Intelligent question and answer data processing method and device and electronic equipment

Abstract

The invention provides an intelligent question and answer data processing method and apparatus, and an electronic device. The method comprises the steps of obtaining a to-be-queried urban rail public work problem; analyzing the urban rail public work problem to be inquired to obtain an analysis result; based on the analysis result, performing intention recognition on the to-be-queried urban rail public work problem to obtain a target intention category to which the to-be-queried urban rail public work problem belongs; judging whether the urban rail public work problem to be inquired belongs to a conventional problem or not; and if the to-be-queried urban rail public work problem belongs to the conventional problem, based on the target intention category, retrieving a preset urban rail public work knowledge base to obtain a target query result corresponding to the to-be-queried urban rail public work problem. In the mode, the urban rail public work knowledge base is generated based on the knowledge graph associated with the urban rail public work, and entity expansion can be realized based on the knowledge graph, so that when the urban rail public work knowledge base is retrieved based on the target intention category to which the to-be-queried urban rail public work problem belongs, the recall coverage rate can be effectively improved, and the query efficiency is improved. And the accuracy of the query result is improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an intelligent question-and-answer data processing method, apparatus, and electronic device. Background Technology

[0002] Retrieval-Augmented Generation (RAG) is a deep learning model architecture that combines retrieval and generation. It aims to address the "illusion" problem that large language models (LLMs) may encounter when faced with unknown questions, generating seemingly reasonable but actually incorrect or inaccurate answers. RAGs attempt to improve the accuracy and relevance of content by retrieving relevant information from a broad document database before the LLM generates an answer, and then using this information to guide the generation process. However, in practical applications, diverse data sources, inconsistent document data processing quality, and uncertain or disjointed user question formats can all negatively impact the accuracy of knowledge retrieval in RAGs. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent question-and-answer data processing method, apparatus, and electronic device to improve the accuracy of query results.

[0004] This invention provides an intelligent question-answering data processing method, comprising: acquiring a question about urban rail transit engineering to be queried; parsing the question about urban rail transit engineering to obtain a parsing result; based on the parsing result, performing intent recognition on the question about urban rail transit engineering to obtain the target intent category to which the question about urban rail transit engineering belongs; determining whether the question about urban rail transit engineering belongs to a routine question; if the question about urban rail transit engineering belongs to a routine question, searching a preset urban rail transit engineering knowledge base based on the target intent category to obtain the target query result corresponding to the question about urban rail transit engineering; wherein, the urban rail transit engineering knowledge base is generated based on a knowledge graph related to urban rail transit engineering.

[0005] Furthermore, the urban rail transit engineering knowledge base is generated in the following manner: The original data source is obtained; dynamic data is collected in real time from multiple preset devices; the original data source and dynamic data are preprocessed according to a preset preprocessing method, and the data preprocessing results and metadata information corresponding to multiple knowledge data are output; the data preprocessing results are divided into blocks to obtain multiple blocks of data; based on the entities and relationships in each block of data, a knowledge graph related to urban rail transit engineering is constructed; vector transformation is performed on each block of data and the graph nodes of the knowledge graph to obtain data vectors corresponding to multiple knowledge data.

[0006] Furthermore, based on the parsing results, the steps for performing intent recognition on the urban rail transit engineering question to obtain the target intent category to which the urban rail transit engineering question belongs include: obtaining a preset intent category set; wherein, the intent category set includes multiple preset intent categories; for each preset intent category, calculating the probability that the parsing result belongs to the preset intent category based on the parsing results; and determining the preset intent category with the highest probability as the target intent category to which the urban rail transit engineering question belongs.

[0007] Furthermore, the steps for determining whether a query about urban rail transit engineering issues belongs to the category of routine issues include: obtaining a preset hot topic knowledge base; wherein, the hot topic knowledge base includes: multiple hot topics and the corresponding results for each hot topic; for each hot topic, based on the parsing results, calculating the matching degree between the query about urban rail transit engineering issues and the hot topic; if at least one matching degree is greater than a preset matching degree threshold, it is determined that the query about urban rail transit engineering issues belongs to the category of high-frequency issues; if each matching degree is less than or equal to the preset matching degree threshold, it is determined that the query about urban rail transit engineering issues belongs to the category of routine issues.

[0008] Furthermore, the method also includes: if the urban rail transit engineering issue to be queried is a high-frequency issue, obtaining the target issue result corresponding to the urban rail transit engineering issue to be queried from the hot topic knowledge base; and determining the target query result corresponding to the urban rail transit engineering issue to be queried based on the target issue result.

[0009] Furthermore, the urban rail transit engineering knowledge base includes: data vectors and metadata information corresponding to multiple knowledge data. The steps for retrieving the target query result corresponding to the urban rail transit engineering question based on the target intent category include: vectorizing the urban rail transit engineering question to obtain a query question vector; selecting multiple first data vectors from the data vectors corresponding to multiple knowledge data based on the query question vector, target intent category, and metadata information; calculating the similarity between the query question vector and the first data vector for each first data vector to obtain a vector similarity result; determining a preset number of target similarity results according to the vector similarity results from high to low; determining the first data vector corresponding to each target similarity result as a candidate data vector; and determining the target query result corresponding to the urban rail transit engineering question based on the knowledge data corresponding to the candidate data vectors.

[0010] Furthermore, the steps for determining the target query result corresponding to the urban rail transit engineering problem based on the knowledge data corresponding to the candidate data vectors include: rearranging and filtering the knowledge data corresponding to the candidate data vectors to output target knowledge data; inputting the target knowledge data and the urban rail transit engineering problem to be queried into a preset answer generation model, and outputting the initial query result through the answer generation model; and performing answer post-processing on the initial query result according to a preset post-processing method to obtain the target query result corresponding to the urban rail transit engineering problem to be queried.

[0011] This invention provides an intelligent question-and-answer data processing device, comprising: an acquisition module for acquiring a question about urban rail transit engineering to be queried; a parsing module for parsing the question about urban rail transit engineering to obtain a parsing result; an identification module for identifying the intent of the question about urban rail transit engineering to be queried based on the parsing result, and obtaining the target intent category to which the question about urban rail transit engineering belongs; a judgment module for judging whether the question about urban rail transit engineering to be queried is a regular question; and a retrieval module for, if the question about urban rail transit engineering to be queried is a regular question, retrieving a preset urban rail transit engineering knowledge base based on the target intent category, and obtaining the target query result corresponding to the question about urban rail transit engineering; wherein, the urban rail transit engineering knowledge base is generated based on a knowledge graph related to urban rail transit engineering.

[0012] The present invention provides an electronic device, including a processor and a memory, wherein the memory stores machine-executable instructions that can be executed by the processor, and the processor executes the machine-executable instructions to implement the intelligent question-answering data processing method described above.

[0013] The present invention provides a machine-readable storage medium storing machine-executable instructions. When the machine-executable instructions are invoked and executed by a processor, the machine-executable instructions cause the processor to implement any of the above-mentioned intelligent question-answering data processing methods.

[0014] The intelligent question-answering data processing method, apparatus, and electronic device provided by this invention acquire a question about urban rail transit engineering to be queried; parse the question to obtain a parsing result; based on the parsing result, perform intent recognition on the question to obtain the target intent category to which the question belongs; determine whether the question is a routine question; if the question is a routine question, search a preset urban rail transit engineering knowledge base based on the target intent category to obtain the target query result corresponding to the question. In this method, since the urban rail transit engineering knowledge base is generated based on a knowledge graph related to urban rail transit engineering, entity expansion can be achieved based on the knowledge graph. Therefore, when searching the urban rail transit engineering knowledge base based on the target intent category to which the question belongs, the recall coverage rate can be effectively improved, thereby improving the accuracy of the query results. Attached Figure Description

[0015] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0016] Figure 1 A flowchart of an intelligent question-answering data processing method provided in an embodiment of the present invention; Figure 2 A flowchart of another intelligent question-answering data processing method provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an intelligent question-and-answer data processing device provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. 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] RAG (Retrieval-Augmented Generation) is a deep learning model architecture that combines retrieval and generation. It aims to solve the "illusion" problem that large language models (LLMs) may encounter when facing unknown problems. Conventional RAGs mainly have the following shortcomings: 1. Unstable retrieval quality: Traditional vector recall is prone to "semantic similarity but business irrelevant", especially when professional terms overlap, false recall will occur.

[0019] 2. Coarse context processing: For long documents, RAG directly divides them into blocks and stores them in the vector library, lacking "domain-related clipping". As a result, the recalled fragments may be context-segmented, leading to missing or incorrect information when the model is generated.

[0020] 3. Low retrieval efficiency: Importing content data from external knowledge bases poses a significant challenge to retrieval efficiency, resulting in long response times and impacting the user experience.

[0021] 4. Low timeliness of knowledge base updates: Generally, documents need to be manually imported to update the knowledge base, and it is difficult to obtain highly time-sensitive data.

[0022] 5. The quality of Q&A is sensitive to the way users ask questions: overly simple or disjointed dialogues pose a significant challenge to the accuracy of retrieval.

[0023] Based on this, embodiments of the present invention provide an intelligent question-and-answer data processing method, apparatus, and electronic device, which can be applied to applications that require automatic querying of answers to urban rail transit engineering questions.

[0024] To facilitate understanding of this embodiment, a method for intelligent question-answering data processing disclosed in this embodiment of the invention will first be introduced, such as... Figure 1 As shown, the method includes the following steps: Step S102: Obtain the urban rail transit maintenance issue to be queried; The aforementioned urban rail transit maintenance issues to be queried can be understood as questions input by dispatchers or maintenance personnel during the operation and maintenance of urban rail transit tracks. For example, during urban rail operation and maintenance, dispatchers might input questions such as "How to handle a switch that cannot be switched?" or maintenance personnel might ask questions such as "What is the safe distance during tunnel maintenance?" In practice, the system can retrieve urban rail transit maintenance issues to be queried that are input in natural language by dispatchers, maintenance personnel, or other relevant urban rail transit maintenance staff.

[0025] Step S104: Analyze the urban rail transit engineering problem to be queried and obtain the analysis result; The query on urban rail transit engineering issues can be parsed. For example, semantic understanding can be performed on the query on urban rail transit engineering issues to extract keywords and constraints, and corresponding parsing results can be obtained. For specific parsing methods, please refer to relevant technologies, which will not be elaborated here.

[0026] Step S106: Based on the parsing results, perform intent recognition on the urban rail transit engineering issue to be queried to obtain the target intent category to which the urban rail transit engineering issue to be queried belongs; In practice, the target intent category of the urban rail transit engineering question can be identified based on the parsing results corresponding to the question. For example, the target intent category could be status query, fault diagnosis, or regulation query. For instance, if the question is "What is the safe distance during tunnel maintenance?", parsing can confirm that the target intent category of the question is "regulation query".

[0027] Step S108: Determine whether the urban rail transit engineering issue to be queried is a routine issue; The aforementioned routine problems can be understood as problems with relatively low query frequency; in actual implementation, after obtaining the urban rail transit engineering problem to be queried, it can be further determined whether the urban rail transit engineering problem to be queried belongs to routine problems.

[0028] Step S110: If the urban rail transit engineering issue to be queried is a routine issue, the preset urban rail transit engineering knowledge base is searched based on the target intent category to obtain the target query result corresponding to the urban rail transit engineering issue to be queried; wherein, the urban rail transit engineering knowledge base is generated based on the knowledge graph related to urban rail transit engineering.

[0029] Knowledge graphs are a technology that uses graph structures to represent knowledge and relationships. By constructing semantic networks through nodes and edges, they aim to transform scattered data into a machine-understandable and reasonable knowledge system. Nodes represent entities, and edges represent relationships between entities. In practical implementation, a knowledge base for urban rail transit engineering can be pre-generated by constructing a knowledge graph. Once it is confirmed that the urban rail transit engineering question to be queried is a routine question, the knowledge base can be retrieved according to the target intent category to obtain the target query results corresponding to the urban rail transit engineering question.

[0030] The aforementioned intelligent question-answering data processing method involves: acquiring the urban rail transit engineering question to be queried; parsing the question to obtain the parsing result; based on the parsing result, performing intent recognition on the question to obtain the target intent category to which the question belongs; determining whether the question is a routine question; and if it is a routine question, searching a pre-defined urban rail transit engineering knowledge base based on the target intent category to obtain the target query result corresponding to the question. In this approach, since the urban rail transit engineering knowledge base is generated based on a knowledge graph related to urban rail transit engineering, entity expansion can be achieved based on the knowledge graph. Therefore, when searching the knowledge base based on the target intent category to which the question belongs, the recall coverage rate can be effectively improved, thereby increasing the accuracy of the query results.

[0031] This invention also provides another intelligent question-answering data processing method, which is implemented based on the method in the above embodiments, and includes the following steps: Step 1: Obtain the urban rail transit engineering issues to be queried; Step two: Analyze the urban rail transit engineering problem to be queried to obtain the analysis results; Step 3: Based on the parsing results, perform intent recognition on the urban rail transit engineering question to be queried, and obtain the target intent category to which the urban rail transit engineering question belongs; This third step can be achieved through the following steps 30 to 32: Step 30: Obtain a preset set of intent categories; wherein the set of intent categories includes multiple preset intent categories; The number of preset intent categories can be set according to actual needs. In actual implementation, multiple preset intent categories can be predefined to form an intent category set. For example, this intent category set can be represented as: ,in, ... This corresponds to n preset intent categories, such as "status query", "rule query", "fault diagnosis", etc.

[0032] Step 31: For each preset intent category, calculate the probability that the parsing result belongs to the preset intent category based on the parsing result; In practical implementation, based on the parsing results corresponding to the urban rail transit engineering question to be queried, and combined with the context, the probability of the parsing result belonging to each preset intent category can be calculated. For example, keywords corresponding to each preset intent category can be predefined. For each preset intent category, the keywords in the parsing result can be compared with the keywords corresponding to that preset intent category. Based on the comparison, the probability of the parsing result belonging to that preset intent category can be determined. Generally, the higher the repetition between the keywords in the parsing result and the keywords corresponding to that preset intent category, the higher the probability; the lower the repetition between the keywords in the parsing result and the keywords corresponding to that preset intent category, the lower the probability.

[0033] Step 32: Determine the preset intent category with the highest probability as the target intent category to which the urban rail transit engineering question to be queried belongs.

[0034] After calculating the probability that the parsed result belongs to each preset intent category, the preset intent category with the highest probability can be determined as the target intent category to which the urban rail transit engineering question belongs. For example, it can be represented by the following expression: ; Where Q represents the natural language question input by the user (corresponding to the urban rail transit engineering question to be queried above); C represents the set of intent categories; and c represents the preset intent category. y represents the probability that the natural language question Q belongs to the preset intent category c; y represents the identified target intent category; argmax represents finding the c that maximizes the value of P from the intent category set C.

[0035] Step 4: Obtain the preset hot topic knowledge base; the hot topic knowledge base includes: multiple hot topics and the corresponding results for each hot topic. The aforementioned trending questions can be understood as questions that have been frequently queried within a recent preset time period; the results can be understood as the answers to these trending questions. In actual implementation, the recently frequently queried trending questions and their corresponding results can be stored in a trending knowledge base.

[0036] Step 5: For each hot topic, based on the analysis results, calculate the matching degree between the urban rail transit engineering issue to be queried and the hot topic issue; In practice, after obtaining the hot topic knowledge base, the matching degree between the urban rail transit engineering question to be queried and each hot topic question can be calculated based on the parsing results corresponding to the above-mentioned urban rail transit engineering question. For example, the matching degree can be confirmed based on the text similarity and semantic similarity between the urban rail transit engineering question to be queried and each hot topic question.

[0037] Step 6: If there is at least one matching degree greater than the preset matching degree threshold, it is determined that the urban rail transit engineering issue to be queried is a high-frequency issue. Step 7: If each matching degree is less than or equal to the preset matching degree threshold, it is determined that the urban rail transit engineering problem to be queried belongs to the category of routine problems.

[0038] The aforementioned preset matching threshold can be set according to actual needs. In actual implementation, when one or more matching degrees are greater than the preset matching threshold, the urban rail transit engineering problem to be queried can be considered a high-frequency problem, that is, a problem that is queried relatively many times; when each matching degree is less than or equal to the preset matching threshold, the urban rail transit engineering problem to be queried can be considered a regular problem, that is, a problem that is queried relatively few times.

[0039] Step 8: If the urban rail transit engineering question to be queried is a routine question, the preset urban rail transit engineering knowledge base is searched based on the target intent category to obtain the target query result corresponding to the urban rail transit engineering question to be queried; wherein, the urban rail transit engineering knowledge base is generated based on the knowledge graph related to urban rail transit engineering.

[0040] The urban rail transit engineering knowledge base is generated through the following steps A through F: Step A: Obtain the original data source; The original data source can include various types of information such as design drawings, engineering regulations, maintenance manuals, historical maintenance records, accident cases, and inspection reports.

[0041] Step B involves collecting dynamic data in real time from multiple preset devices. The aforementioned preset equipment terminals can be track line sensors, turnout monitoring devices, sleeper settlement detection devices, etc. In actual implementation, multiple preset equipment terminals can be pre-connected to collect real-time dynamic data by accessing the streaming data generated by multiple preset equipment terminals.

[0042] Step C: Preprocess the original data source and dynamic data according to the preset preprocessing method, and output the data preprocessing results and metadata information corresponding to multiple knowledge data. In practical implementation, the original data sources and dynamic data obtained above can be preprocessed according to preset preprocessing methods, such as noise reduction, cleaning, unified formatting, and standardization. Normalization can be performed by combining a professional terminology dictionary in the engineering field to ensure the consistency of subsequent knowledge representation. Through this processing, the data preprocessing results and metadata information corresponding to multiple knowledge data are obtained. The multiple knowledge data can be understood as each knowledge data in the original data source and dynamic data. The metadata information can include the source, timestamp, applicable route, standard number, and other metadata of the knowledge data for filtering and constraint during retrieval. The applicable route can be understood as the profession and scenario to which the document is applicable; the standard number can be understood as the document number.

[0043] Step D involves dividing the preprocessed data into blocks to obtain multiple data blocks. In practice, the data preprocessing results are usually long documents. The data preprocessing results can be divided into semantic units to ensure that each block meets the length limit of the subsequent processing while maintaining the integrity of the knowledge semantics.

[0044] Step E: Based on the entities and relationships in each data block, construct a knowledge graph related to urban rail transit engineering. In practical implementation, after obtaining multiple data chunks, named entity recognition and relation extraction techniques, combined with civil engineering ontology rules, can be used to generate a knowledge graph containing entity nodes such as equipment, processes, faults, standards, and detection methods, along with their semantic relationships. The constructed knowledge graph not only supplements the modeling of the chunked documents but also aims to "improve the relevance of the retrieval stage." By modeling civil engineering entities and their relationships using the knowledge graph, the system can automatically include fragments highly relevant to the entities involved in the user's question into the candidate results during the retrieval stage. For example, when a user queries "Which components are affected by sleeper settlement?", even if the document does not explicitly mention "settlement," the system can retrieve relevant fragments through the graph relationships (sleeper settlement → gauge change → rail stress anomaly). The ultimate goal is to improve the entity expansion capability during the retrieval stage, resulting in higher candidate fragment coverage and thus improving the accuracy and completeness of the generated answers.

[0045] The principle is to extract engineering entities and relationships from segmented data to form a knowledge graph structure. The formula can be expressed as: knowledge graph ; Among them, entity set For example, "rail sleeper", "settlement", "turnout"; set of relationships For example, "settlement - leading to - gauge change".

[0046] Step F involves performing vector transformation on each data block and the graph node of the knowledge graph to obtain data vectors corresponding to multiple knowledge data.

[0047] A domain-optimized semantic embedding model can be used to convert each data block and knowledge graph node into a vector representation, resulting in data vectors corresponding to multiple knowledge data. For details, please refer to relevant technologies; they will not be elaborated upon here. The data vectors corresponding to the multiple knowledge data obtained above can be saved to an urban rail transit engineering knowledge base, which supports efficient similarity retrieval.

[0048] This step eight can be achieved through the following steps 80 to 85: Step 80: Perform vector transformation on the urban rail transit engineering problem to be queried to obtain the problem vector; In practice, the urban rail transit engineering problem to be queried can be transformed into a vector representation to obtain the problem vector, which is then used for similarity matching.

[0049] Step 81: Based on the question vector to be queried, the target intent category, and metadata information, select multiple first data vectors from the data vectors corresponding to multiple knowledge data. In practice, data vectors corresponding to multiple knowledge data are usually categorized according to different intent types. For example, there might be data vectors for knowledge data related to status queries, or data vectors for knowledge data related to rule queries. Based on the vector of the query question, combined with the target intent type and metadata information, filtering and selection can be performed to identify multiple relevant primary data vectors.

[0050] Step 82: For each first data vector, calculate the similarity between the query vector and the first data vector to obtain the vector similarity result; In practice, the similarity between the query vector and each first data vector can be calculated to obtain the vector similarity result for each first data vector. For details on how to calculate the similarity, please refer to relevant technologies; these will not be elaborated upon here.

[0051] Step 83: Determine a preset number of target similarity results according to the vector similarity results from high to low. Step 84: Determine the first data vector corresponding to each target similarity result as the candidate data vector; The preset number can be set according to actual needs. The preset number of target similarity results can be selected in descending order of vector similarity, and the first data vector corresponding to each target similarity result can be used as the candidate data vector. For example, candidate Top-K related document fragments can be obtained.

[0052] Step 85: Based on the knowledge data corresponding to the candidate data vectors, determine the target query result corresponding to the urban rail transit engineering problem to be queried.

[0053] This step 85 can be achieved through steps 850 to 852: Step 850: Rearrange and filter the knowledge data corresponding to the selected data vector, and output the target knowledge data; The system can sort the knowledge data corresponding to the retrieved candidate data vectors according to relevance and contextual consistency, and output the selected target knowledge data, which can be a part or all of the knowledge data corresponding to the candidate data vectors.

[0054] Step 851: Input the target knowledge data and the urban rail transit engineering question to be queried into the preset answer generation model, and output the initial query results through the answer generation model; Based on the target knowledge data and the urban rail transit engineering question to be queried, an initial query result is generated using a preset answer generation model.

[0055] Step 852: According to the preset post-processing method, perform answer post-processing on the initial query results to obtain the target query results corresponding to the urban rail transit engineering problem to be queried.

[0056] The initial query results can be post-processed according to a preset post-processing method. For example, the initial query results can be verified for facts, marked with references, and formatted for output to obtain the target query result corresponding to the urban rail transit engineering issue to be queried. The target query result can then be fed back to the user.

[0057] Step 9: If the urban rail transit engineering issue to be queried is a high-frequency issue, obtain the target issue result corresponding to the urban rail transit engineering issue to be queried from the hot topic knowledge base; Step 10: Determine the target query result corresponding to the urban rail transit engineering problem to be queried based on the target problem result.

[0058] In practical implementation, since the hot topic knowledge base stores frequently queried questions and their corresponding results, when it is confirmed that the urban rail transit engineering question to be queried is a high-frequency question, the target question result can be directly retrieved from the hot topic knowledge base. This reduces redundant retrieval and generation processes, ensuring response efficiency in high-concurrency scenarios. In high-concurrency, highly repetitive application scenarios such as urban rail transit engineering operation and maintenance, this significantly improves system response speed and throughput.

[0059] For example, a hot topic knowledge base can maintain a cache table H of hot topics, where the user inputs a natural language question, Q. The formula can be expressed as: If Then the cached answer can be output directly: ; Where: Q represents the natural language question input by the user; H represents the hot question cache table, which stores recent high-frequency questions and their corresponding answers; H[Q] represents the cached answer retrieved from the hot question cache table H based on the natural language question Q; Ans(Q) represents the final answer returned to the user (corresponding to the target query result mentioned above).

[0060] In practical applications, the target question result can be directly returned to the user as the target query result. When there are multiple target question results, the target question results can also be rearranged and filtered according to the steps 850 to 852 above to output the final target query result.

[0061] For ease of understanding, see Figure 2 The flowchart shown illustrates another intelligent question-answering data processing method. The offline knowledge base construction module is used to process, structure, and semantically model the raw multi-source data to form a knowledge support system that can be invoked by online question answering. This module includes: (1) Original data source; (2) Real-time data stream acquisition module; The real-time data stream acquisition module is used to access streaming data generated by track sensors, turnout monitoring devices, and sleeper settlement detection devices in real time. Through a stream processing framework, it performs incremental cleaning, anomaly detection, and index extraction on the data, enabling dynamic supplementation of static document knowledge. This module dynamically supplements the urban rail transit knowledge base by accessing and processing real-time data from track sensors and turnout monitoring. Unlike traditional methods that rely solely on static knowledge bases, this module processes streaming data into structured fragments and writes them into the vector knowledge base through embedded models and incremental updates of knowledge graphs. For example, it supplements information such as the data reception time and size according to standard structured representations. This achieves "real-time dynamic incremental knowledge updates," ensuring that when users query the current track status, the query results contain the latest data, avoiding the problem of "insufficient timeliness" in the knowledge base.

[0062] This module performs window aggregation and incremental updates on sensor stream data, transforming it into searchable knowledge fragments. The formula can be expressed as: given a sequence of sensor stream data... During the time window Internal computational statistical characteristics: ; in, Indicates within the time window Internally computed sensor stream data sequence Statistical characteristics; The mean, The variance is used to detect anomalies, and max represents the variance within the time window. Sensor stream data sequence detected inside The maximum value in the time window, min represents the maximum value in the time window. Sensor stream data sequence detected inside The minimum value of the data.

[0063] Incremental updates to the knowledge base: ; in, The vector knowledge base represents the contents at time t. Indicates the time step The content of the updated vector knowledge base; X represents real-time streaming data collected from track line sensors, turnout monitoring devices, sleeper settlement detection devices, etc. Represented as a window function, it is used to aggregate and extract features from streaming data within a sliding time window, and output statistical features (such as mean, variance, peak value, trend coefficient, etc.). This is a feature embedding function used to map window features into a vector space representation, often based on embedding models (such as BERT, Word2Vec, or structured coding networks). The union operation represents a set, used to incrementally write newly generated embedding vectors into a knowledge base; The time update step size is used to control the refresh frequency of the knowledge base.

[0064] This ensures that the knowledge base is dynamically updated over time, achieving "real-time dynamic incremental knowledge updates".

[0065] (3) Data preprocessing module The function of this data preprocessing module can be referred to the relevant description in the foregoing embodiments, and will not be repeated here.

[0066] (4) Text segmentation module The functionality of this text segmentation module can be found in the descriptions in the foregoing embodiments, and will not be repeated here.

[0067] (5) Knowledge Graph Construction Module The functionality of this knowledge graph construction module can be found in the descriptions in the foregoing embodiments, and will not be repeated here.

[0068] During retrieval, if the user's question vector is q, then a set of candidate fragments is obtained through entity expansion: ; Where: D represents the set of candidate knowledge fragments; d represents a candidate knowledge fragment; KB represents the vector knowledge base, which stores the embedded knowledge fragments; The knowledge graph is represented by V, where V is the set of entity nodes (e.g., “sleeper”, “settlement”, “turnout”), E is the set of relations (e.g., “settlement-cause-gauge change”); q represents the user question vector (corresponding to the above query question vector), which represents the question representation after semantic encoding. This represents the vector of the i-th entity node in the graph; The relevance function between the user question vector q and the i-th entity node vector can be cosine similarity or dot product similarity. This represents the semantic relevance threshold coefficient, used to control the strength of entity expansion, with a typical value range of [0.6, 0.9].

[0069] When satisfied When, it indicates the relationship between the user question vector q and the vector of the i-th entity node. If a knowledge fragment d is semantically significantly relevant, the system includes the related knowledge fragment d into the candidate knowledge fragment set D, achieving entity-based semantic expansion retrieval. This mechanism enables the system not only to recall semantically similar fragments but also to retrieve content with implicit connections to the question entity (such as chain-like reasoning of "sleeper settlement → gauge change → rail stress anomaly"), thereby significantly improving the accuracy and completeness of answer generation.

[0070] (6) Embedded model module The embedding model module uses a domain-optimized semantic embedding model to convert chunked text and graph nodes into vector representations.

[0071] (7) Vector knowledge base Used to store the embedded vector representation, supporting efficient similarity retrieval.

[0072] (8) Metadata storage module It is used to store metadata such as data source, timestamp, applicable lines, and standard number, so that it can be filtered and constrained during retrieval.

[0073] The online Q&A module receives user queries, parses the question's intent, and combines offline knowledge bases with real-time data to generate accurate answers. This module includes: (1) User query input unit Users input questions in natural language.

[0074] (2) Problem Analysis Module The system performs semantic understanding on user input, identifies intent categories (such as status query, fault diagnosis, and regulation query), extracts keywords and constraints, and obtains the parsing results.

[0075] (3) Search Intent Optimizer The retrieval intent optimizer dynamically selects retrieval strategies based on the parsed results. By combining question category, contextual information, metadata constraints, and semantic paths from the knowledge graph, it determines whether to prioritize retrieving real-time monitoring data, regulatory documents, or historical cases. This module breaks through the traditional single retrieval method based on vector similarity, significantly improving retrieval accuracy and result adaptability. Simultaneously, it performs feature comparison and pattern detection on the user-input question. If it is determined to be a high-frequency, hot topic question, it directly jumps to the hot topic cache module, querying from the hot topic knowledge base there, without performing a regular retrieval, thus shortening the response path. This ensures retrieval accuracy while shortening the process for high-frequency questions, improving question-answering efficiency.

[0076] First, through expressions Determine the target intent category. Then, perform dynamic search optimization, determining the search strategy based on the question category. The formula is as follows: ; in: This indicates that results will be retrieved primarily from the dynamic knowledge base corresponding to the real-time monitoring data stream; This indicates the search results retrieved from static documents of the regulations and standards category; This indicates the final retrieval strategy path.

[0077] This module can automatically switch search channels based on intent categories and optimize the retrieval order by combining knowledge graph path information to ensure that the source of the answer best matches the semantic requirements of the question.

[0078] Additionally, for high-frequency question identification, it can determine whether a user's question belongs to the high-frequency category. The formula is as follows: ; in: This represents a question similarity matching function, used to calculate the matching degree between the urban rail transit engineering question to be queried and the hot topic questions; This indicates the preset matching threshold, which is dynamically set based on system statistics (typical value 0.8–0.9). If... If the condition is not met, the process will directly switch to the hotspot caching module.

[0079] (4) Embedded model module The urban rail transit engineering problem to be queried is transformed into a vector representation for similarity matching.

[0080] (5) Search module By combining vector similarity, metadata filtering, and knowledge graph semantic expansion, candidate Top-K related document fragments are obtained.

[0081] (6) High-frequency hot topic knowledge caching module It maintains a hot topic knowledge base, which is used to retrieve the target question result corresponding to the urban rail transit engineering question when it is confirmed that the urban rail transit engineering question to be queried is a high-frequency question.

[0082] (7) Rearrangement module The retrieved candidate fragments are sorted according to relevance and contextual consistency, and selected document fragments (corresponding to the target knowledge data mentioned above) are output.

[0083] (8) Generative Model Module Based on selected snippets and user-submitted urban rail transit engineering questions, an answer generation model is used to generate answers, i.e., initial query results.

[0084] (9) Answer post-processing module The generated initial query results are validated for facts, annotated with references, and formatted for output to obtain the target query results.

[0085] (10) User output unit The target query results are returned to the user.

[0086] In this approach, the overall architecture of the urban rail transit engineering knowledge Q&A system consists of two core modules: offline knowledge base construction and online Q&A. The offline component first acquires data from the original data source and an innovative real-time data stream acquisition module. After unified data preprocessing and text segmentation, the data is processed through an embedding model and stored in a vector knowledge base. Simultaneously, the processed metadata is stored in dedicated storage. An innovative knowledge graph construction module is introduced to enhance semantic relationships. The online Q&A module receives user queries, first parsing the question, then using an innovative retrieval intent optimizer to deeply understand and optimize the query. The optimized query is then converted into a vector through the embedding model and enters the retrieval module, which retrieves Top-K relevant document fragments from the vector knowledge base and metadata storage. Furthermore, an innovative high-frequency hot topic knowledge caching module can quickly respond to common questions. The retrieval results are refined by a rearrangement module and sent to a generation model along with the original query to generate a preliminary answer. Finally, the answer is polished and formatted through a post-processing stage, ultimately returning a clear and accurate answer to the user. The entire process integrates multiple innovative designs to achieve efficient and accurate knowledge services.

[0087] The knowledge graph construction module ensures that the retrieval stage goes beyond semantic similarity, incorporating entity expansion based on the knowledge graph to improve recall and coverage. The real-time data stream acquisition module enables real-time dynamic incremental updates to the knowledge base, guaranteeing the timeliness and credibility of answers. The retrieval intent optimizer balances accuracy and efficiency through dynamic retrieval optimization and high-frequency question identification. The high-frequency hot topic knowledge caching module provides a fast hit mechanism, improving question-and-answer response speed and meeting the real-time needs of the scheduling center and frontline operations.

[0088] The aforementioned intelligent question-answering data processing method proposes a knowledge-based question-answering system for urban rail transit engineering. With the rapid development of artificial intelligence technology, large-scale language models (LLMs) have shown great potential in question-answering systems. However, traditional RAGs still have certain limitations in handling knowledge retrieval and real-time information acquisition in the urban rail transit engineering field. This approach, based on RAG technology and combined with an external urban rail transit engineering knowledge base, further enhances knowledge retrieval, user intent recognition, high-frequency knowledge acquisition, and real-time knowledge base updates through innovative methods such as knowledge graphs, dynamic incremental knowledge updates, search intent optimization, and hot topic knowledge caching. This significantly improves the accuracy, timeliness, and query result quality of the urban rail transit engineering knowledge-based question-answering system, accurately responding to user intent. In the field of urban rail transit engineering, the application of this patent will help improve information retrieval efficiency, fault diagnosis accuracy, and the intelligence level of maintenance decision-making.

[0089] To facilitate understanding of this solution, two implementation methods are provided below: Example 1: Troubleshooting Q&A Scenario During urban rail transit operation and maintenance, dispatchers entered the question: "How to handle the situation where the turnout cannot be switched?" (1) The system first determines that it is a "fault diagnosis" problem through the retrieval intent optimizer.

[0090] (2) The dynamic knowledge incremental update module extracts relevant records from the latest maintenance logs and monitoring data and adds them to the knowledge base.

[0091] (3) The retrieval module calls the vector database and combines it with the knowledge graph constraints to find document fragments related to “turnout jamming, mechanical failure and electrical failure”.

[0092] (4) Frequent hot topic knowledge cache returns common processing measures (such as manual reset steps).

[0093] (5) The final generation module comprehensively retrieves the content and returns a structured answer, including: possible causes (mechanical jamming / abnormal electrical signal) and recommended operations (check motor power supply, manual switching, report to dispatch).

[0094] This embodiment demonstrates how the system combines dynamic knowledge with knowledge graphs for accurate retrieval, ensuring the professionalism and practicality of the answers.

[0095] Example 2: Scenario for Inquiring about Rules and Regulations The maintenance personnel asked: "What is the safe distance during tunnel maintenance?" (1) The system identified it as a "regulation query" problem.

[0096] (2) The retrieval module calls the vector database and combines the knowledge graph to quickly locate the document entry of "Urban Rail Tunnel Maintenance Specification".

[0097] (3) The high-frequency hot topic knowledge caching module directly returns the cached relevant entries.

[0098] (4) The generation module outputs the answer: "According to Article X of the 'Urban Rail Tunnel Maintenance Specification', maintenance personnel should maintain a safe distance of not less than X meters from the running track." This embodiment demonstrates the system's efficient response capability in answering questions related to laws and regulations.

[0099] This invention provides an intelligent question-and-answer data processing device, such as... Figure 3 As shown, the device includes: an acquisition module 30 for acquiring a query question related to urban rail transit engineering; a parsing module 31 for parsing the query question to obtain a parsing result; an identification module 32 for identifying the intent of the query question based on the parsing result to obtain the target intent category to which the query question belongs; a judgment module 33 for judging whether the query question is a routine question; and a retrieval module 34 for retrieving, if the query question is a routine question, searching a preset urban rail transit engineering knowledge base based on the target intent category to obtain the target query result corresponding to the query question; wherein, the urban rail transit engineering knowledge base is generated based on a knowledge graph related to urban rail transit engineering.

[0100] The aforementioned intelligent question-answering data processing device, because the urban rail transit engineering knowledge base is generated based on a knowledge graph related to urban rail transit engineering, and entity expansion can be achieved based on the knowledge graph, can effectively improve the recall coverage when searching the urban rail transit engineering knowledge base based on the target intent category to which the question to be queried belongs, thereby improving the accuracy of the query results.

[0101] Furthermore, the urban rail transit engineering knowledge base is generated in the following manner: The original data source is obtained; dynamic data is collected in real time from multiple preset devices; the original data source and dynamic data are preprocessed according to a preset preprocessing method, and the data preprocessing results and metadata information corresponding to multiple knowledge data are output; the data preprocessing results are divided into blocks to obtain multiple blocks of data; based on the entities and relationships in each block of data, a knowledge graph related to urban rail transit engineering is constructed; vector transformation is performed on each block of data and the graph nodes of the knowledge graph to obtain data vectors corresponding to multiple knowledge data.

[0102] Furthermore, the identification module is also used to: obtain a preset set of intent categories; wherein the set of intent categories includes multiple preset intent categories; for each preset intent category, calculate the probability that the parsing result belongs to the preset intent category based on the parsing result; and determine the preset intent category with the highest probability as the target intent category to which the urban rail transit engineering question to be queried belongs.

[0103] Furthermore, the judgment module is also used to: obtain a preset hot topic knowledge base; wherein, the hot topic knowledge base includes: multiple hot topics and the corresponding results for each hot topic; for each hot topic, based on the parsing results, calculate the matching degree between the urban rail transit engineering question to be queried and the hot topic question; if at least one matching degree is greater than a preset matching degree threshold, determine that the urban rail transit engineering question to be queried is a high-frequency question; if each matching degree is less than or equal to the preset matching degree threshold, determine that the urban rail transit engineering question to be queried is a regular question.

[0104] Furthermore, the judgment module is also used to: if the urban rail transit engineering issue to be queried is a high-frequency issue, obtain the target issue result corresponding to the urban rail transit engineering issue to be queried from the hot topic knowledge base; and determine the target query result corresponding to the urban rail transit engineering issue to be queried based on the target issue result.

[0105] Furthermore, the retrieval module is also used for: vector transformation of the urban rail transit engineering question to be queried to obtain a question vector; filtering out multiple first data vectors from the data vectors corresponding to multiple knowledge data based on the question vector, target intent category, and metadata information; calculating the similarity between the question vector and the first data vector for each first data vector to obtain a vector similarity result; determining a preset number of target similarity results in descending order of vector similarity results; determining the first data vector corresponding to each target similarity result as a candidate data vector; and determining the target query result corresponding to the urban rail transit engineering question based on the knowledge data corresponding to the candidate data vectors.

[0106] Furthermore, the retrieval module is also used to: rearrange and filter the knowledge data corresponding to the selected data vector, and output the target knowledge data; input the target knowledge data and the urban rail transit engineering question to be queried into the preset answer generation model, and output the initial query results through the answer generation model; and perform answer post-processing on the initial query results according to the preset post-processing method to obtain the target query results corresponding to the urban rail transit engineering question to be queried.

[0107] The intelligent question-and-answer data processing device provided in this embodiment of the invention has the same implementation principle and technical effects as the aforementioned intelligent question-and-answer data processing method embodiment. For the sake of brevity, any parts not mentioned in the intelligent question-and-answer data processing device embodiment can be referred to the corresponding content in the aforementioned intelligent question-and-answer data processing method embodiment.

[0108] This invention also provides an electronic device, see [link to relevant documentation]. Figure 4 As shown, the electronic device includes a processor 130 and a memory 131. The memory 131 stores machine-executable instructions that can be executed by the processor 130, which executes the machine-executable instructions to implement the above-described XX method.

[0109] Furthermore, Figure 4 The electronic device shown also includes a bus 132 and a communication interface 133, with the processor 130, the communication interface 133 and the memory 131 connected via the bus 132.

[0110] The memory 131 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 133 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 132 may be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 4 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0111] Processor 130 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 130 or by instructions in software form. Processor 130 may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 131, and processor 130 reads the information in memory 131 and, in conjunction with its hardware, completes the steps of the method described in the foregoing embodiments.

[0112] This invention also provides a machine-readable storage medium storing machine-executable instructions. When these machine-executable instructions are invoked and executed by a processor, they cause the processor to implement the aforementioned intelligent question-and-answer data processing method. For specific implementation details, please refer to the method embodiments, which will not be repeated here.

[0113] The computer program products of the intelligent question-answering data processing method, apparatus and electronic device provided in the embodiments of the present invention include a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here.

[0114] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions 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. An intelligent question-answering data processing method, characterized by, The method comprises: acquiring a city rail maintenance problem to be queried; analyzing the city rail maintenance problem to be queried to obtain an analysis result; based on the analysis result, performing intent recognition on the city rail maintenance problem to be queried to obtain a target intent category to which the city rail maintenance problem to be queried belongs; determining whether the city rail maintenance problem to be queried is a regular problem; if the city rail maintenance problem to be queried is a regular problem, based on the target intent category, searching a preset city rail maintenance knowledge base to obtain a target query result corresponding to the city rail maintenance problem to be queried; wherein the city rail maintenance knowledge base is generated based on a knowledge graph associated with city rail maintenance.

2. The method of claim 1, wherein, The city rail maintenance knowledge base is generated in the following manner: acquiring an original data source; collecting dynamic data in real time from a plurality of preset device ends; preprocessing the original data source and the dynamic data in a preset preprocessing manner to output a data preprocessing result and metadata information corresponding to a plurality of knowledge data; performing block processing on the data preprocessing result to obtain a plurality of block data; based on entities and relationships in each block data, a knowledge graph associated with city rail maintenance is constructed; performing vector conversion on each block data and a graph node of the knowledge graph to obtain data vectors corresponding to a plurality of knowledge data.

3. The method of claim 1, wherein, The step of performing intent recognition on the city rail maintenance problem to be queried based on the analysis result to obtain a target intent category to which the city rail maintenance problem to be queried belongs comprises: acquiring a preset intent category set; wherein the intent category set includes a plurality of preset intent categories; for each preset intent category, calculating a probability that the analysis result belongs to the preset intent category according to the analysis result; determining the preset intent category with the highest probability as the target intent category to which the city rail maintenance problem to be queried belongs.

4. The method of claim 1, wherein, The step of determining whether the city rail maintenance problem to be queried is a regular problem comprises: acquiring a preset hot knowledge base; wherein the hot knowledge base includes a plurality of hot problems and a problem result corresponding to each hot problem; for each hot problem, calculating a matching degree between the city rail maintenance problem to be queried and the hot problem based on the analysis result; if there is at least one matching degree greater than a preset matching degree threshold, determining that the city rail maintenance problem to be queried is a high-frequency problem; if each matching degree is less than or equal to the preset matching degree threshold, determining that the city rail maintenance problem to be queried is a regular problem.

5. The method of claim 4, wherein, The method further comprises: if the city rail maintenance problem to be queried is a high-frequency problem, acquiring a target problem result corresponding to the city rail maintenance problem to be queried from the hot knowledge base; based on the target problem result, determining a target query result corresponding to the city rail maintenance problem to be queried.

6. The method of claim 2, wherein, The step of searching a preset city rail maintenance knowledge base based on the target intent category to obtain a target query result corresponding to the city rail maintenance problem to be queried comprises: performing vector conversion on the city rail maintenance problem to be queried to obtain a query problem vector; According to the to-be-queried question vector, the target intent category, and the metadata information, a plurality of first data vectors are screened out from a plurality of data vectors corresponding to a plurality of knowledge data; For each first data vector, a similarity between the to-be-queried question vector and the first data vector is calculated to obtain a vector similarity result; A preset number of target similarity results are determined in a descending order of the vector similarity results; Each first data vector corresponding to each target similarity result is determined as a to-be-selected data vector; Based on the to-be-selected data vector corresponding to the knowledge data, a target query result corresponding to the to-be-queried urban rail maintenance problem is determined.

7. The method of claim 6, wherein, The step of determining the target query result corresponding to the to-be-queried urban rail maintenance problem based on the knowledge data corresponding to the to-be-selected data vector comprises: The knowledge data corresponding to the to-be-selected data vector is rearranged and filtered to output target knowledge data; The target knowledge data and the to-be-queried urban rail maintenance problem are input into a preset answer generation model, and an initial query result is output by the answer generation model; The initial query result is processed by a preset post-processing manner to obtain the target query result corresponding to the to-be-queried urban rail maintenance problem. 8.An intelligent question-answering data processing apparatus, characterized by comprising: The device comprises: An acquisition module configured to acquire a to-be-queried urban rail maintenance problem; An analysis module configured to analyze the to-be-queried urban rail maintenance problem to obtain an analysis result; An identification module configured to identify an intent of the to-be-queried urban rail maintenance problem based on the analysis result to obtain a target intent category to which the to-be-queried urban rail maintenance problem belongs; A judgment module configured to judge whether the to-be-queried urban rail maintenance problem is a regular problem; A retrieval module configured to, if the to-be-queried urban rail maintenance problem is a regular problem, retrieve a preset urban rail maintenance knowledge base based on the target intent category to obtain a target query result corresponding to the to-be-queried urban rail maintenance problem, wherein the urban rail maintenance knowledge base is generated based on a knowledge graph associated with urban rail maintenance.

9. An electronic device, comprising: A processor and a memory, the memory stores machine executable instructions that can be executed by the processor, and the processor executes the machine executable instructions to implement the intelligent question and answer data processing method in any one of claims 1-7.

10. A machine-readable storage medium, characterized in that, The machine readable storage medium stores machine executable instructions, and when the machine executable instructions are called and executed by the processor, the machine executable instructions cause the processor to implement the intelligent question and answer data processing method in any one of claims 1-7.

Images