Adaptive complex scene information analysis method, electronic device, storage medium and program product

Abstract

The present disclosure provides an adaptive complex scene information analysis method, electronic device, storage medium and program product, the method comprising: acquiring scene data from a plurality of data sources respectively through a data source module; converting the acquired scene data into a preset data format through a data conversion module, and storing the data in a data storage module; querying, in response to receiving an analysis request, from a history storage module historical analysis requests whose similarity with the analysis request is greater than a first similarity threshold through an analysis module, and resolving the analysis request into a query task with reference to analysis records of the queried historical analysis requests; calling, through an adaptive query module, a query tool in a query tool library according to the query task, performing a data query on the data storage module to obtain query result data; and analyzing and processing the query result data through the analysis module to obtain analysis result data.

Description

Technical Field

[0001] The present disclosure generally relates to the field of data analysis technology, and more specifically, to an adaptive complex scene information analysis method, electronic device, storage medium, and program product. Background Art

[0002] In complex scenarios like emergency management, smart cities, and enterprise risk management, data analysis faces multiple challenges, including large data volumes, high heterogeneity, strong real-time requirements, and complex event correlations. Traditional analytical methods often rely on fixed rules for data query and analysis, which makes it difficult to meet the demand for efficient, intelligent, and adaptive analysis in these scenarios, hindering the intelligent development of emergency management, smart cities, and other fields.

[0003] Therefore, it is urgent to develop an intelligent analysis system that can integrate multi-source data, support real-time analysis, and have adaptive learning capabilities. Summary of the Invention

[0004] The present disclosure provides an adaptive complex scene information analysis method, electronic device, storage medium and program product, which are used to solve at least one of the above problems.

[0005] According to a first aspect of an embodiment of the present disclosure, an adaptive complex scene information analysis method is provided. The adaptive complex scene information analysis method is used in an information analysis system, wherein the information analysis system includes a data source module, a data conversion module, a data storage module, an analysis module, a query tool library, an adaptive query module, and a historical storage module for storing analysis records of historical analysis requests. The adaptive complex scene information analysis method includes: obtaining scene data from multiple data sources respectively through the data source module; converting the obtained scene data into a preset data format through the data conversion module and storing it in the data storage module; in response to receiving an analysis request, the analysis module queries the historical storage module for historical analysis requests whose similarity with the analysis request is greater than a first similarity threshold, and resolves the analysis request into a query task with reference to the analysis records of the queried historical analysis requests; calling the query tool in the query tool library through the adaptive query module according to the query task, performing a data query on the data storage module to obtain query result data; and analyzing and processing the query result data through the analysis module to obtain analysis result data.

[0006] Optionally, the acquired scene data is converted into a preset data format by the data conversion module and stored in the data storage module, including: for the scene data of each object, the object identifier of the object is generated as a key by the data conversion module, and the scene data of the object is used as a value, and stored in the data storage module in the format of a key-value pair.

[0007] Optionally, in response to receiving the analysis request, the analysis module queries the historical analysis requests whose similarity with the analysis request is greater than a first similarity threshold from the historical storage module, and resolves the analysis request into a query task with reference to the analysis records of the queried historical analysis requests, including: in response to receiving the analysis request, the analysis module evaluates whether the analysis request is a complex request, wherein the complex request is an analysis request that needs to be resolved into at least N query tasks, and N is an integer greater than 1; if the analysis request is not the complex request, resolving the analysis request into a query task; if the analysis request is the complex request, querying the historical analysis requests whose similarity with the analysis request is greater than the first similarity threshold from the historical storage module, and resolving the analysis request into the query task with reference to the analysis records of the queried historical analysis requests.

[0008] Optionally, when the analysis request is the complex request, querying the historical analysis request whose similarity with the analysis request is greater than the first similarity threshold from the historical storage module, and resolving the analysis request into the query task with reference to the analysis record of the queried historical analysis request, including: when the analysis request is the complex request, determining whether there is a historical analysis request that meets preset conditions in the historical storage module, wherein the preset conditions include: the difference between the request time of the corresponding historical analysis request and the request time of the analysis request is less than the time difference threshold, and the similarity between the corresponding historical analysis request and the analysis request is greater than a second similarity threshold, wherein the second similarity threshold is greater than the first similarity threshold; if there is no historical analysis request that meets the preset conditions, querying the historical analysis request whose similarity with the analysis request is greater than the first similarity threshold from the historical storage module, and resolving the analysis request into the query task with reference to the analysis record of the queried historical analysis request; if there is a historical analysis request that meets the preset conditions, abandoning the query task parsing, and obtaining analysis result data of the historical analysis request that meets the preset conditions as the analysis result data of the analysis request.

[0009] Optionally, parsing the analysis request into a query task includes: using a large language model to parse the analysis request into the query task; and / or using a predefined template to parse the analysis request into the query task; and / or parsing the analysis request into the query task based on a knowledge graph.

[0010] Optionally, calling the query tool in the query tool library according to the query task through the adaptive query module includes: determining the query tool requirement according to the query task through the adaptive query module, wherein the query tool requirement represents the function of the query tool to be called; retrieving the query tool that meets the query tool requirement from the query tool library; and dynamically generating the query tool that meets the query tool requirement in case of retrieval failure.

[0011] Optionally, the query tool library also records the call information of the dynamically generated query tools, and the call information includes the number of calls and the call time. The adaptive complex scene information analysis method also includes: for the query tools dynamically generated in the query tools, determining the call frequency of the corresponding query tools according to the call information; deleting the dynamically generated query tools whose call frequency meets the preset deletion conditions, wherein the preset deletion conditions are used to represent low-frequency calls.

[0012] Optionally, the information analysis system also includes a domain knowledge base, wherein the adaptive query module calls the query tool in the query tool library according to the query task, performs data query on the data storage module, and obtains query result data, including: calling the query tool in the query tool library according to the query task by the adaptive query module to query the data storage module as preliminary query data; analyzing whether the preliminary query data meets the analysis requirements of the analysis request; if the analysis requirements of the analysis request are met, using the preliminary query data as the query result data; if the analysis requirements of the analysis request are not met, analyzing and reasoning the preliminary query data in combination with the data in the data storage module and the domain knowledge base by the analysis module to obtain an updated query task, re-calling the query tool in the query tool library by the adaptive query module according to the updated query task, performing data query on the data storage module to update the preliminary query data, and repeating the step of analyzing whether the preliminary query data meets the analysis requirements of the analysis request until the query result data is obtained or the maximum number of updates is reached.

[0013] According to a second aspect of an embodiment of the present disclosure, an adaptive complex scene information analysis device is provided. The adaptive complex scene information analysis device is used in an information analysis system. The information analysis system includes a data source module, a data conversion module, a data storage module, an analysis module, a query tool library, an adaptive query module, and a history storage module for storing analysis records of historical analysis requests. The adaptive complex scene information analysis device includes: an acquisition unit configured to respectively acquire scene data from multiple data sources through the data source module; a conversion unit configured to convert the acquired scene data into a preset data format through the data conversion module and store the data in the data storage module; a parsing unit configured to, in response to receiving an analysis request, query the history storage module for historical analysis requests whose similarity with the analysis request is greater than a first similarity threshold, and parse the analysis request into a query task with reference to the analysis records of the queried historical analysis requests; a query unit configured to, through the adaptive query module, call a query tool in the query tool library according to the query task, perform a data query on the data storage module to obtain query result data; and an analysis unit configured to analyze and process the query result data through the analysis module to obtain analysis result data.

[0014] According to a third aspect of an embodiment of the present disclosure, an electronic device is provided, comprising: at least one processor; and at least one memory storing computer-executable instructions, wherein the computer-executable instructions, when executed by the at least one processor, prompt the at least one processor to execute an adaptive complex scene information analysis method according to an exemplary embodiment of the present disclosure.

[0015] According to a fourth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided. When the instructions in the computer-readable storage medium are executed by at least one processor, the at least one processor is prompted to execute the adaptive complex scene information analysis method according to the exemplary embodiment of the present disclosure.

[0016] According to a fifth aspect of an embodiment of the present disclosure, a computer program product is provided, comprising computer instructions, which, when executed by at least one processor, prompt the at least one processor to execute an adaptive complex scene information analysis method according to an exemplary embodiment of the present disclosure.

[0017] The technical solutions provided by the embodiments of the present disclosure bring at least the following beneficial effects: According to the adaptive complex scene information analysis method, electronic device, storage medium and program product disclosed in the present disclosure, by obtaining scene data from multiple data sources and uniformly converting and storing the data, it is possible to effectively integrate multi-source data, which helps to improve data query efficiency. By storing analysis records of different analysis requests that have been processed previously in the historical storage module, searching for similar historical analysis requests when an analysis request is received, and parsing the currently received analysis request with reference to the analysis records of the historical analysis requests, it is possible to adaptively learn how to optimize task parsing and obtain query tasks that better meet the needs, which helps to reduce uncertainty in the reasoning process and query the required scene data to meet the real-time analysis needs of complex scenes.

[0018] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The accompanying drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure, and do not constitute an improper limitation of the present disclosure.

[0020] Figure 1 is a flowchart of an adaptive complex scene information analysis method according to an exemplary embodiment of the present disclosure.

[0021] Figure 2 FIG. 4 is a general architecture diagram of an information analysis system according to an exemplary embodiment of the present disclosure.

[0022] Figure 3 is a flowchart of adaptive query processing according to a specific embodiment of the present disclosure.

[0023] Figure 4 is a block diagram of an adaptive complex scene information analysis device according to an exemplary embodiment of the present disclosure.

[0024] Figure 5 is a block diagram of an electronic device according to an exemplary embodiment of the present disclosure. DETAILED DESCRIPTION

[0025] In order to enable ordinary persons in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0026] It should be noted that the terms "first," "second," and the like in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or precedence. It should be understood that the numbers used in this manner are interchangeable where appropriate so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The implementation methods described in the following examples do not represent all implementation methods consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with certain aspects of the present disclosure as detailed in the appended claims.

[0027] It should be noted that the phrase "at least one of the several items" in this disclosure includes three types of parallel situations: "any one of the several items", "a combination of any multiple of the several items", and "all of the several items". For example, "including at least one of A and B" includes the following three parallel situations: (1) including A; (2) including B; (3) including A and B. For another example, "performing at least one of step 1 and step 2" means the following three parallel situations: (1) performing step 1; (2) performing step 2; and (3) performing both step 1 and step 2.

[0028] In areas such as emergency management, smart cities, and enterprise risk control, data analysis systems must process multi-source heterogeneous data to enable real-time analysis and dynamic reasoning. However, existing technologies still have many shortcomings in data integration, real-time response, interactive query, and complex event correlation reasoning, making it difficult to meet the needs of highly dynamic and uncertain scenarios.

[0029] Next, a variety of existing analytical techniques are introduced one by one. It should be noted that the following is only an introduction to existing technologies and does not imply that the analytical method disclosed in this disclosure can solve all the problems mentioned below.

[0030] Traditional model-driven approaches analyze data using predefined rules or mathematical models, such as optimizing evacuation routes based on traffic flow or predicting fire spread using fire spread models. While these approaches offer clear logic and interpretable results, their applicability is limited to specific domains and they struggle to adapt flexibly to emergencies and complex environments. Furthermore, they lack the ability to integrate multimodal data, making them particularly difficult to process unstructured data such as social media feeds or video surveillance data. Furthermore, these approaches often rely on manually defined parameters, making it difficult to automatically adjust analysis strategies based on real-time data.

[0031] Traditional machine learning methods rely on historical data to train models for pattern recognition, classification, and prediction, such as traffic congestion forecasting or optimizing emergency resource allocation. While these methods can uncover underlying patterns in data and improve prediction accuracy, they rely heavily on high-quality datasets, making them inadequate for scenarios with scarce data or emergencies. Furthermore, existing machine learning methods lack the ability to deeply model complex relationships between events, making them unable to adapt to new situations. Furthermore, traditional machine learning models are primarily trained on batch data, making them inadequate for online learning and real-time analysis.

[0032] Digital twin technology builds digital models of cities or facilities, enabling real-time monitoring, analysis, and simulation. For example, disaster monitoring systems use sensor data to construct disaster scenarios. However, this approach is costly to deploy and relies on high-precision data collection and fixed visualization models, making it difficult to quickly build models in response to emergencies. Furthermore, digital twin systems are typically based on fixed data schemas and cognitive models, making them difficult to learn from new environmental changes in real time and limiting their ability to predict the evolving trends of complex events.

[0033] With the development of large language models (LLMs), large-model-based intelligent analysis technologies are widely used in complex data analysis, intelligent query, and situational awareness scenarios. This approach can combine multimodal data to achieve comprehensive perceptual analysis of text, images, and sensor information. However, existing large-model-based methods primarily rely on fixed prompt word templates or specially trained domain models, lacking adaptive fine-grained task decomposition and optimization mechanisms, making it difficult to dynamically adjust analysis strategies based on the needs of different scenarios.

[0034] Large-scale inference model technology, exemplified by models like OpenAI's o1, demonstrates strong capabilities in intelligent analysis, dynamic querying, and decision-making reasoning, but limitations remain. First, when reasoning over long texts and facing complex task chains, these models may suffer from contextual information loss and decreased reasoning consistency. Second, dynamic task decomposition still relies on static prompts and rule templates, lacking adaptive optimization capabilities for diverse analysis scenarios. Third, multimodal data fusion accuracy is limited, particularly for cross-modal understanding of image, text, and sensor data, which still requires additional data preprocessing or specialized model support. Fourth, tool invocation lacks intelligent intelligence, relying primarily on static matching and single-step invocation, making multi-round task optimization difficult. This limits the usability and efficiency of complex queries. In highly dynamic and uncertain application scenarios, more robust adaptive task planning, intelligent tool management, and associative inference optimization techniques are needed to enhance model practicality and decision accuracy.

[0035] Large-model tool invocation technology utilizes a large language model to invoke tools, providing greater flexibility in data query, analysis, and complex task execution. Existing methods primarily rely on natural language matching functions for tool invocation, but semantic biases can lead to incorrect invocations, compromising query accuracy. Furthermore, most methods rely on static tool libraries or simple registration mechanisms, lacking dynamic optimization capabilities. This makes it difficult to effectively control tool invocation frequency, impacting computing resource utilization. Even systems such as AutoGPT offer some dynamic tool creation capabilities, but these approaches often employ crude tool management, failing to effectively control tool invocation frequency, performance evaluation, and update strategies. This results in redundant tools occupying resources while critical tools are not prioritized, hindering their practical application. Furthermore, existing methods primarily rely on single-query triggering, making it difficult to automatically decompose complex tasks or optimize query order, resulting in inefficient analysis and wasted computing resources. In multi-round interactive query scenarios, insufficient state preservation capabilities lead to poor coherence between consecutive analysis tasks, impacting the comprehensiveness and accuracy of decision-making.

[0036] In summary, existing technologies still have many shortcomings in complex scenario analysis. Traditional model-driven methods have poor adaptability and are difficult to handle dynamic environments; machine learning methods rely on high-quality data and lack associative reasoning and adaptive adjustment capabilities; digital twin technology is limited in real-time updates during emergencies; and although large models have improved multimodal data integration and intelligent analysis capabilities, existing methods rely on fixed prompt words or static models, making it difficult to perform fine-grained task decomposition and dynamic optimization. In addition, current large model tool calling technology still lacks accurate tool matching, dynamic management, and multi-round interaction support, making it difficult to efficiently execute complex queries and associative reasoning. Overall, existing methods find it difficult to achieve real-time, accurate, and adaptive analysis and decision support in highly heterogeneous and dynamically changing scenarios, which restricts the intelligent development of emergency management, smart cities, and other fields.

[0037] Hereinafter, an adaptive complex scene information analysis method, an electronic device, a storage medium, and a program product according to exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings.

[0038] Figure 1 The present invention is a flowchart of an adaptive complex scene information analysis method according to an exemplary embodiment of the present invention. The method can be executed on an electronic device with sufficient computing power, and an information analysis system is configured on the electronic device. Figure 2 FIG. 4 is a general architecture diagram of an information analysis system according to an exemplary embodiment of the present disclosure.

[0039] Reference Figure 2 The information analysis system includes a data source module, a data conversion module, a data storage module (for example, a cache module, corresponding to Figure 2Data cache in), analysis module, query tool library, adaptive query module and historical storage module (for example, cache form, corresponding Figure 2 The data source module is mainly used to obtain scene data from multiple data sources; the data conversion module is mainly used to convert data formats; the data storage module is mainly used to store the data after format conversion; the analysis module (also known as the analysis agent) is mainly used to obtain and parse analysis requests, and hand over the query tasks obtained by the parsing to the adaptive query module (also known as the adaptive query agent). The adaptive query module calls the query tools in the query tool library to query the required data from the scene data stored in the data storage module, and then hands it over to the analysis module for analysis and processing to obtain analysis result data (for example, query results or a summary of query results). The historical storage module is mainly used to store analysis records of historical analysis requests, which can provide a reference for the task parsing of the analysis module.

[0040] like Figure 2 As shown, the analysis module can also be connected to the user interaction module and the downstream business module to interact with the user and the downstream business module to obtain analysis requests and deliver analysis result data.

[0041] Specifically, users can initiate analysis requests through the user interaction module to obtain intelligent summaries or query data. Users can also adjust the analysis direction of their requests based on the feedback provided by the system. The results of the interaction between users and the analysis module can be further used as analysis input to optimize the behavior of the intelligent agent, thereby achieving the effect of active learning and improvement of the system, making it more in line with actual business needs.

[0042] The downstream business module represents the scalability of the system, allowing the analysis result data to be used in other business systems (such as scheduling systems, decision support systems, etc.).

[0043] It should be understood that the user interaction module and the downstream business module can exist at the same time or be used selectively, and this disclosure does not impose any restrictions on this. As an example, if the system is mainly used for expert decision support, the external interface can be omitted, and only user visual interaction functions can be provided. This approach can simplify the system architecture and reduce the complexity of external system integration. In highly automated systems (such as intelligent scheduling systems), users often do not directly operate the information analysis system. Instead, the downstream business module receives the analysis result data and automatically performs related operations, such as adjusting the deployment of rescue resources and optimizing logistics routes. This approach can improve the system's degree of automation, but reduces the user's flexibility in adjusting the analysis strategy.

[0044] As will be introduced later, the information analysis system can further include Figure 2 The domain knowledge base in , in order to provide more information reference, will not be expanded here.

[0045] Next, an adaptive complex scene information analysis method according to an exemplary embodiment of the present disclosure is introduced.

[0046] Reference Figure 1 In step S101, scene data is obtained from multiple data sources respectively through the data source module.

[0047] In complex information analysis scenarios, data sources are often highly heterogeneous, including structured data such as sensors and dispatch systems, unstructured data such as social media and news reports, and multimodal information such as images and videos. These data formats vary and are updated at inconsistent frequencies. Traditional analysis methods are difficult to integrate efficiently, resulting in information fragmentation, which in turn affects the accuracy and timeliness of decision-making. The data source module of the present disclosure can continuously collect situational data from multiple data sources (for example, including but not limited to sensors, monitoring systems, dispatch systems, social media, etc.) to support decision-making.

[0048] In step S102, the acquired scene data is converted into a preset data format by a data conversion module and stored in a data storage module.

[0049] The data conversion module standardizes the acquired scene data (i.e., converts it to a preset data format) to conform to unified data format specifications, effectively integrating multi-source data and improving data query efficiency. This process includes parsing structured data, extracting unstructured text using natural language processing (NLP), and converting multimodal data (such as videos and images). For multimodal data, large multimodal models (such as GPT4o and QwenVL) can be combined to generate corresponding natural language descriptions for subsequent query and analysis.

[0050] As an example, a large multimodal model can be used to process scene data, including structured data, unstructured data, and visual data (images, videos). In some scenarios, other smaller models can be used to replace the large multimodal model to reduce computing resource overhead. For example, if the input scene data is mainly text (such as social media, reports, logs, etc.), you can use a pure large language model to process all text data instead of a large multimodal model. This can reduce the cost of multimodal training and inference, but will sacrifice the ability to understand visual data. For some stable visual scenes (such as disaster detection, target recognition, etc.), a dedicated visual model can be used to replace the large multimodal model to reduce computational complexity.

[0051] The converted scene data is stored in the data storage module. The data storage module can support efficient indexing and retrieval mechanisms (such as high-speed storage based on Redis) to reduce direct queries to the original data source and improve query efficiency. In addition, through the TTL (Time-to-Live) mechanism of Redis, it supports automatic expiration and cleanup of data, which can ensure the management efficiency of real-time data and prevent redundant or outdated data from occupying storage resources. As an example, the data storage module specifically adopts the form of cache, corresponding to Figure 2 The data cache in the system helps better meet the real-time requirements of information analysis. It specifically includes a scene data cache and a scene description cache. The former is used to store scene data obtained from different data sources and serves as the primary target for data queries. The latter is used to store key information and / or descriptive information extracted from scene data at different time periods. This extracted information can be applied to scenarios such as trend analysis and event backtracking. This enables the system to not only process current situations but also use historical data for inference and prediction, allowing the system to conduct deeper analysis based on historical data. The data storage module can also be combined with the vector database (Facebook AI Similarity Search, or FAISS) for data storage, or other solutions can be selected based on storage and performance requirements. For example, if the analysis task does not require high-frequency queries but relies primarily on historical data, it can be combined with persistent storage (such as MySQL), using a database such as MySQL as the primary storage, and combining it with Redis or other high-speed caching mechanisms as a cache for hot data. This approach is suitable for long-term data storage and reduces memory usage, but the query speed may be slower. For example, for applications that rely on similarity search (such as situation analysis and case matching), only a vector database (such as Milvus) can be used without Redis. This approach is suitable for scenarios that require efficient retrieval of similar events, but is not suitable for data storage with high-frequency updates.

[0052] In addition, in the related art, the correlation between different data sources is weak, making it difficult to unify modeling and real-time analysis, which reduces the efficiency of data utilization. In some embodiments of the present disclosure, optionally, step S102 includes: for the scene data of each object, the object identifier of the object is generated as a key through the data conversion module, and the scene data of the object is used as a value, and stored in the data storage module in the format of a key-value pair. By taking the object as a clue, a unique object identifier is generated for each object as a key (Key), and the scene data for the same object is used as a value (Value), and stored in a structured form of a key-value pair (KV pair). This storage method can adapt to heterogeneous data types, while maintaining good scalability, it also provides efficient query and aggregation capabilities, which helps to improve the efficiency of data utilization.

[0053] The following is an example of a key-value pair:

[0054]

[0055] In this example, the key is the object ID of the tent, rescue_tent:001, and the value is the scenario data related to the tent, including its name, type, location, inventory level, and demand level.

[0056] In step S103, in response to receiving the analysis request, the analysis module searches for historical analysis requests whose similarity with the analysis request is greater than a first similarity threshold from the history storage module, and refers to the analysis records of the searched historical analysis requests to parse the analysis request into a query task.

[0057] The user interaction module or downstream business module can initiate analysis requests at any time. These requests may involve data query, trend prediction, event backtracking, risk assessment and other needs. The system can parse the analysis request, refer to similar historical analysis requests and their analysis records in the historical storage module, and combine the tool functions in the query tool library to determine the optimal query method to obtain the query task. For complex requests involving multiple query targets or different data sources, they can be broken down into multiple query tasks and executed in sequence. By storing the analysis records of different analysis requests that have been processed before in the historical storage module, searching for similar historical analysis requests when receiving an analysis request, and parsing the currently received analysis request with reference to the analysis records of the historical analysis request, it can adaptively learn how to optimize task parsing and obtain query tasks that better meet the needs, which helps to reduce the uncertainty in the reasoning process and query the required scene data to meet the real-time analysis needs of complex scenes.

[0058] For example, when calculating the similarity between two analysis requests, the similarity between the embedded features of the corresponding request texts can be calculated, such as cosine similarity. Furthermore, a reasonable first similarity threshold can be set to ensure that historical analysis requests with a similarity greater than the first similarity threshold are found.

[0059] In step S104, the adaptive query module calls the query tool in the query tool library according to the query task, performs data query on the data storage module, and obtains query result data.

[0060] The adaptive query module is a crucial component of the system and the core of the dynamic query mechanism. Based on query tasks parsed by the analysis module (which can be executed by sending query requests to the adaptive query module), it registers / deregisters specific query patterns and invokes appropriate query tools (such as, but not limited to, query functions) from a query tool library to intelligently query the scene data cached in the data storage module, effectively responding to user interaction or analysis needs or requests from downstream business modules. Query tasks or query requests can, for example, be in natural language. The adaptive query module supports dynamic registration of query tools to ensure that query logic can flexibly adapt to changing business needs. For example, the adaptive query module can adjust query parameters based on historical analysis records and current scene data to improve the accuracy and real-time nature of query results. This approach maintains efficiency in complex analytical tasks while enhancing decision support capabilities and making the query process more intelligent and adaptive.

[0061] The query tool library contains various query tools that can be called by the adaptive query module to support various query requirements. It allows query tools to be registered or deregistered on demand, enabling the system to automatically adjust query capabilities as analytical requirements change.

[0062] In step S105, the query result data is analyzed and processed by the analysis module to obtain analysis result data.

[0063] The query result data is passed to the analysis module for summary and refinement based on user needs, historical scenario descriptions, domain knowledge and other information to ensure that the output information is focused, logically clear and readable.

[0064] As an example, for different types of analysis requests, summaries of different granularities can be provided, including but not limited to overview summaries and detailed analyses. The former can extract key information and provide main conclusions and recommendations, while the latter can list relevant data, historical cases, reasoning processes, and attach decision-making reference recommendations.

[0065] As an example, a large language model can be used to generate an intelligent summary as the analysis result data, or an automatic summary based on keyword extraction can be used, such as using TF-IDF (Term Frequency-Inverse Document Frequency), TextRank (graph-based ranking model) and other algorithms to extract key information without relying on a large language model, which can reduce computing costs. A pre-trained summary model can also be used to generate a summary, such as using T5 (Text-to-Text Transfer Transformer), BERT (Bidirectional Encoder Representations from Transformers) and other models to generate a summary without relying on a general large language model. This method can improve the summary speed in specific fields. In actual implementation, the generation method can be selected as needed, and this disclosure does not limit this.

[0066] The final analysis summary and query results are returned to the requester (including the user and / or downstream business modules). Based on the results, the requester can further adjust the query requirements, such as refining the query conditions (such as focusing on risk changes in a specific area) or exploring the causal relationship between related events (such as analyzing the potential impact of a disaster on the supply chain).

[0067] If the system cannot fully satisfy a requested analysis request, it can provide specific reasons for the failure and suggested solutions. For example, if data is insufficient, the system can suggest adjusting the query scope or providing more relevant input. If query tools are limited, the system can recommend more advanced analytical methods or dynamically invoke new tools. Furthermore, this feedback can be simultaneously communicated to relevant users or downstream business modules to further optimize business decision-making processes.

[0068] The system also uses a historical storage module to record analysis requests, their corresponding analysis results, and feedback from the requester to optimize subsequent query processes. This adaptive mechanism ensures that the system can continuously optimize query logic over the long term, making it more aligned with actual business needs and improving intelligent analysis capabilities and overall operational efficiency.

[0069] Next, step S103 is further introduced, that is, how to parse and analyze the request to obtain the query task.

[0070] In some embodiments, optionally, step S103 includes: evaluating, by the analysis module, in response to receiving the analysis request, whether the analysis request is a complex request, wherein a complex request is an analysis request that needs to be parsed into at least N query tasks, where N is an integer greater than 1; if the analysis request is not a complex request, parsing the analysis request into a query task; if the analysis request is a complex request, querying from the history storage module historical analysis requests whose similarity with the analysis request is greater than a first similarity threshold, and referencing the analysis records of the retrieved historical analysis requests to parse the analysis request into a query task. By not querying and using historical analysis requests for simple requests and directly performing task parsing, and querying and referencing historical analysis requests for complex requests for task parsing, the computing resources consumed by querying historical analysis requests can be reduced. Of course, in other embodiments, historical analysis requests can also be queried and referenced for simple requests. Alternatively, based on the actual situation of the simple request, historical analysis requests may not be queried and referenced when a query task can be directly parsed, and historical analysis requests may be queried and referenced when a query task cannot be directly parsed. This is also an implementation of the present disclosure.

[0071] Further optionally, in step S103, when the analysis request is a complex request, historical analysis requests whose similarity with the analysis request is greater than a first similarity threshold are searched from the historical storage module, and the analysis request is parsed into a query task with reference to the analysis records of the queried historical analysis requests. The operation includes: when the analysis request is a complex request, determining whether there is a historical analysis request that meets preset conditions in the historical storage module, wherein the preset conditions include: the difference between the request time of the corresponding historical analysis request and the request time of the analysis request is less than the time difference threshold, and the similarity between the corresponding historical analysis request and the analysis request is greater than a second similarity threshold, wherein the second similarity threshold is greater than the first similarity threshold; if there is no historical analysis request that meets the preset conditions, historical analysis requests whose similarity with the analysis request is greater than the first similarity threshold are searched from the historical storage module, and the analysis request is parsed into a query task with reference to the analysis records of the queried historical analysis requests; if there is a historical analysis request that meets the preset conditions, the query parsing task is abandoned, and the analysis result data of the historical analysis request that meets the preset conditions is obtained as the analysis result data of the analysis request. By further configuring preset conditions for complex requests to represent the same or highly similar historical analysis requests that have been processed recently, and prioritizing the search for historical analysis requests that meet the preset conditions, the analysis result data can be reused first, eliminating the need to perform task parsing and subsequent data query and analysis, thereby significantly reducing the consumption of computing resources and greatly improving the response rate of information analysis.

[0072] As an example, regarding the first condition in the preset conditions, namely the time condition, for different types of analysis requests, different time difference thresholds can be set according to different sensitivities to time, so as to obtain historical analysis requests with more reference value and ensure the reliability of the reused analysis result data.

[0073] As an example, if a historical analysis request is found that meets the preset condition of a similarity greater than a second similarity threshold but does not meet the time condition, that is, an earlier identical or highly similar historical analysis request is found, then on the one hand, the historical analysis request can be used as a reference for parsing the query task, and on the other hand, after obtaining the final analysis result data in the subsequent step S105, the content that is different from the analysis result data of the historical analysis request can be marked in the analysis result data to highlight the change in the situation.

[0074] As an example, additional conditions can be added to the preset conditions as needed to meet more needs and enhance the flexibility of the solution. For example, specific context information can be configured for different types of analysis requests, and the preset conditions can include the similarity between the specified context information of the corresponding historical analysis request and the specified context information of the analysis request being greater than a third similarity threshold to improve the reliability of the reused analysis results.

[0075] In summary, the history storage module is a core module in the system specifically used to store, reuse, and optimize historical analysis requests. It not only improves query efficiency and reduces repeated calculations, but also provides contextual information for multiple rounds of conversations, enhancing the intelligence and coherence of the system in interactive analysis.

[0076] During system operation, the historical storage module stores past analysis requests, analysis results, and task breakdown records. This information is combined with the parsed query tasks, query results, and data sources to support intelligent reuse. Furthermore, the historical storage module records the query tools used and user feedback on the output analysis results to optimize future task breakdown and query strategies.

[0077] As an example, the history storage module specifically adopts the cache form, corresponding to Figure 2 The historical cache in the system helps better meet the real-time requirements of information analysis. To improve storage and query efficiency, the historical storage module can be implemented by combining a vector database (such as FAISS) with a memory queue. The vector database is responsible for storing the embedded vectors of analysis requests and supports efficient similarity queries, enabling the system to quickly retrieve similar historical analysis records and improve the accuracy of analysis strategies. The memory queue is mainly used to store recent analysis requests and results, such as the most recent set time period or a certain number of recent analysis requests. It also provides contextual information for multiple rounds of interactions to support continuous dialogue and help ensure the consistency of continuous queries.

[0078] In addition to combining a vector database with an in-memory queue, the historical storage module can also be implemented using other methods as needed. For example, for systems with a relatively small history of analysis requests, log file-based analysis reuse can be used to directly record request logs and parse them using a rules engine (such as the ELK Stack), replacing vector database storage. Another example is that since some large language models (such as GPT-4o) support long-term memory, historical analysis requests can be directly stored through an API (Application Programming Interface) based on the memory mechanism of the large language model, rather than using a local cache. This approach is suitable for SaaS (Software as a Service) deployments, but relies on an external large language model service and makes it difficult to flexibly customize the memory strategy.

[0079] When executing a new analysis request, the system will first query the historical storage module to determine whether existing analysis results can be directly reused or incrementally updated based on historical data. For example, when querying the rescue situation in a certain area, if the system finds that current weather, infrastructure status, and other information have not changed significantly in the recent period, it can reuse historical descriptions, optimizing computing resources and speeding up analysis.

[0080] The historical storage module also features change detection capabilities, automatically identifying key change points by comparing current query data with historical cached records. For example, if the system detects significant fluctuations in relief supply inventory or the addition of new high-risk areas, it automatically annotates these changes, allowing users to quickly identify the latest developments when accessing analysis results. The vector database's similarity calculation, combined with a time window mechanism, effectively extracts the most impactful changes in the short term, ensuring real-time and accurate analysis.

[0081] To improve storage management efficiency, the historical storage module regularly cleans up redundant or infrequently used data. For example, it uses a TTL mechanism to control the data lifecycle and combines the LRU (Least Recently Used) algorithm to dynamically optimize the records of historical analysis requests to ensure efficient use of storage space and the timeliness of the analysis result data of historical analysis requests.

[0082] When executing step S103 in conjunction with the historical storage module, the system first assesses the complexity of an analysis request submitted by a user or downstream business module to determine the optimal processing method. Simple requests, such as "Query the current location of a rescue vehicle" or "Get the current inventory at a warehouse," can be directly parsed into a single query task. This query is then accessed through the data storage module or the query tool library without further decomposition, improving response speed and reducing computing resource consumption.

[0083] However, for more complex analysis requests such as "obtaining the overall situation of the current rescue scene", the system needs to perform a deeper analysis. To this end, the historical storage module can be queried first to check whether there are recent analysis records, so as to determine whether the existing analysis results can be reused. If it is found that the same or highly similar analysis requests have been processed recently, the historical analysis results can be returned directly to avoid repeated calculations. If the recent analysis records cannot be reused directly, a similarity query can be further performed in the vector database to find historical analysis records that are close to the content of the current analysis request. Through cosine similarity calculation, the system can quickly filter out the most relevant analysis cases and optimize the current analysis strategy based on these historical data. For example, if the current analysis request is "analyze the impact of floods in a certain area", the system may find past analysis records for adjacent areas or similar disasters and use them as references to reduce uncertainty in the reasoning process.

[0084] In complex scenarios, due to the massive volume of data, especially in sudden or emergency situations, relevant information accumulates rapidly, making it difficult for traditional analysis tools to filter and process it in a timely manner. This requires decision makers to spend a significant amount of time manually extracting key information. Due to the lack of effective automatic summarization and priority screening mechanisms, redundant information can easily interfere with the analysis process, making it difficult to quickly obtain core data, thereby impacting the efficiency of emergency response and the timeliness of decision-making. In terms of event correlation analysis, existing systems primarily rely on static rules or simple statistical associations, lacking in-depth modeling and reasoning capabilities for the relationships between complex events. For example, traffic congestion in a certain area may affect the dispatch of rescue vehicles, and negative public opinion on social media may reflect delays in rescue efforts. However, these potential correlations are often difficult for traditional analysis tools to automatically discover and utilize, resulting in a lack of a global perspective in the formulation of emergency response strategies, potentially missing key influencing factors. Furthermore, online analysis is highly dynamic and interactive. However, in a dynamically changing environment, existing systems often rely on pre-set, fixed analysis models and employ fixed query patterns. These systems often lack adaptive learning capabilities and the ability to predict event trends. They are unable to automatically optimize analysis processes based on changing needs, making them difficult to adapt to complex and changing analytical requirements. This limits their applicability in long-term emergency management and risk assessment. For example, command personnel may need to query information such as "the nearest available rescue vehicle to a rescue point," "the rescuer with the longest working hours," or "the latest distress message and its key elements" in real time. Traditional methods typically require pre-defined query logic, making it difficult to flexibly respond to temporary user needs. Furthermore, as an incident evolves, the focus may shift from the scope of the disaster to resource allocation and then to post-disaster recovery. Existing systems often require manual adjustments to analysis logic and are unable to adaptively adjust data processing strategies. This limitation makes traditional analytical tools difficult to adapt to rapidly changing business scenarios in practice, resulting in reduced analytical efficiency and difficulty meeting the adaptability requirements of long-term operations, reducing the system's adaptability and user experience.

[0085] When the present disclosure parses analysis requests in complex scenarios with reference to analysis records of historical analysis requests, as an example, the task decomposition method of the queried historical analysis requests can be directly used; the content of the multiple query tasks obtained by decomposition can also be adjusted based on the changes in time, content, background information, etc. of the current analysis request relative to the queried historical analysis requests (regarding time, for example, whether the event has developed to different stages can be estimated based on the length of the interval time, thereby adjusting the focus of the query task; regarding content, for example, the relevant content in the query task can be adjusted based on the difference in content; regarding background information, for example, for analysis requests for formulating rescue plans, the focus of the query task can be adjusted by referring to changes in background information such as the weather and infrastructure status of the day, and the background information can be stored in the scene description cache in the data storage module introduced above); the reasoning ability of the large language model can also be used, and prompt words can be used to enable the large language model to automatically decompose tasks with reference to the analysis records of the queried historical analysis requests.

[0086] Regarding the aforementioned changes in background information, as an example, for the embodiment described above where the preset condition includes specifying that the similarity of the background information is greater than a third similarity threshold, if a historical analysis request is found that does not meet this preset condition but meets the preset condition that the similarity is greater than a second similarity threshold, then although the analysis result data of this historical analysis request cannot be directly reused, this historical analysis request can be directly used as a reference for resolving the query task (regardless of whether it meets the time condition in the preset condition), because this historical analysis request necessarily meets the condition that the similarity is greater than the first similarity threshold, thereby reducing the computing resource consumption of querying the historical analysis request. At the same time, the query task obtained by the analysis can be adjusted based on the changes in the specified background information to improve the quality of task resolution.

[0087] Regarding task parsing, in some embodiments, optionally, the operation of parsing the analysis request into a query task in step S103 includes: using a large language model to parse the analysis request into a query task. When processing complex requests, the system can combine the reasoning capabilities of the large language model to automatically break down the analysis request into more fine-grained query tasks using predefined prompt word templates. This task breakdown method can improve query accuracy and optimize analysis strategies using historical data and similar cases, making the broken-down tasks more consistent with business logic, thereby improving the flexibility and reliability of task breakdown.

[0088] Taking "get the overall situation of the current rescue scene" as an example, the system can use prompt words similar to the following:

[0089]

[0090] Based on these prompt words, the large language model can break down high-level analysis requests into multiple specific query tasks, and the analysis module then sends the broken down query tasks to the adaptive query module. Still taking "obtaining the overall situation of the current rescue scenario" as an example, it can be broken down into the following specific query tasks: 1. Obtain the total number of available rescue supplies: including statistics on the quantity of food, water, medicine, tents, etc. 2. Obtain the total number of rescue personnel: including on-site rescue personnel, logistics support personnel, and their location, status, and professional skills distribution. 3. Obtain weather and environmental conditions: including current weather, future forecasts, and the possibility of potential secondary disasters. 4. Obtain the scale of the affected population: including the number of people affected, missing people, casualties, etc. 5. Obtain key rescue targets: information on key facilities that require priority protection, such as hospitals, schools, and community centers.

[0091] It should be understood that the large language models used here include both pure large language models and multimodal large models. As technology advances, the performance of multimodal large models and machines will gradually improve. Queries can also leverage the capabilities of multimodal large models, using conditions or results from multiple modalities as query elements. From a low-level implementation perspective, query tool parameters can also be represented using multimodal data.

[0092] Regarding task parsing, in some other embodiments, optionally, the operation of parsing the analysis request into a query task in step S103 includes parsing the analysis request into a query task using a predefined template. For fixed query requirements (such as inventory queries and personnel scheduling), query efficiency can be improved by using predefined query templates, such as SQL queries or API endpoints, rather than dynamically parsing queries using a large language model.

[0093] In further embodiments, optionally, the operation of parsing the analysis request into a query task in step S103 includes parsing the analysis request into a query task based on a knowledge graph. By employing knowledge graph-based query optimization, in scenarios with complex relational data (e.g., disaster impact propagation paths), "knowledge graph + SQL" queries can be employed. Alternatively, a knowledge graph that supports large language models (e.g., GraphRAG) can be used to replace the free text parsing of large language models, thereby enhancing causal reasoning capabilities.

[0094] Generally speaking, step S103 can achieve the following aspects.

[0095] Cache content management: The embedded vectors of analysis requests are stored in a vector database, and the recent query tasks are stored in a memory queue to improve query efficiency.

[0096] Intelligent reuse and incremental updates: When parsing a new task, the cache is first queried to determine whether existing results can be directly reused or incremental optimization can be performed based on existing data, thereby reducing computational overhead.

[0097] Change detection: Combining vector database similarity calculation and time window mechanism, it identifies key change points, such as fluctuations in rescue resources and changes in high-risk areas, and annotates key information to highlight dynamic evolution trends.

[0098] Task decomposition optimization: Based on historical task completion status, prompt words are automatically iterated and optimized to improve the accuracy of task decomposition and the adaptability of query tasks.

[0099] Multi-round interaction support: When users conduct continuous queries, the conversation context information is stored to ensure the consistency of the query, eliminating the need for users to repeatedly describe the query background, thus improving the intelligent interaction experience.

[0100] The above mechanism provides the system with an adaptive scenario analysis method.

[0101] Next, step S104 is further introduced, that is, how to call the query tool to perform data query through the adaptive query module.

[0102] In some embodiments, the operation of invoking a query tool from the query tool library based on the query task in step S104 optionally includes: determining query tool requirements based on the query task, wherein the query tool requirements represent the functionality of the query tool to be invoked; retrieving a query tool from the query tool library that meets the query tool requirements; and dynamically generating a query tool that meets the query tool requirements if the retrieval fails. Clarifying the query tool requirements provides a basis for determining the query tool. By checking the query tool library to see if a suitable query tool exists to directly execute the query, and dynamically generating a new query tool if no matching query tool is registered in the query tool library, this allows for fully utilizing existing query tools while enabling real-time querying and adaptive expansion capabilities in complex scenarios. For example, after dynamically generating a new query tool, its logical correctness and computational efficiency may be further tested to ensure its usability. For example, the adaptive query module may receive a natural language query request, retrieve or dynamically generate a query tool using a large language model (e.g., GPT4o, Qwen2.5, etc.), and then invoke the appropriate tool to extract the required data from the real-time situation data cache and historical data storage in the data storage module.

[0103] It should be understood that dynamically generated new query tools can also be stored in the query tool library, expanding the library. Initially, the query tool library can store a number of commonly used predefined query tools. This means the query tool library includes both predefined and dynamically generated query tools. Since dynamically generated query tools are generated for specific analysis request query tasks, their potential for reuse varies. Storing difficult-to-reuse tools in the query tool library for an extended period of time would waste storage space and increase computing resource consumption and search time for the tools.

[0104] To address this issue, the query tool library optionally also records call information for dynamically generated query tools, including call counts and call times. The adaptive complex scene information analysis method according to an exemplary embodiment of the present disclosure further includes: determining the call frequency of dynamically generated query tools in the query tool library based on the call information; and deleting dynamically generated query tools whose call frequencies meet a preset deletion condition, where the preset deletion condition is used to indicate low-frequency calls. By counting the call frequencies of dynamically generated query tools and deleting low-frequency call tools, that is, removing infrequently used query tools, storage space can be freed up, thereby improving the availability and call efficiency of tools in the library. It should be understood that this operation can be performed by the query tool library or by the adaptive query module. In actual implementation, as an example, a dynamically generated query tool can be assigned a dynamic generation flag, or dynamically generated query tools and predefined query tools can be stored separately to facilitate determining the call frequency and deleting tools only for dynamically generated query tools. A default call frequency can also be assigned to predefined query tools, and the default value can be set to a sufficiently large value. This disclosure is not limited to this. The preset deletion conditions can be set as needed, for example, including when the call frequency is below a frequency threshold, or when the call frequency ranks a certain number of places low among all dynamically generated query tools (the specific number of places can be determined based on the number of tools to be deleted), and this disclosure does not impose any restrictions on this. The timing of executing this operation can also be set as needed, including but not limited to periodic cleanup and cleanup when a new tool is generated, and this disclosure does not impose any restrictions on this.

[0105] In addition to using large language models to dynamically generate query tools, in other embodiments, if query requirements are relatively fixed, a fixed tool set can be used, with a predefined set of query tools that are manually updated regularly, without the need for dynamic tool generation. This approach is suitable for systems with high stability requirements.

[0106] In some further embodiments, plug-in-based tool management can be adopted, whereby tools can be deployed as independent plug-ins (e.g., Python APIs) and support dynamic loading, rather than relying entirely on large language model generation. This approach increases flexibility but requires additional tool management mechanisms.

[0107] In addition, in some embodiments, the information analysis system optionally further includes a domain knowledge base. Step S104 includes: invoking a query tool in a query tool library according to a query task by the adaptive query module to query the data storage module for data as preliminary query data; analyzing whether the preliminary query data meets the analysis requirements of the analysis request; if the analysis requirements of the analysis request are met, using the preliminary query data as query result data; if the analysis requirements of the analysis request are not met, analyzing and reasoning the preliminary query data by the analysis module in combination with the data in the data storage module and the domain knowledge base to obtain an updated query task; re-calling the query tool in the query tool library according to the updated query task by the adaptive query module to query the data storage module for data to update the preliminary query data, and repeating the step of analyzing whether the preliminary query data meets the analysis requirements of the analysis request until query result data is obtained or the maximum number of updates is reached. By not directly applying the queried data but first analyzing it to determine whether it meets the analysis requirements, the queried data can be verified in advance, and if it does not meet the analysis requirements, a new query task can be re-inferred by combining multiple information. This cycle can effectively improve the quality of the analysis results.

[0108] Domain knowledge bases store industry-specific knowledge to enhance the reasoning capabilities of analysis modules. For example, they include standard rescue operation procedures, equipment dispatching rules, and public opinion analysis methods. Domain knowledge bases are typically implemented using RAG (Retrieval-Augmented Generation) or knowledge graph technology (e.g., GraphRAG).

[0109] After completing the initial query, the system first checks whether the query result data meets the analysis requirements of the analysis request. If the results meet the analysis requirements, the system directly proceeds to the next step; otherwise, the system determines whether further reasoning is required to explore deeper information or optimize the analysis logic.

[0110] When further reasoning is needed, the system leverages the scene data cache, scene description cache, and domain knowledge base in the data storage module to conduct comprehensive analysis and deep reasoning on the current preliminary query data to identify potential correlates. For example, when analyzing whether traffic congestion in a certain area will affect rescue operations, the system may invoke a large language model to perform multi-step reasoning, automatically generating new queries or adjusting existing analysis logic to ensure that the reasoning process is more realistic. During the reasoning process, the system dynamically assesses whether the current information is sufficient to support the final conclusion. If critical data is insufficient, the system automatically backtracks to the adaptive query process, triggering a new data query task to supplement the missing information. This backtracking mechanism ensures the continuity of query and reasoning while avoiding analytical bias caused by insufficient information. The system also controls the rational use of computing resources based on a preset maximum query / inference depth, ensuring analytical integrity and avoiding meaningless reasoning loops. This reflective reasoning mechanism not only improves the system's intelligent adaptability to complex analytical tasks but also enhances the accuracy and explainability of its decisions.

[0111] In addition to using large language model reasoning for complex event analysis, other approaches can be employed as needed. For example, rule-based decision support can be used. Specifically, in some areas (such as disaster warning and logistics scheduling), expert rule engines (such as Drools) can be used instead of large language model reasoning, which can improve interpretability to a certain extent. In event correlation analysis, causal reasoning models (such as Bayesian networks and Granger causality analysis) can be used instead of large language model reasoning, improving the transparency and interpretability of causal reasoning.

[0112] Figure 3 is a flowchart of adaptive query processing according to a specific embodiment of the present disclosure.

[0113] In this specific embodiment, the query tool library stores all predefined and dynamically generated query tools and records the description and invocation method of each tool to support the matching process. The main contents of the query tool library include the tool description (Description) and the tool invocation method (Invocation). The former briefly describes the tool's function and applicable scenarios, and the latter provides a specific tool invocation format. The query tool library manages query tools through metadata. Metadata includes, for example, the tool type (predefined or dynamically generated), invocation count, and last invocation time. The system can use metadata to dynamically optimize the tool functions in the query tool library, such as removing infrequently used tools (primarily dynamically generated query tools), thereby improving tool availability and tool invocation efficiency.

[0114] After receiving the query task, the adaptive query module decides which query tool to use for the query. The system supports dynamic registration / deregistration of the query tool library, and can adjust the use of query tools according to task requirements, making the query method more flexible. For example, if a query is received for "querying idle vehicles in disaster area A" or "checking recent inventory consumption trends", the following will be triggered: Figure 3 The adaptive query processing flow shown in the figure below first converts a natural language query request into a query tool call request and searches the query tool library for available tools. If no matching tool exists, the large language model is used to dynamically generate a new query tool, which is then tested and verified for logical correctness and performance. After a tool is matched or generated, the system calls the corresponding tool with the necessary input parameters (such as "vehicle model = truck") and retrieves the query result data from the data storage module in real time. Simultaneously, the system updates metadata such as the tool's call count and timestamp, and uses a least-recently-used (LRU) algorithm to prune infrequently used tools to ensure efficient operation of the tool library.

[0115] Next, the details of the adaptive query processing flow of this specific embodiment are introduced.

[0116] Reference Figure 3 In step S301, a natural language query request is received.

[0117] Users can initiate analysis requests through natural language descriptions. For simple requests, they can be directly used as natural language query requests. For complex requests, they need to go through the following steps: Figure 1 Step S103 is parsed into multiple query tasks. These query tasks can also be in the form of natural language query requests. The natural language query request interface converts the request into a call requirement of the query tool. For example, for the query request "Query for idle vehicles in disaster area A", the function call generated by the large language model is as follows:

[0118]

[0119] In step S302, a tool matching prompt word is constructed.

[0120] Extract the manuals and call methods of all tools from the query tool library, splice them into Chinese prompts, and append the query request. For example, if you enter "Query for available vehicles in disaster area A", the spliced ​​prompts may be as follows:

[0121]

[0122] The large model selects the best query tool based on the spliced ​​Chinese prompt words, or returns the "<:NO_MATCH>" mark.

[0123] In step S303, matching tools.

[0124] The large language model returns a matching result based on the prompt word. If the tool is successfully matched, the tool is called and the process proceeds to step S307. If it fails, the process proceeds to step S304 to start the dynamic tool generation process.

[0125] Query tools include predefined query tools by type ( ) and the Dynamic Query Generation Tool ( ) two. Predefined query tools ( ) includes the system's built-in common query tool functions, handles common needs, and resides for a long time; dynamically generates query tools ( ) includes tool functions dynamically generated by the large language model according to specific query requirements, and supports an automatic cleanup mechanism (see step S309 for details).

[0126] In step S304, a tool is dynamically generated.

[0127] Calling large language models to dynamically generate new query tools , the tool logic is combined with the input parameters of the query request. The generated new query tool function is, for example:

[0128]

[0129] In step S305, the newly generated tool is tested and verified.

[0130] For newly generated query tools, the large language model generates validation procedures that cover a variety of input scenarios and boundary conditions. For example:

[0131]

[0132] The tests are executed in a separate process, and mechanisms such as execution timeout checks are set up to ensure the robustness of the system.

[0133] In step S306, failure feedback and reflection and improvement are performed.

[0134] When a newly generated tool fails the test, the system will record the failure reason and test feedback information. The failure reasons usually include syntax errors, logical errors, performance issues, and errors out of range. Among them, logic errors refer to the generated tool not meeting the query requirements, syntax errors refer to the generated tool function having syntax errors and cannot be compiled and run normally, performance issues refer to the tool execution time or resource consumption exceeding expectations, and error issues refer to the tool output results With expected results The error exceeds the allowable range , the available formula express.

[0135] Based on error feedback, the system modifies the tool to generate prompt words for improvement. For example, the system combines the failure cause and feedback information to construct a new prompt word. The prompt word can be supplemented with the following content:

[0136]

[0137] Based on the improved prompt words, the failure reasons and test feedback can be passed to the large language model for reflection and improvement, starting the iterative process of "recreate -> test", that is, the large language model regenerates an improved version of the tool. If the newly generated tool still fails the test program verification, this process is repeated until the newly generated tool passes the test or the preset maximum number of retries is reached.

[0138] In step S307, the tool function is executed.

[0139] This step calls the matching tool or the newly generated tool, passes in the required parameters, and extracts the results from the data storage module. The process can be expressed as follows:

[0140]

[0141] In the above formula, Indicates the query results. Represents the query tool logic function, Indicates additional input parameters provided by the user, such as "Vehicle Model = Truck", Represents the situation data cache in the data storage module (such as Redis data).

[0142] During execution, the query tool will combine and Extract the results that meet the criteria.

[0143] In step S308, the tool metadata is updated.

[0144] After the query tool is executed, the tool metadata is updated, such as the call count and the last call time.

[0145] In step S309, the tool is cleaned.

[0146] Over time, the query tool library may accumulate too many generated tools, potentially degrading the performance of large language models. For example, excessively long prompt words can affect the model's parsing ability; at the same time, the large language model's comprehension accuracy can also decrease when processing long texts. To prevent the excessive number of generated tools from affecting tool invocation accuracy, the system uses a preset tool limit combined with a least-recently-used (LRU) algorithm. Each time a new tool is generated, the system automatically checks and removes infrequently used dynamically generated query tools, ensuring the efficient operation of the query tool library.

[0147] The conditional formula for tool cleaning is as follows:

[0148]

[0149] In this formula, Indicates the total number of tools in the current query tool library. Indicates the upper limit of the number of tools. Exceed When , the cleanup mechanism is triggered to remove the least recently used tool. It should be understood that when the total number of tools reaches the upper limit of the tool book, this condition will be triggered every time a new tool is generated.

[0150] The LRU cleaning priority formula is defined as:

[0151]

[0152] In this formula, Represents the collection of tools to be cleaned, Represents a collection of dynamically generated tools. Display Tools The calling frequency is defined as:

[0153]

[0154] In this formula, Display Tools The number of calls, Display Tools The last call time, Display Tools Registration time, Indicates in the tool The number of times a function is called per unit time during its lifetime.

[0155] Collection of tools to be cleaned Frequency of tool calls The decision was made, and the registration time of the tools was taken into consideration. The tools with the lowest call frequency were removed first.

[0156] The cleanup operation can be expressed as:

[0157]

[0158] Through the above formula, the system limits the number of dynamically generated tools while ensuring the efficient operation of the query tool library, effectively reducing the risk of large language model parsing tool errors caused by overly long tool call prompt words.

[0159] In general, considering that in complex scenario analysis, data sources are diverse and analysis requirements are highly dynamic, traditional fixed query patterns are difficult to meet the needs of real-time adjustment. According to the adaptive complex scenario information analysis method of the exemplary embodiment of the present disclosure, a combination of adaptive query and intelligent analysis is adopted to dynamically parse the analysis requests of users or downstream business modules, and combine historical data, domain knowledge and intelligent reasoning to optimize the query strategy.

[0160] Specifically, the present disclosure employs adaptive analysis of complex scenario data. By integrating a large multimodal model, it achieves comprehensive analysis and understanding of structured and unstructured text and visual data. Compared with traditional methods, this adaptive analysis mechanism significantly improves data processing efficiency and real-time performance. The system can automatically adjust the analysis process based on real-time data to ensure the accuracy and timeliness of decisions. For example, when processing multi-source heterogeneous data, the system can automatically identify data changes and dynamically adjust analysis strategies, avoiding decision delays caused by data fragmentation in traditional methods.

[0161] This disclosure also utilizes a dynamic query tool management mechanism. The system uses a large language model to dynamically generate query tools, supporting real-time generation or adjustment of tool sets based on varying analytical requirements. Compared to traditional fixed tool sets, this dynamic generation mechanism reduces manual intervention, improves response speed, and enhances system scalability. The system automatically generates new query tools based on user needs and automatically clears infrequently used tools after tasks are completed, ensuring efficient operation of the query tool library. This mechanism enables the system to quickly adapt to changing analytical needs and improves overall flexibility.

[0162] This disclosure also utilizes adaptive scenario summarization. By combining data caching with a vector database, the system achieves efficient caching, retrieval, and dynamic management. It also generates analytical summaries based on a large language model, offering significant advantages in relevance and accuracy. The system adaptively optimizes summary generation based on historical analysis requests and contextual information, enabling users to quickly access key information and enhancing the intelligence and coherence of analysis.

[0163] The present disclosure also adopts flexible configuration of alternative solutions. The system provides a variety of alternative solutions, such as retaining only the downstream business module or the user interaction module, and choosing to use a pure large language model or a multimodal large model based on the characteristics of the data. This flexibility enables the system to better adapt to different application scenarios. For example, in scenarios with small data volumes, the use of a pure large language model can reduce computing costs, while in scenarios with rich visual data, the application of a multimodal large model can significantly improve the accuracy of the analysis.

[0164] This disclosure also innovates the overall system architecture. The system adopts a modular design, which is highly scalable. This design enables the system to easily integrate new data sources and query tools, adapting to future technological developments and changing business needs. For example, the system's scalability makes it easy to integrate new data sources or query analysis algorithms, improving the system's data processing capabilities and adaptability.

[0165] In summary, the exemplary embodiment of the present disclosure adopts a large language model combined with adaptive tool calling, dynamic task decomposition, dynamic tool registration and multimodal data fusion technology to build an efficient intelligent analysis system. This system can not only automatically extract key information, optimize the query process, and realize the reasoning and correlation analysis of complex events, but also solve the shortcomings of existing technologies in data integration, real-time analysis, interactive query and complex event reasoning, further improve the overall performance and scope of application of the system, but also has adaptive learning capabilities, and can dynamically adjust the analysis strategy according to changes in scenarios and actual needs, and provide more intelligent decision support, thereby improving information utilization efficiency and decision support capabilities in complex environments, and meeting the needs of highly dynamic and high-uncertainty application scenarios such as emergency management and smart cities. It should be understood that this article introduces different implementation methods for each step. In actual applications, different solutions can be selected according to specific needs and resource conditions.

[0166] Figure 4 This is a block diagram of an adaptive complex scene information analysis device according to an exemplary embodiment of the present disclosure. The device is used in an information analysis system, which includes a data source module, a data conversion module, a data storage module, an analysis module, a query tool library, an adaptive query module, and a history storage module for storing analysis records of historical analysis requests. Figure 4 The adaptive complex scene information analysis device 400 includes an acquisition unit 401, a conversion unit 402, a parsing unit 403, a query unit 404, and an analysis unit 405.

[0167] The acquisition unit 401 is configured to respectively acquire scene data from multiple data sources through a data source module.

[0168] The conversion unit 402 is configured to convert the acquired scene data into a preset data format through the data conversion module, and store the data in the data storage module.

[0169] The parsing unit 403 is configured to respond to receiving an analysis request through the analysis module, query historical analysis requests whose similarity with the analysis request is greater than a first similarity threshold from the history storage module, and refer to the analysis records of the queried historical analysis requests to parse the analysis request into a query task.

[0170] The query unit 404 is configured to call the query tool in the query tool library according to the query task through the adaptive query module, perform data query on the data storage module, and obtain query result data.

[0171] The analysis unit 405 is configured to analyze and process the query result data through the analysis module to obtain analysis result data.

[0172] Optionally, the conversion unit 402 is further configured to generate an object identifier of the object as a key and the scene data of the object as a value for each object scene data in a key-value pair format in the data storage module through the data conversion module.

[0173] Optionally, the parsing unit 403 is further configured to: evaluate whether the analysis request is a complex request in response to receiving the analysis request through the analysis module, wherein a complex request is an analysis request that needs to be parsed into at least N query tasks, and N is an integer greater than 1; if the analysis request is not a complex request, parse the analysis request into a query task; if the analysis request is a complex request, query historical analysis requests whose similarity with the analysis request is greater than a first similarity threshold from the historical storage module, and refer to the analysis records of the queried historical analysis requests to parse the analysis request into a query task.

[0174] Optionally, the parsing unit 403 is further configured to: in the case that the analysis request is a complex request, determine whether there is a historical analysis request that meets the preset conditions in the historical storage module, wherein the preset conditions include: the difference between the request time of the corresponding historical analysis request and the request time of the analysis request is less than the time difference threshold, and the similarity between the corresponding historical analysis request and the analysis request is greater than a second similarity threshold, wherein the second similarity threshold is greater than the first similarity threshold; if there is no historical analysis request that meets the preset conditions, query the historical analysis request whose similarity with the analysis request is greater than the first similarity threshold from the historical storage module, and refer to the analysis records of the queried historical analysis requests to parse the analysis request into a query task; if there is a historical analysis request that meets the preset conditions, abandon the parsing query task, and obtain the analysis result data of the historical analysis request that meets the preset conditions as the analysis result data of the analysis request.

[0175] Optionally, the parsing unit 403 is further configured to: parse the analysis request into a query task using a large language model; and / or parse the analysis request into a query task using a predefined template; and / or parse the analysis request into a query task based on a knowledge graph.

[0176] Optionally, the query unit 404 is further configured to: determine the query tool requirements based on the query task through the adaptive query module, wherein the query tool requirements represent the functions of the query tool to be called; retrieve the query tool that meets the query tool requirements from the query tool library; and dynamically generate a query tool that meets the query tool requirements in the event of retrieval failure.

[0177] Optionally, the query tool library also records the call information of the dynamically generated query tools, and the call information includes the number of calls and the call time. The adaptive complex scene information analysis device 400 also includes a tool management unit (not shown in the figure), which is configured to: for the query tools dynamically generated in the query tools, determine the call frequency of the corresponding query tools according to the call information; delete the dynamically generated query tools whose call frequency meets the preset deletion conditions, wherein the preset deletion conditions are used to represent low-frequency calls.

[0178] Optionally, the information analysis system also includes a domain knowledge base, and the query unit 404 is further configured to: call the query tool in the query tool library according to the query task through the adaptive query module, and perform data query on the data storage module as preliminary query data; analyze whether the preliminary query data meets the analysis requirements of the analysis request; if the analysis requirements of the analysis request are met, the preliminary query data is used as the query result data; if the analysis requirements of the analysis request are not met, the preliminary query data is analyzed and inferred by the analysis module in combination with the data in the data storage module and the domain knowledge base to obtain an updated query task, and re-call the query tool in the query tool library through the adaptive query module according to the updated query task to perform data query on the data storage module to update the preliminary query data, and repeat the step of analyzing whether the preliminary query data meets the analysis requirements of the analysis request until the query result data is obtained or the maximum number of updates is reached.

[0179] Regarding the apparatus in the above embodiment, the specific manner in which each unit performs operations has been described in detail in the embodiment of the method, and will not be elaborated on here.

[0180] Figure 5 FIG. 5 shows a structural block diagram of an electronic device 500 according to an exemplary embodiment of the present disclosure.

[0181] Reference Figure 5 The electronic device 500 includes: at least one memory 501 and at least one processor 502, wherein the at least one memory 501 stores computer executable instructions. When the computer executable instructions are executed by the at least one processor 502, the at least one processor is prompted to execute the adaptive complex scene information analysis method as described in the above exemplary embodiment.

[0182] As an example, electronic device 500 may be a PC, tablet device, personal digital assistant, smartphone, or other device capable of executing the aforementioned instruction set. Here, electronic device 500 is not necessarily a single electronic device 500, but may also be any collection of devices or circuits capable of executing the aforementioned instructions (or instruction set) individually or in combination. Electronic device 500 may also be part of an integrated control system or system manager, or may be configured as a portable electronic device 500 that is interconnected locally or remotely (e.g., via wireless transmission) via an interface.

[0183] In electronic device 500, processor 502 may include a central processing unit (CPU), a graphics processing unit (GPU), a programmable logic device, a dedicated processor system, a microcontroller, or a microprocessor. By way of example and not limitation, processor 502 may also include an analog processor, a digital processor, a microprocessor, a multi-core processor, a processor array, a network processor, etc.

[0184] The processor 502 can execute instructions or codes stored in the memory 501, wherein the memory 501 can also store data. Instructions and data can also be sent and received over a network via a network interface device, wherein the network interface device can use any known transmission protocol.

[0185] The memory 501 may be integrated with the processor 502, for example, by placing RAM or flash memory within an integrated circuit microprocessor or the like. Furthermore, the memory 501 may comprise a separate device, such as an external disk drive, a storage array, or any other storage device usable by a database system. The memory 501 and the processor 502 may be operatively coupled or may communicate with each other, for example, via an I / O port, a network connection, or the like, such that the processor 502 can access files stored in the memory.

[0186] In addition, the electronic device 500 may further include a video display (such as a liquid crystal display) and a user interaction interface (such as a keyboard, a mouse, a touch input device, etc.) All components of the electronic device 500 may be connected to each other via a bus and / or a network.

[0187] According to an exemplary embodiment of the present disclosure, a computer-readable storage medium storing instructions may also be provided, wherein the instructions, when executed by at least one processor, prompt the at least one processor to execute the adaptive complex scene information analysis method as described in the above exemplary embodiment. Examples of computer-readable storage media include: read-only memory (ROM), random access programmable read-only memory (PROM), electrically erasable programmable read-only memory (EEPROM), random access memory (RAM), dynamic random access memory (DRAM), static random access memory (SRAM), flash memory, non-volatile memory, CD-ROM, CD-R, CD+R, CD-RW, CD+RW, DVD-ROM, DVD-R, DVD+R, DVD-RW, DVD+RW, DVD-RAM, BD-ROM, BD-R, BD-R LTH, BD-RE, Blu-ray or optical disk storage, hard disk drive (HDD), solid state drive (SSD), card storage (such as a multimedia card, secure digital (SD) card or extreme digital (XD) card), magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid state disk and any other device configured to store a computer program and any associated data, data files and data structures in a non-transitory manner and provide the computer program and any associated data, data files and data structures to a processor or computer so that the processor or computer can execute the computer program. The computer program in the above-mentioned computer-readable storage medium can be executed in an environment deployed in a computer device such as a client, a host, an agent device, a server, etc. In addition, in one example, the computer program and any associated data, data files and data structures are distributed on a networked computer system so that the computer program and any associated data, data files and data structures are stored, accessed and executed in a distributed manner by one or more processors or computers.

[0188] According to an exemplary embodiment of the present disclosure, a computer program product may also be provided, including computer instructions, which, when executed by at least one processor, execute the adaptive complex scene information analysis method as described in the above exemplary embodiment.

[0189] Other embodiments of the present disclosure will readily occur to those skilled in the art after considering the specification and practicing the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only, with the true scope and spirit of the present disclosure being indicated by the appended claims.

[0190] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes can be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. An adaptive complex scene information analysis method, characterized in that: The adaptive complex scene information analysis method is used in an information analysis system, which includes a data source module, a data conversion module, a data storage module, an analysis module, a query tool library, an adaptive query module, and a history storage module for storing analysis records of historical analysis requests. The adaptive complex scene information analysis method includes: Acquire scene data from multiple data sources respectively through the data source module; The acquired scene data is converted into a preset data format by the data conversion module, and stored in the data storage module; In response to receiving the analysis request, the analysis module evaluates whether the analysis request is a complex request, wherein the complex request is an analysis request that needs to be parsed into at least N query tasks, where N is an integer greater than 1; In the case where the analysis request is the complex request, determining whether there is a historical analysis request that meets a preset condition in the history storage module, wherein the preset condition includes: a difference between a request time of a corresponding historical analysis request and a request time of the analysis request is less than a time difference threshold, and a similarity between the corresponding historical analysis request and the analysis request is greater than a second similarity threshold, wherein the second similarity threshold is greater than the first similarity threshold; If there is a historical analysis request that meets the preset conditions, abandon the parsing query task and obtain the analysis result data of the historical analysis request that meets the preset conditions as the analysis result data of the analysis request; If there is no historical analysis request that meets the preset condition, querying the historical analysis request whose similarity with the analysis request is greater than the first similarity threshold from the historical storage module, and resolving the analysis request into a query task with reference to the analysis records of the retrieved historical analysis request; If the analysis request is not the complex request, parsing the analysis request into a query task; The adaptive query module calls the query tool in the query tool library according to the query task analyzed by the analysis module, performs data query on the data storage module, and obtains query result data; The query result data is analyzed and processed by the analysis module to obtain analysis result data.

2. The adaptive complex scene information analysis method according to claim 1, characterized in that: The step of converting the acquired scene data into a preset data format by the data conversion module and storing the data in the data storage module includes: For the scene data of each object, the data conversion module generates an object identifier of the object as a key, and uses the scene data of the object as a value, and stores the object in the data storage module in a key-value pair format.

3. The adaptive complex scene information analysis method according to claim 1, characterized in that: The step of parsing the analysis request into a query task includes: Using a large language model, parsing the analysis request into the query task; and / or Parsing the analysis request into the query task using a predefined template; and / or Based on the knowledge graph, the analysis request is parsed into the query task.

4. The adaptive complex scene information analysis method according to claim 1, characterized in that: The step of calling the query tool in the query tool library according to the query task by the adaptive query module includes: Determining query tool requirements based on the query task by the adaptive query module, wherein the query tool requirements represent functions of the query tool to be called; Retrieving a query tool that meets the query tool requirement from the query tool library; In case of retrieval failure, a query tool that meets the query tool requirements is dynamically generated.

5. The adaptive complex scene information analysis method according to claim 4, characterized in that: The query tool library also records the call information of the dynamically generated query tool, wherein the call information includes the number of calls and the call time. The adaptive complex scene information analysis method further includes: For a dynamically generated query tool among the query tools, determining a call frequency of the corresponding query tool according to the call information; A dynamically generated query tool is provided for deleting calls whose frequency satisfies a preset deletion condition, wherein the preset deletion condition is used to represent low-frequency calls.

6. The adaptive complex scene information analysis method according to any one of claims 1 to 5, characterized in that: The information analysis system further includes a domain knowledge base, wherein the adaptive query module calls a query tool in the query tool library according to the query task to perform a data query on the data storage module to obtain query result data, including: The adaptive query module calls the query tool in the query tool library according to the query task to query the data in the data storage module as preliminary query data; Analyzing whether the preliminary query data meets the analysis requirements of the analysis request; If the analysis requirements of the analysis request are met, the preliminary query data is used as the query result data; If the analysis requirements of the analysis request are not met, the preliminary query data is analyzed and inferred by the analysis module in combination with the data in the data storage module and the domain knowledge base to obtain an updated query task. The adaptive query module re-calls the query tool in the query tool library according to the updated query task, performs data query on the data storage module to update the preliminary query data, and repeats the step of analyzing whether the preliminary query data meets the analysis requirements of the analysis request until the query result data is obtained or the maximum number of updates is reached.

7. An electronic device, characterized in that: include: at least one processor; at least one memory storing computer-executable instructions, Wherein, when the computer executable instructions are executed by the at least one processor, the at least one processor is prompted to execute the adaptive complex scene information analysis method according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that When the instructions in the computer-readable storage medium are executed by at least one processor, the instructions cause the at least one processor to execute the adaptive complex scene information analysis method according to any one of claims 1 to 6.

9. A computer program product comprising computer instructions, characterized in that When the computer instructions are executed by at least one processor, the at least one processor is prompted to execute the adaptive complex scene information analysis method according to any one of claims 1 to 6.

Images