector data with the highest similarity score, and then the final query result is determined based on the candidate vector data with the highest similarity score, or the top N candidate vector data with higher similarity scores are selected, and then the final query result is determined based on the top N candidate vector data with higher similarity scores, wherein N represents a preset quantity threshold, which can be set according to actual conditions and is not specifically limited here.
[0039] In addition, in this embodiment, the initial training process of the target large model is as follows: using a large-scale and diverse multimodal corpus, through the data processing capabilities of the large model itself, automatically complete the data cleaning, word segmentation, labeling and other tasks, reduce manual intervention, and use unsupervised learning and self-supervised learning techniques to allow the model to be trained in massive data. For example, for image and text corpora, image-text pairs are collected from data collection libraries such as LAION-2B, LAION-COCO, LAION-Aesthetics and JourneyDB. It is possible to improve the quality of images and texts, thereby forming a high-quality corpus. In addition, for speech and text corpora, spoken text pairs can be collected from online platforms, crowdsourcing and audiobooks, covering a wide range of speakers, fields and environments.
[0040] In one embodiment, during the training process, the LoRa low-rank matrix is used to efficiently fine-tune the parameters of the model. Compared with the traditional fine-tuning method, the LoRa technology only updates the low-rank matrix, which can greatly reduce the computing resources and time cost required for fine-tuning. Secondly, due to the characteristics of the low-rank matrix, the LoRa technology can better capture the feature information related to the task, thereby improving the generalization ability of the model. The pre-processed expected data is divided into a training set and a validation set (a test set can also be set), and the model is iteratively trained using the training set. The value of the loss function is calculated in each iteration, and the model parameters are updated through the optimization algorithm. At the same time, the performance of the model is evaluated using the validation set to monitor the learning of the model and prevent overfitting.
[0041] Step S23, if the intent recognition result is an API interface query intent, an API interface query request is generated based on the target question information, and a query result is obtained based on the API interface query request.
[0042] Specifically, if the intent recognition result is an API interface query intent, the entity information and keyword information are analyzed; then based on the entity information and keyword information, parameter information is generated, wherein these parameters are parameter data used for API interface calls, and they define the specific details of the API request. Further, based on the parameter information, an API interface query request is generated, wherein the API interface query request is a structured message that contains all the necessary parameters for the request so that the external API can understand and perform the corresponding operation, and then based on the API interface query request, the external API interface is called to perform the query operation according to the parameters in the request, and the query result is returned.
[0043] As a specific example: departure city: New York; arrival city: Los Angeles; departure date: May 1, 2023; ticket class: economy class; airline (optional): no special requirements. The parameters are as follows: API interface name: FlightBookingAPI; request method: GET; request URL: / flights, request parameters: {"origin": "New York", "destination": "Los Angeles", "departureDate": "2023-05-01", "class": "economy"}.
[0044] Through the above scheme, the embodiment of the present invention realizes the use of a large model to quickly and accurately understand the user's intention, and according to the intention recognition result, it determines whether the task involves operations such as SQL database query, knowledge base search, other tool API call query, etc., and can automatically select the most appropriate retrieval and processing strategy, so that the final analysis answer integrates the results obtained from the SQL database, vector knowledge base and other tool platform API call query, ensuring the comprehensiveness and accuracy of the information.
[0045] The network operation and maintenance analysis device provided by the present invention is described below. The network operation and maintenance analysis device described below and the network operation and maintenance analysis method described above can be referenced to each other.
[0046] Figure 3 is a schematic diagram of the structure of the network operation and maintenance analysis device provided by the present invention, such as Figure 3 As shown, a network operation and maintenance analysis device according to an embodiment of the present invention includes: An acquisition module 21 is used to acquire multimodal question information of a target user; The processing module 22 is used to pre-process and split the multimodal question information to obtain at least one target question information; A classification module 23, configured to classify the intent of at least one target question information using the target macro model, and obtain an intent recognition result corresponding to the at least one target question information; A query module 24 is used to query the target question information according to the intention recognition result to obtain the query result; The reasoning module 25 is used to integrate the query results corresponding to the different intent recognition results, and obtain the network operation and maintenance analysis results based on the integration results by using the target large model reasoning; Among them, the target large model is the large model corresponding to the Transformer network architecture.
[0047] The query module 24 is also used for: If the intent recognition result is an SQL database query intent, then a SQL query statement is generated based on the target question information, and a query result is obtained by querying a preset SQL database based on the SQL query statement; If the intent recognition result is a vector database query intent, a multimodal feature vector is extracted based on the target question information, and a query result is obtained by querying in a preset vector database based on the multimodal feature vector; If the intent recognition result is an API interface query intent, an API interface query request is generated based on the target question information, and a query result is obtained based on the API interface query request.
[0048] The query module 24 is also used for: Determine the similarity between the target question information and each historical question in the preset knowledge base; Use the historical questions with the highest similarity as question samples; The target problem information, problem samples and SQL statements associated with the problem samples are input into the target large model to obtain SQL query statements.
[0049] The query module 24 is also used for: The multimodal feature vector includes a text feature vector, an image feature vector and a speech feature vector; When the target question information includes text information, the text information in the target question information is segmented; based on the text information after the segmentation, a word vector, a text vector and a position vector are extracted; the word vector, the text vector and the position vector are added together to perform feature extraction to obtain a text feature vector; When the target problem information includes image information, feature extraction is performed on the image information to obtain image feature information; the image feature information is transformed to obtain an image feature vector; When the target question information includes voice information, voice recognition is performed on the voice information to obtain voice feature information; the voice feature information is transformed to obtain a voice feature vector; wherein the vector dimensions of the voice feature vector, the image feature vector and the text feature vector are the same; Fusing multi-modal feature vectors to obtain a fused feature vector; Based on the fused feature vector, the query result is obtained by querying the vector database.
[0050] The query module 24 is also used for: Based on the target question information, analyze and obtain entity information and keyword information; Generate parameter information based on entity information and keyword information; Based on the parameter information, an API interface query request is generated, and based on the API interface query request, an external API interface query is called to obtain the query result.
[0051] The classification module 23 is also used for: For any target question information: Use the target large model to perform entity recognition and semantic analysis on the target question information to obtain entity recognition results and semantic analysis results; Based on the entity recognition results and semantic analysis results, determine the intent recognition result corresponding to the target question information, where the intent recognition result includes SQL database query intent, vector database query intent and / or API interface query intent.
[0052] The processing module 22 is also used for: Performing a grammar check on the multimodal question information, and correcting the multimodal question information according to the grammar check result; The corrected multimodal question information is split into question pieces to obtain at least one target question information.
[0053] The reasoning module 25 is also used to: Integrate the query results corresponding to different intent recognition results to obtain an integrated result; The integration results are packaged into prompt words in a preset format, and the prompt words are input into the target large model to infer the network operation and maintenance analysis results.
[0054] It should be noted here that the above-mentioned device provided in the embodiment of the present invention can implement all the method steps implemented in the above-mentioned method embodiment, and can achieve the same technical effect. The parts and beneficial effects that are the same as the method embodiment in this embodiment will not be described in detail here.
[0055] Figure 4 is a schematic diagram of the structure of the electronic device provided by the present invention, such as Figure 4 As shown, the electronic device may include: a processor 310, a memory 320, a communication interface 330 and a communication bus 340, wherein the processor 310, the memory 320, and the communication interface 330 communicate with each other through the communication bus 340. The processor 310 may call the logic instructions in the memory 320 to execute the network operation and maintenance analysis method.
[0056] In addition, the logic instructions in the above-mentioned memory 320 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods of various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.
[0057] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the network operation and maintenance analysis method provided by the above-mentioned methods.
[0058] Through the description of the above implementation modes, those skilled in the art can clearly understand that each implementation mode can be implemented by means of software plus a necessary general hardware platform, or of course by hardware. Based on such an understanding, the above technical solution can essentially or in other words be embodied in the form of a software product that contributes to the prior art. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each embodiment or some parts of the embodiment.
[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A network operation and maintenance analysis method, characterized in that: include: Obtain multimodal question information from target users; Preprocessing and splitting the multimodal question information to obtain at least one target question information; Using the target large model to perform intent classification on the at least one target question information, and obtain an intent recognition result corresponding to the at least one target question information; According to the intention recognition result, query the target question information to obtain a query result; Based on the query results corresponding to the different intent recognition results, the network operation and maintenance analysis results are obtained by using the target large model inference; Among them, the target large model is a large model corresponding to the Transformer network architecture.
2. The network operation and maintenance analysis method according to claim 1, characterized in that: The querying of the target question information according to the intention recognition result to obtain the query result includes: If the intention recognition result is an SQL database query intention, then based on the target question information, a SQL query statement is generated, so as to obtain the query result by querying a preset SQL database based on the SQL query statement; If the intention recognition result is a vector database query intention, then based on the target question information, a multimodal feature vector is extracted to obtain the query result by querying in a preset vector database based on the multimodal feature vector; If the intent recognition result is an API interface query intent, an API interface query request is generated based on the target question information, so as to obtain the query result based on the API interface query request.
3. The network operation and maintenance analysis method according to claim 2, characterized in that: The step of generating a SQL query statement based on the target question information includes: Determine the similarity between the target question information and each historical question in a preset knowledge base; Taking the historical question with the highest similarity as a question sample; The target question information, the question sample and the SQL statement associated with the question sample are input into the target big model to obtain the SQL query statement.
4. The network operation and maintenance analysis method according to claim 2, characterized in that: The multimodal feature vector includes a text feature vector, an image feature vector and a speech feature vector; The step of extracting a multimodal feature vector based on the target question information, and querying a preset vector database based on the multimodal feature vector to obtain the query result includes: When the target question information includes text information, the text information in the target question information is segmented; based on the text information after the segmentation, a word vector, a text vector and a position vector are extracted; the word vector, the text vector and the position vector are added together to perform feature extraction to obtain the text feature vector; When the target question information includes image information, extracting features from the image information to obtain image feature information; transforming the image feature information to obtain the image feature vector; When the target question information includes voice information, performing voice recognition on the voice information to obtain voice feature information; converting the voice feature information to obtain the voice feature vector; wherein the vector dimensions of the voice feature vector, the image feature vector and the text feature vector are the same; Fusing the multimodal feature vectors to obtain a fused feature vector; Based on the fused feature vector, the query result is obtained by querying the vector database.
5. The network operation and maintenance analysis method according to claim 2, characterized in that: The step of generating an API interface query request based on the target question information, and querying to obtain the query result based on the API interface query request, includes: Based on the target question information, analyzing and obtaining entity information and keyword information; Generate parameter information based on the entity information and the keyword information; Based on the parameter information, an API interface query request is generated, and based on the API interface query request, an external API interface query is called to obtain the query result.
6. The network operation and maintenance analysis method according to claim 1, characterized in that: The using the target big model to perform intent classification on the at least one target question information to obtain an intent recognition result corresponding to the at least one target question information includes: For any of the above mentioned target question information: Using the target large model to perform entity recognition and semantic analysis on the target question information, to obtain entity recognition results and semantic analysis results; Based on the entity recognition result and the semantic analysis result, determine the intent recognition result corresponding to the target question information, wherein the intent recognition result includes SQL database query intent, vector database query intent and / or API interface query intent.
7. The network operation and maintenance analysis method according to claim 1, characterized in that: The preprocessing and splitting of the multimodal question information to obtain at least one target question information includes: Performing a grammar check on the multimodal question information to correct the multimodal question information according to a grammar check result; The corrected multimodal question information is split into questions to obtain the at least one target question information.
8. The network operation and maintenance analysis method according to claim 1, characterized in that: The query results corresponding to the different intent recognition results are based on the network operation and maintenance analysis results obtained by inferring the target large model, including: Integrate the query results corresponding to the different intent recognition results to obtain an integrated result; The integration result is packaged into a prompt word in a preset format, and the prompt word is input into the target large model to infer the network operation and maintenance analysis result.
9. A network operation and maintenance analysis device, characterized in that: include: The acquisition module is used to obtain the multimodal question information of the target user; A processing module, used for preprocessing and splitting the multimodal question information to obtain at least one target question information; A classification module, used to classify the intent of the at least one target question information by using the target macro model, and obtain an intent recognition result corresponding to the at least one target question information; A query module, used to query the target question information according to the intention recognition result to obtain a query result; The inference module is used to integrate the query results corresponding to different intent recognition results, and to obtain the network operation and maintenance analysis results based on the integration results by using the target large model for inference; wherein the target large model is a large model corresponding to the Transformer network architecture.
10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the network operation and maintenance analysis method as described in any one of claims 1 to 8 is implemented.
Citation Information
Patent Citations
Shared content processing method, device, computer equipment and storage medium
CN111581510A
Intention recognition and question answering method and device, electronic equipment and storage medium
CN117194730A
SQL statement generation method and device
CN117708157A
Multi-modal information retrieval method, device and equipment, readable storage medium and computer program product
CN117909555A
Simulation audio identification method, device and equipment based on multi-modal large model
CN118053451A
Cited By
Operation and maintenance question-answering system and method based on large language model workflow
CN119441410A
Method and device for converting natural language question into SQL statement and medium
CN119597785A
Abnormal information processing method and device, equipment, storage medium and program product
CN119788495A
LLM-based NL2API and NL2SQL fused AI question-answering system and method
CN119829727A
5G private network operation and maintenance method based on large model adaptive capability and related device
CN120264329A
