Defense equipment intelligence automatic report generation method based on large language model

Abstract

The application provides a defense equipment intelligence automatic report generation method based on a large language model, relates to the technical field of intelligence automatic report generation, and comprises the following steps: multi-source heterogeneous intelligence data preprocessing and element extraction, intelligence element association and report content automatic generation, content filling verification and model closed-loop iteration, and report full life cycle traceability optimization. The application can optimize the large language model based on the defense field knowledge graph, access multi-source heterogeneous intelligence data, and complete preprocessing through multi-modal extraction; the intelligence elements are associated with the knowledge graph to generate structured conclusions, which drive the model to generate compliant intelligence reports; through multi-granularity filling and consistency verification, a closed-loop iteration is formed, the report full life cycle traceability optimization is triggered as needed, the model field generation capability is continuously improved, the reliability of the defense equipment intelligence automatic report generation is improved, and the problem of low reliability of the defense equipment intelligence automatic report generation in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of automated intelligence report generation technology, and in particular to a method for generating automated intelligence reports for defense equipment based on a large language model. Background Technology

[0002] In the real-time perception and integration of dynamic intelligence, reports are prone to becoming outdated or missing key dynamics. This is because the underlying system architecture was designed from the outset to rely on static knowledge bases or batch data processing models, and it lacks a technical link for low-latency connection with real-time intelligence streams (such as automatic recognition results of satellite imagery, open-source intelligence crawling data, and sensor data streams). At the same time, it lacks the ability to automatically interpret multi-source real-time data. It also lacks real-time data access protocols optimized for defense scenarios and domain reasoning modules that can quickly link equipment dynamics, deployment changes, and exercise intelligence. As a result, when new intelligence data is generated, it is impossible to complete data cleaning and content integration in a short time. The resulting situation reports, such as weekly situation reports, can only rely on historical static information and cannot reflect the latest battlefield dynamics.

[0003] In the understanding and reasoning of multi-source heterogeneous data, there is a lack of a unified framework for multimodal information extraction and association. On the one hand, the data formats involved in defense intelligence, such as text reports, satellite images, radar signals, social media information, and equipment technical parameter tables, vary greatly. Existing technologies have not established a cross-modal knowledge association mechanism, making it impossible for models to understand the technical details in images of new equipment or the abnormal features in the radar spectrum. On the other hand, the model training phase lacks sufficient multimodal domain-labeled data, making it impossible to learn the complete logical link from visual and signal features to equipment performance inference and combat intent judgment. Therefore, it can only remain at the level of single-modal content generation and cannot build in-depth intelligence analysis supported by multi-source data.

[0004] In the generation and management phases, due to a misalignment in the functional design, the current model management function only focuses on the lifecycle operations of the model itself, such as creation, editing, and deletion. It does not incorporate report generation tasks and report content into a unified management system. It does not record key contextual metadata such as the data source, equipment model version, and generation rules for each report, nor does it establish an index linking report content and the generation chain. Therefore, once a report is exported, subsequent iterations and updates cannot trace back to the original basis, and collaborative reviews cannot restore the generation logic. Ultimately, this completely separates the report content from the generation process, making it difficult to support the high-frequency version iteration and team collaboration needs in defense intelligence work. This results in a problem of low reliability in the automatic generation of defense equipment intelligence reports. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a method for automatically generating defense equipment intelligence reports based on a large language model, which can improve the reliability of automatically generating defense equipment intelligence reports.

[0006] The technical solution of this invention is implemented as follows: Based on the knowledge graph of the defense field, optimize the large language model, access multi-source heterogeneous intelligence data and complete preprocessing through multimodal extraction; associate intelligence elements with the knowledge graph to generate structured conclusions, drive the model to generate compliant intelligence reports; form a closed-loop iteration through multi-granularity filling and consistency verification, trigger the full life cycle traceability optimization of the report as needed, continuously improve the model's domain generation capability, and improve the reliability of automatic intelligence report generation for defense equipment.

[0007] This invention provides a method for automatically generating defense equipment intelligence reports based on a large language model. The method includes: Step 1, based on a constructed defense domain knowledge graph, performing domain pre-training and multiple rounds of fine-tuning on the original large language model to obtain a large language model adapted to the defense equipment intelligence scenario; accessing multi-source heterogeneous intelligence data through a real-time data interface and extracting intelligence elements from the multi-source heterogeneous intelligence data to complete data cleaning, format unification, and standardized annotation; Step 2, associating the preprocessed intelligence elements with the defense domain knowledge graph in real time to generate structured intelligence conclusions; and generating instructions to trigger the large language model based on the structured intelligence conclusions to drive the large model to automatically... Step 3: Generate a report outline and complete content that conforms to the defense equipment intelligence specifications. Step 4: Perform multi-granularity content filling and data consistency verification on the automatically generated report outline and complete content that conforms to the defense equipment intelligence specifications. Feed the verification results back to the large language model to form a closed-loop iterative mechanism to continuously improve the model's domain generation capabilities. Step 5: Based on the verification results of multi-granularity content filling and data consistency, determine whether to trigger the report's full lifecycle source tracing optimization. If so, perform the report's full lifecycle source tracing optimization and synchronously feed it back to the large language model after the source tracing optimization. If not, directly synchronously feed it back to the large language model to continuously improve the model's defense equipment intelligence domain generation capabilities.

[0008] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. By constructing a knowledge graph in the defense domain and combining it with dynamic pre-training and multi-round fine-tuning mechanisms, the adaptability of the large language model in defense equipment intelligence scenarios has been significantly improved. On the one hand, based on the entity association index library and feature standardization mapping technology of the knowledge graph, accurate parsing and association of multi-source heterogeneous intelligence data (text, images, signals, etc.) have been achieved, solving the problems of fragmented intelligence element extraction and semantic alignment difficulties in traditional methods, and ensuring efficient matching of intelligence elements with core nodes such as equipment models and deployment status. On the other hand, by dynamically adjusting the pre-training sampling frequency through weight changes and optimizing the task switching ratio in stages, the model can specifically strengthen entity recognition and relational reasoning capabilities, avoiding capability imbalance caused by single-scenario training. This allows the model's professional understanding and adaptability in the defense domain to far exceed that of general large models, laying a high-quality model foundation for subsequent report generation.

[0009] 2. A standardized intelligence report generation process has been established, significantly improving the quality, efficiency, and standardization of defense equipment intelligence reports. In the report generation stage, the precise integration of scenario-based instruction templates and structured intelligence conclusions ensures that the generated report outline and complete content strictly adhere to the professional specifications and format requirements of different types of reports, such as feasibility studies and weekly situation reports. Furthermore, all intelligence conclusions can be traced back to knowledge graph nodes and reasoning bases, further guaranteeing the report's authority. In the content filling and verification stage, multi-granularity filling dynamically adjusts the hierarchical division threshold to adapt to different precision requirements, avoiding redundant filling or missing information. Data consistency verification achieves a balance between batch processing efficiency and data integrity through time window alignment tolerance optimization. This effectively solves problems such as non-standard content, data contradictions, and insufficient precision in traditional report generation, significantly reducing manual intervention costs and improving report generation efficiency.

[0010] 3. Through a closed-loop iteration and full lifecycle traceability optimization mechanism, continuous improvement in model capabilities and report quality has been achieved, ensuring the stability and reliability of defense equipment intelligence applications. By feeding back multi-granularity filling and data consistency verification results to the model, a normalized closed-loop iteration is formed, driving the model to continuously adapt to the dynamic needs of defense intelligence scenarios. Simultaneously, based on the cumulative percentage of deviations exceeding thresholds at the granular level, traceability optimization is triggered. By dynamically adjusting the report generation lifecycle threshold, output quality fluctuations are corrected in a timely manner, ensuring that continuously generated reports meet the stringent requirements of defense intelligence applications in terms of stability and reliability. This continuous optimization mechanism not only solves the pain point of traditional models having fixed capabilities after training and difficulty adapting to changing scenarios, but also makes the entire report generation process traceable and optimizable, providing long-term, stable, and high-quality intelligence support for defense decision-making. Attached Figure Description

[0011] Figure 1 This is a flowchart of the automatic defense equipment intelligence report generation method based on a large language model provided in this embodiment of the invention; Figure 2 This is a flowchart of the domain pre-training process for the automatic report generation method for defense equipment intelligence based on a large language model provided in this embodiment of the invention. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0014] This invention provides a method for automatically generating defense equipment intelligence reports based on a large language model, such as... Figure 1 The flowchart shown is for an automatic defense equipment intelligence report generation method based on a large language model. The processing flow of this method may include the following steps: Step 1: Based on the existing defense domain knowledge graph, the original large language model is pre-trained and fine-tuned multiple times to obtain a large language model adapted to the intelligence scenario of defense equipment. Through a real-time data interface, multi-source heterogeneous intelligence data is accessed. Multi-modal information extraction technology is used to extract intelligence elements such as entities, attributes, and relationships from the multi-source heterogeneous intelligence data to complete data cleaning, format unification, and standardized labeling. Multi-source heterogeneous intelligence data includes, but is not limited to, text reports, satellite imagery, radar signals, open-source intelligence streams, and sensor data.

[0015] It should be understood that, such as Figure 2 The diagram shows the domain pre-training flowchart of the automatic defense equipment intelligence report generation method based on a large language model provided in this embodiment of the invention. The specific process is as follows: First, the model's weight change value is obtained. Then, the core judgment step, namely "judging the interval to which the weight change value belongs," is entered. When the weight change value is greater than or equal to the upper limit of the weight change reference, the system determines that the model's ability to match and recognize entities and text in the defense domain knowledge graph needs to be strengthened. At this time, the weight change value is input into the weight change value-sampling frequency adjustment coefficient mapping table, and the sampling frequency threshold upward adjustment coefficient is output. Then, the sampling frequency baseline threshold and the upward adjustment coefficient are combined to obtain the target relationship reasoning task sampling frequency, and the process ends. When the weight change value is less than or equal to the lower limit of the weight change reference, the system determines that the model's ability to reason about the relationships between entities in the defense domain knowledge graph is weaker than its entity recognition ability. At this time, the weight change value is input into the weight change value-sampling frequency adjustment coefficient mapping table, and the sampling frequency threshold downward adjustment coefficient is output. Then, the sampling frequency baseline threshold and the downward adjustment coefficient are combined to obtain the target relationship reasoning task sampling frequency, and the process ends. When the weight change value is within the open interval formed by the lower and upper limits of the weight change reference, the system determines that the entity recognition capability and relation reasoning capability of the model are in a matching state. At this time, there is no need to adjust the sampling frequency, and the current relation reasoning task sampling frequency is maintained directly, and the process ends.

[0016] It needs to be explained that the specific steps for performing domain pre-training on the original large language model are as follows: In the real-time detection domain, the weight change value of the entity link prediction task during pre-training is used to establish a mapping table of weight change value and sampling frequency adjustment coefficient. The weight change value of the entity link prediction task represents the difference between the entity link prediction task weight in the next round of training and the current entity link prediction task weight. If the weight change value is greater than or equal to the upper limit of the weight change reference, it indicates that the model's ability to match and recognize entities and text in the knowledge graph of the defense domain needs to be strengthened. Input the current weight change value into the weight change value-sampling frequency adjustment coefficient mapping table, output the sampling frequency threshold adjustment coefficient, and combine the sampling frequency baseline threshold and the sampling frequency threshold adjustment coefficient to obtain the sampling frequency of the target relationship reasoning task. If the weight change value is less than or equal to the lower limit of the weight change reference, it indicates that the model's reasoning ability for the relationship between entities in the knowledge graph of the defense domain is weaker than its entity recognition ability. The current weight change value is input into the weight change value-sampling frequency adjustment coefficient mapping table, and the sampling frequency threshold reduction coefficient is output. The sampling frequency benchmark threshold and the sampling frequency threshold reduction coefficient are combined to obtain the sampling frequency of the target relationship reasoning task.

[0017] Domain pre-training of the original large language model also includes: If the weight change value is within the weight change reference interval, it indicates that the model's entity recognition and relational reasoning capabilities match, and the current relational reasoning task sampling frequency is maintained. The weight change reference interval represents the open interval formed by the lower limit and upper limit of the weight change reference.

[0018] In this embodiment, during the pre-training process in the real-time detection domain (specifically the defense domain in this scenario), it is first necessary to monitor the weight changes of the entity link prediction task in real time. Specifically, this weight change is defined as the difference between the entity link prediction task weight in the next training round and the entity link prediction task weight in the current round. This provides a precise quantitative basis for subsequent dynamic adjustments, avoiding the drawbacks of traditional fixed-weight training where changes in model capabilities cannot be detected in a timely manner, and ensuring the real-time nature and targeted nature of adjustment decisions. Simultaneously, a mapping table of weight change values ​​and sampling frequency adjustment coefficients needs to be established in advance. This mapping table serves as the core association carrier, enabling rapid matching and conversion between weight change values ​​and sampling frequency adjustment coefficients, providing support for the rapid calculation of subsequent sampling frequencies, significantly improving the efficiency of the adjustment process, and avoiding delays and errors caused by manual intervention.After real-time detection of weight changes, the model enters a dynamic adjustment decision-making phase. If the detected weight change is greater than or equal to the preset upper limit of the weight change reference, this signal indicates that the current model's ability to match and recognize entities and text in the defense domain knowledge graph is insufficient and needs further strengthening. In this case, the current weight change is input into a pre-constructed weight change-sampling frequency adjustment coefficient mapping table. The mapping table quickly outputs the corresponding sampling frequency threshold adjustment coefficient. Then, the sampling frequency baseline threshold and this sampling frequency threshold adjustment coefficient are combined (e.g., through multiplication or other reasonable numerical combinations) to finally obtain the sampling frequency for the target relation reasoning task. By increasing the sampling frequency for the relation reasoning task, the training proportion of the model on relation reasoning tasks is increased, thereby balancing the model's entity recognition and relation reasoning capabilities, specifically addressing the shortcomings in entity and text matching and recognition capabilities, and improving the model's accurate matching ability for defense domain knowledge. If the detected weight change is less than or equal to the preset lower limit of the weight change reference, this signal indicates that the current model's ability to reason about the relationships between entities in the defense domain knowledge graph is weaker than its entity recognition ability, indicating a capability imbalance. In this case, the current weight change is also adjusted. The input values ​​are fed into the weight change value-sampling frequency adjustment coefficient mapping table, and the corresponding sampling frequency threshold reduction coefficient is output. Then, the sampling frequency baseline threshold and the sampling frequency threshold reduction coefficient are combined to obtain the target relational reasoning task sampling frequency. By reducing the sampling frequency of the relational reasoning task, its training proportion is appropriately reduced to avoid overtraining of the reasoning task and further exacerbating the ability imbalance. At the same time, more training resources are reserved for consolidating and optimizing entity recognition ability, promoting the balanced development of the model's two core capabilities. If the detected weight change value is within the weight change reference interval (this interval specifically refers to the open interval formed by the lower limit and upper limit of the weight change reference), this numerical signal indicates that the current model's entity recognition ability and relational reasoning ability are in a matching state. There is no need to adjust the sampling frequency; the current relational reasoning task sampling frequency can be maintained. This maintains the stability of training when the model's capabilities are balanced, avoids unnecessary adjustments that may interfere with the model's training effect, and ensures that the model can continuously and stably improve its comprehensive performance during the defense domain pre-training process. Ultimately, this achieves accurate adaptation of the model to entity recognition, entity matching, and relational reasoning in the defense domain knowledge graph, improving the overall quality and efficiency of domain pre-training.

[0019] It should be further explained that the specific steps of the multi-round fine-tuning are as follows: A threshold for the cross-round task switching ratio is set to characterize the switching ratio of defense intelligence task samples between adjacent rounds in multi-round fine-tuning. Based on the total number of fine-tuning rounds, the threshold for the cross-round task switching ratio is dynamically adjusted in stages. Multi-round fine-tuning is divided into three stages based on the total number of fine-tuning rounds, and a mapping relationship between the total number of rounds and the initial threshold adjustment coefficient is established, specifically: If the total number of fine-tuning rounds is less than or equal to the critical lower limit of the number of rounds, it is divided into the basic alignment stage. Then, the current total number of fine-tuning rounds is input into the mapping relationship between the total number of rounds and the initial threshold adjustment coefficient, and the initial threshold adjustment coefficient is output. The initial threshold of the cross-round task switching ratio and the initial threshold adjustment coefficient are combined to obtain the target cross-round task switching ratio threshold.

[0020] Multiple rounds of fine-tuning also included: If the total number of rounds is fine-tuned within the critical interval of the number of rounds, it is divided into the capability deepening stage. The current target cross-round task switching ratio threshold is maintained. The critical interval of the number of rounds represents the open interval formed by the critical lower limit of the number of rounds and the critical upper limit of the number of rounds. If the total number of fine-tuning rounds is greater than or equal to the critical upper limit of the number of rounds, it is divided into the convergence optimization stage. The current total number of fine-tuning rounds is input into the mapping relationship between the total number of rounds and the initial threshold adjustment coefficient. The initial threshold reduction coefficient is output. The initial threshold of the cross-round task switching ratio and the initial threshold reduction coefficient are combined to obtain the target cross-round task switching ratio threshold.

[0021] In this embodiment, during the multi-round fine-tuning of defense intelligence, a threshold for the cross-round task switching ratio must first be set. The core function of this threshold is to characterize the switching ratio of defense intelligence task samples between adjacent rounds during multi-round fine-tuning. This provides a clear quantitative judgment standard for the switching of task samples between adjacent rounds, avoiding disordered and unbalanced task switching between rounds, ensuring the standardization and rationality of task switching during fine-tuning, and laying the foundation for the model to stably learn defense intelligence-related knowledge. To achieve scientific adaptation and dynamic optimization of the threshold, the cross-round task switching ratio threshold needs to be dynamically adjusted in stages based on the total number of fine-tuning rounds. Specifically, the multi-round fine-tuning is divided into three core stages: basic alignment, capability deepening, and convergence optimization, based on the total number of fine-tuning rounds. Simultaneously, a mapping relationship between the total number of rounds and the initial threshold adjustment coefficient is established in advance, constructing a precise correlation bridge between the total number of rounds and the threshold adjustment coefficient. This provides efficient support for the rapid calculation and dynamic adaptation of thresholds at each stage, overcoming the limitation of fixed thresholds adapting to all fine-tuning stages, and achieving targeted and flexible threshold adjustment.After establishing the phase division criteria and core mapping relationship, the threshold dynamic adjustment work is carried out according to the following specific steps: If the current total number of fine-tuning rounds is less than or equal to the preset critical lower limit of the number of rounds, then the round is divided into the basic alignment phase. The core objective of this phase is to enable the model to quickly adapt to the basic requirements of defense intelligence missions and achieve basic capability alignment. At this time, the current total number of fine-tuning rounds is input into the pre-established mapping relationship between the total number of rounds and the initial threshold adjustment coefficient. The corresponding initial threshold adjustment coefficient is quickly output through the mapping relationship. Then, the initial threshold of the cross-round task switching ratio and the initial threshold adjustment coefficient are reasonably combined (such as through multiplication or other numerical operations) to finally obtain the result for this phase. The target is to set a threshold for the cross-round task switching ratio. By increasing this threshold, the switching ratio of defense intelligence task samples between adjacent rounds is improved. This allows the model to access more diverse task samples during the basic alignment phase, quickly familiarize itself with the core scenarios and basic requirements of defense intelligence tasks, and efficiently complete the alignment and consolidation of basic capabilities. This avoids adaptation lag issues caused by the limited number of task samples in the initial stage. If the current total number of rounds for fine-tuning is within the critical interval of the number of rounds (specifically, the open interval formed by the lower and upper critical limits of the number of rounds), then this round is classified as the capability deepening phase. The core objective of this phase is to consolidate the model's basic capabilities and deepen its core business adaptation capabilities. In this case, there is no need to adjust the threshold. Maintaining the current target cross-round task switching ratio threshold is sufficient. During the capability enhancement phase, the threshold should be kept stable. A stable cross-round task switching ratio provides a continuous and coherent training environment for the model, allowing it to gradually solidify its basic capabilities and deepen its understanding and adaptation to the core logic of defense intelligence tasks within a stable task sample input rhythm. This avoids training rhythm disruption caused by frequent threshold adjustments, ensuring the continuity and effectiveness of capability enhancement. If the current total number of fine-tuning rounds is greater than or equal to the preset critical upper limit for the number of rounds, then this round is classified as a convergence optimization phase. The core objective of this phase is to optimize model performance and achieve stable convergence of model results. At this point, the current total number of fine-tuning rounds is adjusted. The mapping relationship between the total number of rounds and the initial threshold adjustment coefficient is input, and the corresponding initial threshold reduction coefficient is output. Then, the initial threshold for the cross-round task switching ratio is combined with the initial threshold reduction coefficient to obtain the target cross-round task switching ratio threshold for this stage. The technical effect of this step is to reduce the switching ratio of defense intelligence task samples between adjacent rounds by lowering the threshold, so that the model can focus on core task samples and key scenarios during the convergence optimization stage, reduce the interference caused by frequent switching of task samples, concentrate training resources to optimize and refine core capabilities, promote the rapid convergence of model performance, and ensure that the final output model can stably adapt to the needs of defense intelligence tasks, thereby improving the overall effect and efficiency of fine-tuning.

[0022] Step two involves linking the pre-processed intelligence elements with nodes such as equipment models, deployment status, and combat systems in the knowledge graph of the defense domain in real time to generate structured intelligence conclusions. Based on these structured intelligence conclusions, instructions are generated to trigger the large language model, which in turn drives the large model to automatically generate a report outline and complete content that conforms to the defense equipment intelligence specifications.

[0023] It should be understood that constructing an entity association index library for a defense domain knowledge graph involves three core node types: equipment model, deployment status, and combat system. Each node is bound to corresponding attribute features and association rules. Specifically, the equipment model node is bound to attributes such as technical parameters, R&D background, and performance indicators, as well as association rules for equipment-equipment and equipment-deployment. The deployment status node is bound to attributes such as deployment area, deployment scale, and deployment time, as well as association rules for deployment-troop and deployment-operation area. The combat system node is bound to attributes such as system composition, combat mission, and coordination logic, as well as association rules for system-equipment and system-deployment. Preprocessed intelligence elements undergo feature extraction and standardization mapping. Entity features, attribute features, spatiotemporal features, and association features are extracted from the intelligence elements. Following the node attribute specifications of the entity association index library, each feature is standardized and mapped into a feature vector that matches the nodes of the defense domain knowledge graph, ensuring dimensional uniformity and semantic alignment between intelligence element features and knowledge graph node features. Based on a vector similarity matching algorithm, the standardized mapped intelligence element feature vectors are compared with those in the entity association index library. The feature vectors of equipment models, deployment status, and combat system nodes are used for real-time similarity calculation. A similarity matching threshold of 0.8-0.95 is set. Intelligence elements with similarity higher than the threshold are precisely associated with corresponding knowledge graph nodes, while those with similarity lower than the threshold are marked as anomalies and included in the manual review process. Multi-dimensional intelligence reasoning is conducted based on the association results. Following the logic of defense equipment intelligence analysis, equipment performance reasoning, deployment dynamic reasoning, and combat status reasoning are performed sequentially. Equipment performance reasoning, based on the association results between intelligence elements and equipment model nodes, combined with the association rules between equipment technical parameters and performance in the knowledge graph, derives the actual combat capability of the equipment. Deployment dynamic reasoning, based on the association results between intelligence elements and deployment status nodes, combined with the association rules between deployment scale and region in the knowledge graph, derives the tactical intent of equipment deployment. Combat status reasoning, based on the association results between equipment performance reasoning, deployment dynamic reasoning, and combat system nodes, combined with the combat system coordination logic in the knowledge graph, derives the overall defense equipment status. Structured intelligence conclusions are generated based on the multi-dimensional intelligence reasoning results. These structured intelligence conclusions use triple entity- The standardized format for constructing relationships and attributes includes conclusions at the equipment, deployment, and operational status levels. Each type of conclusion is labeled with its intelligence source, data credibility, and reasoning basis, with data credibility ranging from 0.0 to 1.0. Quantitative scoring, with reasoning clearly linked to specific nodes and rules in the defense domain knowledge graph; constructing a scenario-based command generation template library, pre-setting command templates for different types of defense equipment intelligence reports such as feasibility study reports and weekly situation reports. Each command template includes four core modules: report type identifier, core intelligence dimension, content generation specifications, and format requirements. The core intelligence dimension corresponds one-to-one with the equipment, deployment, and operational status dimensions of the structured intelligence conclusion. The content generation specifications match the professional writing requirements of defense equipment intelligence, and the format requirements match the chapter division and content arrangement rules of the report outline. Based on template matching and content filling, a large language model is generated to trigger commands. According to the target report type, the corresponding command template is retrieved from the scenario-based command generation template library, and the specific content of each dimension in the structured intelligence conclusion is generated according to the core intelligence dimension of the template. The goal is to accurately fill in the corresponding positions in the template, while embedding report type identifiers, content generation specifications, and format requirements to form standardized instructions that can directly trigger the large language model. These instructions include clear generation objectives, core intelligence evidence, professional standard requirements, and format constraints. The generated standardized instructions are then input into the large language model adapted to defense equipment intelligence scenarios. The model generates content based on the core intelligence evidence in the instructions, ensuring the professionalism and logical coherence of the writing based on content generation specifications, and generating a report outline that conforms to defense equipment intelligence standards based on format requirements. Finally, based on the report outline and structured intelligence conclusions, a complete defense equipment intelligence report with comprehensive content, complete dimensions, and rigorous logic is generated. Furthermore, all intelligence conclusions in the generated report outline and complete content can be traced back to structured intelligence conclusions and corresponding defense domain knowledge graph nodes.

[0024] Step 3: Perform multi-granularity content filling and data consistency verification on the automatically generated report outline and complete content that conform to the defense equipment intelligence specifications. Feed back the verification results to the large language model to form a closed-loop iterative mechanism to continuously improve the model's domain generation capabilities.

[0025] It should be noted that the specific steps for performing multi-granularity content population are as follows: Step 101: Obtain the actual filling step length during the first filling process. The actual filling step length is the single sliding span of the filling unit on the granular content carrier. Based on the current filling step length, establish a mapping relationship between the filling step length and the adjustment amount of the granular level division threshold, and dynamically adjust the granular level division threshold. Step 102: If the current fill step size is less than or equal to the lower bound of the step size reference, it indicates that the corresponding fill granularity level division threshold is set too low and the fill granularity is too fine, which has caused the fill efficiency to decrease and there is a lot of invalid redundant fill. It cannot adapt to the actual fill accuracy requirements of the granular content. Input the current fill step size into the fill step size-level division threshold adjustment amount mapping relationship, output the level division threshold gain amount, and superimpose the initial threshold of fill granularity level division with the level division threshold gain amount to obtain the target fill granularity level division threshold. Step 103: If the current fill step size is greater than or equal to the upper limit of the step size reference, the corresponding fill granularity level division threshold is set too high, the fill granularity is too coarse, and the fill information matching degree is seriously insufficient, which cannot meet the minimum fill accuracy requirements of the granular content. Input the current fill step size into the fill step size-level division threshold adjustment mapping relationship, output the level division threshold reduction amount, and superimpose the initial threshold of fill granularity level division with the level division threshold reduction amount to obtain the target fill granularity level division threshold.

[0026] Performing multi-granularity content filling also includes: Step 104: If the current filling step size is within the step size reference interval, it indicates that the current filling step size is adapted to the basic filling requirements. Maintain the initial threshold for the filling granularity level division. The step size reference interval represents the open interval formed by the lower bound of the step size reference and the upper bound of the step size reference. Step 105: Based on the target filling granularity level division threshold and the current filling step size, execute the next round of granularity content iteration filling, repeating steps 101-104 to ensure the continuity and rationality of the thresholds at each level.

[0027] In this embodiment, when performing the multi-granularity content filling process, the following steps are required to monitor the filling step length and dynamically adjust the filling granularity level division threshold: Step 101, firstly, the actual filling step length during the initial filling process is obtained. The actual filling step length is specifically defined as the single sliding span of the filling unit on the granular content carrier. Its core technical effect is to provide accurate quantitative input basis for subsequent threshold adjustment, realize targeted adjustment based on actual filling conditions, and avoid blind threshold setting that is detached from the actual filling. After obtaining the current filling step length, a mapping relationship between the filling step length and the level division threshold adjustment amount is established based on the current filling step length. The granularity level division threshold is dynamically adjusted through this mapping relationship. The construction of this mapping relationship is to build a precise correlation bridge between the filling step length and the threshold adjustment amount, realize the rapid matching and output of the threshold adjustment amount, greatly improve the efficiency and accuracy of threshold adjustment, and get rid of the drawbacks of traditional fixed thresholds that cannot adapt to dynamic filling scenarios. Step 102: After completing the step size acquisition and mapping relationship construction, step size interval judgment is performed: If the current fill step size is less than or equal to the preset step size reference lower bound, the numerical signal clearly indicates that the corresponding fill granularity level division threshold is set too low, resulting in excessively fine fill granularity. Such excessively fine fill granularity will not only significantly reduce fill efficiency, but also generate a large amount of invalid redundant fill, ultimately failing to meet the actual fill accuracy requirements of the granular content. At this time, the current fill step size is input into the pre-established fill step size-level division threshold adjustment mapping relationship. The corresponding level division threshold gain is quickly output through the mapping relationship. Then, the initial threshold of fill granularity level division is superimposed with the level division threshold gain (i.e., the initial threshold plus the gain) to finally obtain the target fill granularity level division threshold. The threshold gain is used to adjust the fill granularity level division threshold upward, thereby thickening the fill granularity, effectively reducing invalid redundant fill, improving fill efficiency, and ensuring that the fill accuracy can meet the actual requirements, balancing the relationship between fill efficiency and fill accuracy. Step 103: If the current fill step size is greater than or equal to the preset upper limit of the step size reference, the numerical signal indicates that the corresponding fill granularity level division threshold is set too high, resulting in an excessively coarse fill granularity. An excessively coarse fill granularity will cause a serious lack of fill information matching and will not meet the minimum fill accuracy requirements of the granular content. At this time, the current fill step size is also input into the fill step size-level division threshold adjustment mapping relationship, and the corresponding level division threshold reduction amount is output. Then, the initial threshold of the fill granularity level division is superimposed with the level division threshold reduction amount (i.e., the initial threshold minus the reduction amount) to obtain the target fill granularity level division threshold. The threshold reduction achieves the lowering of the fill granularity level division threshold, refines the fill granularity, thereby improving the matching degree of the fill information, ensuring that the fill effect can meet the minimum fill accuracy requirements of the granular content, and guaranteeing the fill quality.The multi-granularity content filling process also includes step 104: If the current filling step size is within the step size reference interval (the step size reference interval specifically refers to the open interval formed by the lower and upper bounds of the step size reference), the numerical signal indicates that the current filling step size can well adapt to the basic filling requirements. At this time, there is no need to adjust the filling granularity level division threshold. It is sufficient to maintain the initial threshold of the filling granularity level division. Under the condition of filling step size adaptation, the threshold stability is maintained, avoiding unnecessary threshold adjustments from interfering with the continuity of the filling process, ensuring that the filling work can proceed smoothly and efficiently, and ensuring that the filling accuracy and efficiency are in a balanced state. Step 105: After determining the target filling granularity level division threshold, the next round of granular content iterative filling operation is performed based on the target threshold and the current filling step size. Then, the complete process from Step 101 to Step 104 is repeated. This achieves continuous dynamic adaptation between the filling step size and the filling granularity level division threshold, ensuring the continuity and rationality of the level threshold in each round of filling, avoiding adaptation deviations caused by single-round adjustments, and ultimately achieving controllable accuracy and optimal efficiency throughout the multi-granularity content filling process, ensuring the quality and stability of the overall filling work.

[0028] It should be further explained that the specific steps for data consistency verification are as follows: The target incremental data set for the first round of incremental updates is collected, the timestamp of each data is extracted, the first round time window interval is divided based on the baseline time window as the period, the baseline time window is divided into sub-time windows on an average basis based on the initial incremental update batch, and the actual effective interval of each sub-time window is determined based on the alignment tolerance of the initial time window. The first round time window interval represents the closed interval formed by the start timestamp and end timestamp of the first round of updates. The target incremental data is assigned to the corresponding sub-time window of the actual effective interval according to the timestamp to complete the first round of batch division and processing. The batch processing efficiency during the first round of processing is recorded, and a batch processing efficiency-alignment tolerance adjustment factor mapping table is established. The batch processing efficiency represents the average efficiency of processing a single batch of data. If the batch processing efficiency is less than the lower limit of the processing efficiency benchmark, it indicates that the amount of incremental data to be processed in a single batch is too large and cannot meet the timeliness requirements of incremental updates of defense equipment intelligence. In this case, the current batch processing efficiency is input into the batch processing efficiency-alignment tolerance adjustment factor mapping table, and the alignment tolerance reduction factor is output. The alignment tolerance reduction factor is then interactively processed with the initial time window alignment tolerance to obtain the target time window alignment tolerance, so as to reduce the amount of data processed in a single batch and improve the overall batch processing efficiency.

[0029] Data consistency verification also includes: If the batch processing efficiency is within the processing efficiency benchmark range, it indicates that the current time window alignment tolerance and the matching degree of incremental update batches are suitable for the processing needs of incremental updates of defense equipment intelligence. In this case, the current time window alignment tolerance is maintained to ensure the stability of processing efficiency. The processing efficiency benchmark range represents the closed interval formed by the lower limit of the processing efficiency benchmark and the upper limit of the processing efficiency benchmark. If the batch processing efficiency is greater than the upper limit of the processing efficiency benchmark, it indicates that the amount of incremental data to be processed in a single batch is too small and the redundancy overhead of batch processing is too high. In this case, the current batch processing efficiency is input into the batch processing efficiency-alignment tolerance adjustment factor mapping table, and the alignment tolerance adjustment factor is output. The alignment tolerance adjustment factor is then interacted with the initial time window alignment tolerance to obtain the target time window alignment tolerance, so as to improve the amount of data processed in a single batch, optimize the utilization rate and ensure the integrity of the data.

[0030] In this embodiment, firstly, the target incremental data set for the first round of incremental updates is collected, and the timestamp corresponding to each data point is fully extracted. The technical effect of timestamp extraction is to provide a precise time dimension basis for subsequent time window division and data classification, ensuring the timeliness and accuracy of data processing, and meeting the stringent requirements of defense equipment intelligence data for time correlation. Subsequently, the first round of time window intervals are divided with a base time window as a fixed period. Specifically, the first round of time window intervals represents the closed interval formed by the start timestamp and end timestamp of the first round of updates. The technical effect is to clarify the data time range of the first round of incremental updates, providing a basis for subsequent sub-window division and data batch processing. The process involves clearly defining boundaries to avoid data omissions or processing across different ranges. Based on a preset initial incremental update batch, the baseline time window is divided into several sub-time windows. This division effectively achieves refined batch splitting of the first round of incremental data, laying the foundation for efficient batch processing and avoiding efficiency bottlenecks caused by excessively large batch sizes. Simultaneously, based on the initial time window alignment tolerance, the actual effective range of each sub-time window is determined. The introduction of alignment tolerance and the definition of the effective range accommodate minor deviations in data timestamps, preventing data classification errors caused by slight timestamp fluctuations and ensuring the accuracy and fault tolerance of batch division. After defining the sub-time windows and their effective intervals, the target incremental data is assigned to the corresponding sub-time windows of the effective intervals according to their timestamps. This completes the first round of batch division and processing, enabling the orderly distribution of incremental data and allowing data processing to proceed according to batch specifications, thus improving the standardization and controllability of the processing flow. Simultaneously, the batch processing efficiency during the first round of processing is accurately recorded. This efficiency indicator specifically represents the average efficiency of processing a single batch of data. The technical effect of this recording is to provide a core quantitative evaluation basis for subsequent tolerance adjustments, enabling dynamic optimization decisions based on actual processing efficiency. Based on the recorded batch processing efficiency data, a batch processing efficiency-alignment tolerance adjustment factor mapping table is established. The technical effect of this mapping table is to build a precise correlation bridge between efficiency indicators and tolerance adjustment factors, enabling rapid matching and output of adjustment factors, significantly improving the efficiency and targeting of tolerance adjustments, and overcoming the shortcomings of traditional fixed tolerances that cannot adapt to dynamic processing scenarios.After batch processing and efficiency recording are completed, the dynamic adjustment of the time window alignment tolerance based on batch processing efficiency begins. If the current batch processing efficiency is less than the preset lower limit of the processing efficiency benchmark, this signal clearly indicates that the amount of incremental data to be processed in a single batch is too large, resulting in insufficient processing efficiency and failing to meet the stringent timeliness requirements of incremental updates for defense equipment intelligence (defense equipment intelligence needs to be updated quickly to support decision-making, and timeliness is a core requirement). At this point, the current batch processing efficiency is input into a pre-established batch processing efficiency-alignment tolerance adjustment factor mapping table. The corresponding alignment tolerance reduction factor is quickly output through the mapping table. Then, the alignment tolerance reduction factor and the initial time window alignment tolerance are reasonably interacted (e.g., through multiplication or other numerical calculations) to finally obtain the target time window alignment tolerance. The technical effect of this adjustment step is to reduce the actual effective range of each sub-time window by reducing the alignment tolerance, thereby reducing the amount of data to be processed in a single batch, avoiding data congestion, improving the overall batch processing efficiency, and ensuring that the incremental update work can be completed on time to meet the timeliness requirements. The data consistency verification process also includes the following two adjustment scenarios: If the current batch processing efficiency is within the processing efficiency benchmark range (specifically, the closed interval formed by the lower and upper limits of the processing efficiency benchmark), this numerical signal indicates that the current time window alignment tolerance and incremental update batch match well, accurately adapting to the processing needs of incremental updates of defense equipment intelligence. In this case, there is no need to adjust the alignment tolerance; simply maintain the current time window alignment tolerance. With stable processing efficiency and adaptation to requirements, maintaining the stability of the tolerance and batch division avoids unnecessary adjustments that could disrupt the continuity of the processing flow, while ensuring data processing consistency and reliability, and ensuring a balance between the quality and efficiency of each batch of data processing. If the current batch processing efficiency is greater than the preset processing efficiency benchmark upper limit, this numerical signal indicates that the current single-batch processing efficiency is within the processing efficiency benchmark range. If the amount of incremental data to be processed in a batch is too small, the redundancy overhead of batch processing will be too high (too many small batches will increase the redundancy costs of process connection, resource scheduling, etc.), resulting in wasted processing resources and potentially affecting data integrity. In this case, the current batch processing efficiency is input into the batch processing efficiency-alignment tolerance adjustment factor mapping table, and the corresponding alignment tolerance adjustment factor is output. Then, the alignment tolerance adjustment factor is interacted with the initial time window alignment tolerance to obtain the target time window alignment tolerance. By adjusting the alignment tolerance, the actual effective range of each sub-time window is expanded, thereby increasing the amount of data processed per batch, reducing redundancy overhead, optimizing the utilization rate of processing resources, and ensuring the integrity of data processing. This avoids the data fragmentation problem caused by excessively fine splitting of small batches, ensuring that the incremental update of defense equipment intelligence is carried out efficiently, completely, and economically.

[0031] Step 4: Based on the verification results of multi-granularity content filling and data consistency, determine whether the report lifecycle tracing optimization is triggered. If yes, perform the report lifecycle tracing optimization and synchronously feed it back to the big language model after the tracing optimization. If no, directly synchronously feed it back to the big language model to continuously improve the model's ability to generate defense equipment intelligence.

[0032] It should be noted that the specific steps for determining whether the report's full lifecycle tracing optimization has been triggered are as follows: A corresponding accuracy deviation threshold is preset for each granularity level. If the multi-granularity accuracy deviation of a level is greater than the accuracy deviation threshold, it is recorded as an over-threshold deviation for that level. During the continuous report generation process, the cumulative percentage of over-threshold deviation for each granularity level is calculated to characterize the stability and reliability of the overall output quality. If the cumulative percentage of deviations exceeding the threshold is less than or equal to the cumulative percentage benchmark value, then the full lifecycle source tracing optimization of the report will not be triggered. If the cumulative percentage of deviations exceeding the threshold is greater than the cumulative percentage benchmark value, the current cumulative percentage of deviations exceeding the threshold is input into the mapping relationship through the established cumulative percentage-lifetime threshold mapping relationship, and the target report generation lifetime threshold is output to reduce the report generation lifetime and restore the stability and reliability of the output.

[0033] In this embodiment, during the continuous generation of multi-granularity level reports, to accurately control the stability and reliability of the overall output quality, quality monitoring and optimization triggering work needs to be carried out according to the following specific steps. First, a corresponding accuracy deviation threshold is preset for each granularity level. This threshold serves as the core judgment standard for whether the accuracy of each level is qualified, providing a clear quantitative definition of the accuracy deviation of different granularity levels. This avoids the ambiguity in quality assessment caused by the lack of a unified judgment standard, ensuring the pertinence and standardization of accuracy monitoring at each level, and meeting the differentiated accuracy requirements of different levels in multi-granularity scenarios. During the continuous generation of reports, the multi-granularity accuracy deviation of each granularity level is monitored in real time. If the multi-granularity accuracy deviation of a certain level is greater than the preset accuracy deviation threshold for that level, it is recorded as an over-threshold deviation for that level. The technical effect of this judgment operation is to achieve real-time capture and recording of accuracy anomalies at each granularity level, promptly detect the accuracy out-of-control problem of a single level, accumulate accurate abnormal data for subsequent overall quality assessment, and avoid the overall output quality deterioration caused by the accumulation of accuracy deviations. Simultaneously, throughout the entire report generation process, the cumulative percentage of out-of-threshold deviations at each granular level is continuously calculated. This percentage is specifically the ratio of the cumulative number of out-of-threshold deviations at a particular level to the total number of reports generated at that level. This quantitatively characterizes the stability and reliability of the report output quality at each level and the overall system, transforming isolated out-of-threshold deviation events into macroscopic quality assessment indicators. This facilitates a direct assessment of whether the overall output quality is within a controllable range, providing core quantitative support for subsequent optimization decisions. After calculating the cumulative percentage of out-of-threshold deviations, the quality optimization trigger judgment stage begins: if the cumulative percentage of out-of-threshold deviations at a certain level is less than or equal to a preset cumulative percentage benchmark value, this numerical signal indicates that the stability and reliability of the report output quality at that level and the overall system are within acceptable limits. The current precision control measures effectively guarantee output quality, and no additional optimization process needs to be initiated. In this case, the report's entire lifecycle traceability optimization is not triggered. Under the condition of acceptable quality, the continuity and stability of the report generation process are maintained, avoiding unnecessary traceability optimization from interfering with generation efficiency. Simultaneously, optimization resources are saved, ensuring efficient progress in report generation.If the cumulative percentage of deviations exceeding the threshold at a certain level exceeds the preset cumulative percentage benchmark value, this numerical signal clearly indicates that the cumulative accuracy at that level is severely abnormal, affecting the stability and reliability of the overall report output quality. If not intervened in time, it may lead to a continuous decline in the quality of subsequent reports. At this time, an optimization process needs to be initiated: through the pre-constructed cumulative percentage-lifecycle threshold mapping relationship (this mapping relationship has been pre-configured to accurately associate the cumulative percentage of deviations exceeding the threshold with the report generation lifecycle threshold), the current cumulative percentage of deviations exceeding the threshold is input into this mapping relationship. The mapping relationship will quickly output the corresponding target report generation lifecycle threshold, which is shorter than the original lifecycle threshold. By reducing the report generation lifecycle, the time consumption and process span of each stage of report generation are compressed, the time window for the accumulation of accuracy deviations is reduced, the core problem causing the deviations exceeding the threshold is quickly located and corrected, thereby restoring the stability and reliability of the report output, ensuring that the subsequently generated reports can meet the preset accuracy and quality requirements, and ensuring that the overall output quality is always within a controllable range.

[0034] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0035] It should be understood that determining B based on A does not mean determining B solely based on A; it also means determining B based on A and / or other information.

[0036] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0037] The above description is only an optional embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for automatically generating defense equipment intelligence reports based on a large language model, characterized in that, The method includes: Step 1: Based on the existing defense domain knowledge graph, the original large language model is pre-trained and fine-tuned in multiple rounds to obtain a large language model that is adapted to the intelligence scenario of defense equipment. Through the real-time data interface, multi-source heterogeneous intelligence data is accessed, and intelligence elements in the multi-source heterogeneous intelligence data are extracted to complete data cleaning, format unification and standardized labeling. Step 2: The pre-processed intelligence elements are linked with the knowledge graph of the defense domain in real time to generate structured intelligence conclusions. Based on the structured intelligence conclusions, instructions are generated to trigger the big language model, so as to drive the big model to automatically generate a report outline and complete content that conforms to the intelligence specifications of defense equipment. Step 3: Perform multi-granularity content filling and data consistency verification on the automatically generated report outline and complete content that conform to the defense equipment intelligence specifications, and feed the verification results back to the large language model to form a closed-loop iterative mechanism to continuously improve the model's domain generation capabilities. Step 4: Based on the verification results of the multi-granularity content filling and data consistency, determine whether the report lifecycle tracing optimization is triggered. If yes, perform the report lifecycle tracing optimization and synchronously feed it back to the large language model after the tracing optimization. If no, directly feed it back to the large language model to continuously improve the model's ability to generate defense equipment intelligence.

2. The method for automatically generating defense equipment intelligence reports based on a large language model as described in claim 1, characterized in that, The specific steps for performing domain pre-training on the original large language model are as follows: In the real-time detection domain pre-training process, the weight change value of the entity link prediction task is obtained, and a mapping table of weight change value and sampling frequency adjustment coefficient is established. The weight change value of the entity link prediction task represents the difference between the entity link prediction task weight in the next round of training and the current entity link prediction task weight. If the weight change value is greater than or equal to the upper limit of the weight change reference, it indicates that the model's ability to match and recognize entities and text in the knowledge graph of the defense domain needs to be strengthened. Input the current weight change value into the weight change value-sampling frequency adjustment coefficient mapping table, output the sampling frequency threshold adjustment coefficient, and combine the sampling frequency baseline threshold and the sampling frequency threshold adjustment coefficient to obtain the sampling frequency of the target relationship reasoning task. If the weight change value is less than or equal to the lower limit of the weight change reference, it indicates that the model's reasoning ability for the relationship between entities in the knowledge graph of the defense domain is weaker than its entity recognition ability. The current weight change value is input into the weight change value-sampling frequency adjustment coefficient mapping table, and the sampling frequency threshold reduction coefficient is output. The sampling frequency benchmark threshold and the sampling frequency threshold reduction coefficient are combined to obtain the sampling frequency of the target relationship reasoning task.

3. The method for automatically generating defense equipment intelligence reports based on a large language model as described in claim 2, characterized in that, The domain pre-training of the original large language model also includes: If the weight change value is within the weight change reference interval, it indicates that the model's entity recognition and relational reasoning capabilities match, and the current relational reasoning task sampling frequency is maintained. The weight change reference interval represents the open interval formed by the lower limit of the weight change reference and the upper limit of the weight change reference.

4. The method for automatically generating defense equipment intelligence reports based on a large language model as described in claim 1, characterized in that, The specific steps of the multi-round fine-tuning are as follows: A threshold for the cross-round task switching ratio is set to characterize the switching ratio of defense intelligence task samples between adjacent rounds in multi-round fine-tuning. Based on the total number of fine-tuning rounds, the threshold for the cross-round task switching ratio is dynamically adjusted in stages. Multi-round fine-tuning is divided into three stages based on the total number of fine-tuning rounds, and a mapping relationship between the total number of rounds and the initial threshold adjustment coefficient is established, specifically: If the total number of fine-tuning rounds is less than or equal to the critical lower limit of the number of rounds, it is divided into the basic alignment stage. Then, the current total number of fine-tuning rounds is input into the mapping relationship between the total number of rounds and the initial threshold adjustment coefficient, and the initial threshold adjustment coefficient is output. The initial threshold of the cross-round task switching ratio and the initial threshold adjustment coefficient are combined to obtain the target cross-round task switching ratio threshold.

5. The method for automatically generating defense equipment intelligence reports based on a large language model as described in claim 4, characterized in that, The multi-round fine-tuning also includes: If the total number of rounds is fine-tuned within the critical interval of the number of rounds, it is divided into the capability deepening stage, and the current target cross-round task switching ratio threshold is maintained. The critical interval of the number of rounds refers to the open interval formed by the critical lower limit of the number of rounds and the critical upper limit of the number of rounds. If the total number of fine-tuning rounds is greater than or equal to the critical upper limit of the number of rounds, it is divided into the convergence optimization stage. The current total number of fine-tuning rounds is input into the mapping relationship between the total number of rounds and the initial threshold adjustment coefficient. The initial threshold reduction coefficient is output. The initial threshold of the cross-round task switching ratio and the initial threshold reduction coefficient are combined to obtain the target cross-round task switching ratio threshold.

6. The method for automatically generating defense equipment intelligence reports based on a large language model as described in claim 1, characterized in that, The specific steps for performing multi-granularity content filling are as follows: Step 101: Obtain the actual filling step length during the first filling process. The actual filling step length is the single sliding span of the filling unit on the granular content carrier. Based on the current filling step length, establish a mapping relationship between the filling step length and the hierarchical division threshold adjustment amount, and dynamically adjust the granular hierarchical division threshold. Step 102: If the current fill step size is less than or equal to the lower bound of the step size reference, it indicates that the corresponding fill granularity level division threshold is set too low and the fill granularity is too fine, which has caused the fill efficiency to decrease and there is a lot of invalid redundant fill. It cannot adapt to the actual fill accuracy requirements of the granular content. Input the current fill step size into the fill step size-level division threshold adjustment amount mapping relationship, output the level division threshold gain amount, and superimpose the initial threshold of fill granularity level division with the level division threshold gain amount to obtain the target fill granularity level division threshold. Step 103: If the current fill step size is greater than or equal to the upper limit of the step size reference, the corresponding fill granularity level division threshold is set too high, the fill granularity is too coarse, and the fill information matching degree is seriously insufficient, which cannot meet the minimum fill accuracy requirements of the granular content. Input the current fill step size into the fill step size-level division threshold adjustment mapping relationship, output the level division threshold reduction amount, and superimpose the initial threshold of fill granularity level division with the level division threshold reduction amount to obtain the target fill granularity level division threshold.

7. The method for automatically generating defense equipment intelligence reports based on a large language model as described in claim 6, characterized in that, The process of performing multi-granularity content filling also includes: Step 104: If the current filling step size is within the step size reference interval, it indicates that the current filling step size is adapted to the basic filling requirements. Maintain the initial threshold of the filling granularity level division. The step size reference interval represents the open interval formed by the lower bound of the step size reference and the upper bound of the step size reference. Step 105: Based on the target filling granularity level division threshold and the current filling step size, execute the next round of granularity content iteration filling, repeating steps 101-104 to ensure the continuity and rationality of the thresholds at each level.

8. The method for automatically generating defense equipment intelligence reports based on a large language model as described in claim 1, characterized in that, The specific steps for the data consistency verification are as follows: The target incremental data set for the first round of incremental updates is collected, the timestamp of each data is extracted, the first round time window interval is divided based on the baseline time window as the period, the baseline time window is divided into sub-time windows on an average basis based on the initial incremental update batch, and the actual effective interval of each sub-time window is determined based on the alignment tolerance of the initial time window. The first round time window interval represents the closed interval formed by the start timestamp of the first round update and the end timestamp of the first round update. The target incremental data is assigned to the corresponding sub-time window of the actual effective interval according to the timestamp to complete the first round of batch division and processing, and the batch processing efficiency during the first round of processing is recorded. A batch processing efficiency-alignment tolerance adjustment factor mapping table is established, wherein the batch processing efficiency represents the average efficiency of processing a single batch of data. If the batch processing efficiency is less than the lower limit of the processing efficiency benchmark, it indicates that the amount of incremental data to be processed in a single batch is too large and cannot meet the timeliness requirements of incremental updates of defense equipment intelligence. In this case, the current batch processing efficiency is input into the batch processing efficiency-alignment tolerance adjustment factor mapping table, and the alignment tolerance reduction factor is output. The alignment tolerance reduction factor is then interactively processed with the initial time window alignment tolerance to obtain the target time window alignment tolerance, so as to reduce the amount of data processed in a single batch and improve the overall batch processing efficiency.

9. The method for automatically generating defense equipment intelligence reports based on a large language model as described in claim 8, characterized in that, The data consistency verification also includes: If the batch processing efficiency is within the processing efficiency benchmark range, it indicates that the current time window alignment tolerance and the matching degree of the incremental update batch are suitable for the processing requirements of incremental updates of defense equipment intelligence. Therefore, the current time window alignment tolerance is maintained to ensure the stability of the processing efficiency. The processing efficiency benchmark range refers to the closed interval formed by the lower limit of the processing efficiency benchmark and the upper limit of the processing efficiency benchmark. If the batch processing efficiency is greater than the upper limit of the processing efficiency benchmark, it indicates that the amount of incremental data to be processed in a single batch is too small and the redundancy overhead of batch processing is too high. In this case, the current batch processing efficiency is input into the batch processing efficiency-alignment tolerance adjustment factor mapping table, and the alignment tolerance adjustment factor is output. The alignment tolerance adjustment factor is then interacted with the initial time window alignment tolerance to obtain the target time window alignment tolerance, so as to improve the amount of data processed in a single batch, optimize the utilization rate and ensure the integrity of the data.

10. The method for automatically generating defense equipment intelligence reports based on a large language model as described in claim 1, characterized in that, The specific steps for determining whether the report's full lifecycle source tracing optimization has been triggered are as follows: A corresponding accuracy deviation threshold is preset for each granularity level. If the multi-granularity accuracy deviation of a level is greater than the accuracy deviation threshold, it is recorded as an over-threshold deviation for that level. During the continuous report generation process, the cumulative percentage of over-threshold deviation for each granularity level is calculated to characterize the stability and reliability of the overall output quality. If the cumulative percentage of deviations exceeding the threshold is less than or equal to the cumulative percentage benchmark value, then the full lifecycle source tracing optimization of the report will not be triggered. If the cumulative percentage of deviations exceeding the threshold is greater than the cumulative percentage benchmark value, the current cumulative percentage of deviations exceeding the threshold is input into the mapping relationship through the established cumulative percentage-lifetime threshold mapping relationship, and the target report generation lifetime threshold is output to reduce the report generation lifetime and restore the stability and reliability of the output.

Images