A method and system for generating an underwriting risk assessment report

Abstract

The application discloses a kind of underwriting risk assessment report generation method and system, comprising: the text data of the insured person, insurance vehicle video stream and external environment data are acquired in real time, and the physical defect feature set of each vehicle video stream is extracted;Text features, the physical defect feature set and initial environmental risk factor score are fused through attention mechanism layer across modalities, generate fusion feature vector, and the target surface damage quantitative value, target component rust area proportion and target environmental risk factor score obtained by decoding the fusion feature vector;Through the target large language model of pre-setting, in autoregressive manner item by item generate structure integrity evaluation item, mechanical failure probability prediction item and environmental damage risk trend item, until all evaluation items are generated, obtain the underwriting risk assessment report.The present application can dynamically capture the physical state change of insurance object, and then improve the timeliness and explainability of risk assessment.

Description

Technical Field

[0001] This invention relates to the field of insurance business, and in particular to a method and system for generating underwriting risk assessment reports. Background Technology

[0002] In the insurance business, changes in a vehicle's physical condition are crucial for accurate risk assessment. A vehicle's physical condition encompasses various factors, including the integrity of its body structure, the wear and tear of its components, and its mileage. Dynamically capturing these changes allows for timely understanding of the vehicle's actual condition under different times and usage scenarios. For example, after a collision, the vehicle's structure may change, and components may be damaged, all of which affect the vehicle's safety and future usability. By dynamically capturing these changes and generating underwriting reports, insurance companies can more accurately assess a vehicle's risk level, reasonably determine insurance rates, and ensure the sustainable development of their insurance business.

[0003] Currently, there are several technical means to understand changes in a vehicle's physical condition. One is through regular vehicle inspections. For example, insurance companies require car owners to regularly take their vehicles to designated repair shops or inspection stations where professionals inspect the vehicle's exterior, chassis, engine, and other components, recording relevant data. Another method is using the vehicle's built-in sensors. Some vehicles are equipped with sensors such as tire pressure sensors and collision sensors. These sensors can send signals when the vehicle experiences specific conditions, informing the insurance company of changes in its physical condition. However, regular inspections often have long intervals, meaning changes in the vehicle's physical condition between inspections may not be detected in time. This prevents insurance companies from obtaining the most accurate and up-to-date information on the vehicle's physical condition. Furthermore, the vehicle's built-in sensors have limited monitoring range, only monitoring specific physical conditions and unable to comprehensively understand overall changes in the vehicle's physical condition, affecting the timeliness of risk assessment and ultimately hindering insurance companies from making accurate and rapid underwriting decisions. Summary of the Invention

[0004] This invention provides a method and system for generating underwriting risk assessment reports, which can dynamically capture changes in the physical state of the insured object, thereby improving the timeliness and interpretability of risk assessment.

[0005] An embodiment of the present invention provides a method for generating an underwriting risk assessment report, comprising:

[0006] The system acquires the insured's text data, the insured vehicle's video stream, and external environment data in real time. It performs time-series segmentation on the insured vehicle's video stream to obtain several video segments and extracts the physical defect feature set of each video segment. The physical defect feature set includes the initial surface damage quantification value and the initial component corrosion area percentage.

[0007] The text features, the physical defect feature set, and the initial environmental risk factor score are fused across modally through an attention mechanism layer to generate a fused feature vector. The fused feature vector is then decoded to obtain the target surface damage quantification value, the target component corrosion area ratio, and the target environmental risk factor score. The text features are obtained by feature extraction from the text data, and the initial environmental risk factor score is obtained based on the analysis of the external environmental data.

[0008] The target surface damage quantification value, the target component corrosion area ratio, and the target environmental risk factor score are input into a preset target large language model. The model generates structural integrity assessment items, mechanical failure probability prediction items, and environmental damage risk trend items in an autoregressive manner until all assessment items are generated. The resulting report elements are then combined according to the template order to obtain the underwriting risk assessment report.

[0009] This application's embodiments, through time-series segmentation processing, can promptly capture changes in the vehicle's physical state across different time periods, thereby extracting physical defect feature sets for each time period. This facilitates subsequent risk assessments based on rich and detailed dynamic data. Cross-modal fusion fully considers the combined impact of vehicle physical state changes, related textual information, and environmental information on risk, eliminating semantic barriers between multi-source data and achieving effective integration of multi-source information. This allows for a more comprehensive assessment of underwriting risks, laying the foundation for improved interpretability of risk assessments. Decoding into numerical values ​​enables risk assessors to intuitively understand the vehicle's risk status in different aspects, thus improving the interpretability of the risk assessment results. Autoregressive generation and item-by-item refinement allow for the gradual improvement and refinement of each part of the assessment report based on the generated content and core assessment parameters, making the report content more complete and accurate in reflecting the vehicle's dynamic risk status. Combining according to template order makes the report content clear and organized, facilitating rapid understanding and use of the assessment results by risk assessors and relevant decision-makers, thereby enhancing the interpretability and practicality of the risk assessment report. Compared with existing technologies, this application can improve the timeliness and interpretability of risk assessment.

[0010] Furthermore, the step of performing time-series segmentation on the video stream of the insured vehicle to obtain several video segments is as follows:

[0011] The pixel gradient change in consecutive frames of the video stream of the insured vehicle is calculated using a three-frame differential gradient algorithm.

[0012] When the pixel gradient change exceeds a first preset threshold and the overlap between the change area and the pre-labeled vehicle component mask is greater than a second preset threshold, it is marked as a target keyframe representing a sudden change in physical state.

[0013] A dynamic time window is established with adjacent target keyframes as boundaries, and the segmentation unit is the keyframe from the start to the end of the window to generate a video segment. The length of the time window is negatively correlated with the amount of pixel gradient change.

[0014] By using time-series segmentation, we can capture the changes in the physical state of the vehicle at different time periods in a timely manner, thereby enabling the extraction of physical defect feature sets for each time period. This facilitates subsequent risk assessment based on rich and detailed dynamic data.

[0015] Furthermore, the extraction of the physical defect feature set of each video segment specifically involves:

[0016] Extract the initial local texture feature vector of each frame in the video segment, and align the same local region of consecutive frames in the same video segment to obtain the local texture feature vector;

[0017] For the local texture feature vectors of the same scale within the same video segment, max pooling is performed along the time dimension to generate aggregated feature vectors of the corresponding scale, and the aggregated feature vectors of all scales are concatenated to form a global aggregated feature matrix.

[0018] The global aggregated feature matrix is ​​used as the initial feature input of the nodes to the pre-trained graph convolutional defect recognition model. After iteratively updating the node features, several component defect probability vectors are obtained, and the physical defect feature set is determined based on each component defect probability vector.

[0019] Extracting physical defect feature sets in this way facilitates subsequent risk assessments based on rich and detailed dynamic data.

[0020] Furthermore, the process of fusing text features, the physical defect feature set, and the initial environmental risk factor score across modalities through an attention mechanism layer to generate a fused feature vector is as follows:

[0021] The text features are then subjected to word embedding encoding to obtain a text feature vector;

[0022] Normalize each defect feature in the physical defect feature set to obtain normalized feature values, and then concatenate the normalized feature values ​​to obtain a physical feature vector.

[0023] The initial environmental risk factor scores are numerically vectorized and encoded to obtain environmental feature vectors;

[0024] The text feature vector, the physical feature vector, and the environmental feature vector are input into a preset attention mechanism layer to calculate the attention weights among the three modalities, and then weighted and fused based on the attention weights to obtain the fused feature vector.

[0025] By performing cross-modal fusion, the combined impact of changes in vehicle physical state, relevant textual information, and environmental information on risk can be fully considered, eliminating semantic barriers between multi-source data and achieving effective integration of multi-source information. This lays the foundation for improving the interpretability of risk assessment in the future.

[0026] Furthermore, the step of inputting the target surface damage quantification value, the target component corrosion area ratio, and the target environmental risk factor score into a preset target large language model, and generating structural integrity assessment items, mechanical failure probability prediction items, and environmental damage risk trend items item by item in an autoregressive manner, specifically involves:

[0027] The target surface damage quantification value, the target component corrosion area ratio, and the target environmental risk factor score are input into a preset target large language model;

[0028] Based on the target surface damage quantification value, the risk level is calculated by calling the preset structural defect mapping rule library, and the large language model generates a structural integrity assessment item containing structural damage description text and risk level according to the first prompt word template;

[0029] The probability of mechanical failure is determined based on the risk level and the proportion of rusted area of ​​the target component. Combined with the probability of mechanical failure, the large language model generates a mechanical failure probability prediction item based on the second prompt word template.

[0030] Based on the mechanical failure probability and the target environmental risk factor score, the environmental risk value and risk trend curve are determined by a dynamic weighting formula. Combining the target environmental risk factor score and the risk trend curve, the large language model generates an environmental damage risk trend item based on the third prompt word template.

[0031] By generating and refining the data through autoregression, the various parts of the assessment report can be gradually improved and refined based on the generated content and core assessment parameters, making the report more complete and accurate in reflecting the dynamic risk status of the vehicle.

[0032] Furthermore, the preset target large language model is a pre-trained language model based on the Transformer architecture, including GPT or T5, and the target large language model is obtained after fine-tuning and training with insurance industry knowledge.

[0033] Furthermore, the training process of the target large language model is as follows:

[0034] Obtain insurance report samples and manually annotate the insurance report samples to obtain annotation results. Use the annotation results to conduct preliminary training on the large language model through supervised learning to obtain an initial large language model.

[0035] The target surface damage quantification value, the target component corrosion area ratio, and the target environmental risk factor score are input into the initial large language model, and a simulation report is output.

[0036] The reward value is calculated based on the simulation report. Based on the reward value, the parameters of the initial large language model are updated using the PPO policy gradient algorithm. When the iteration termination condition is met, the target large language model is obtained.

[0037] Furthermore, the text data includes credit data, historical insurance data, and historical claims data, and the text features are the results obtained after feature extraction from the text data, specifically:

[0038] The text data is acquired, and the text data is cleaned and preprocessed using natural language processing technology to obtain the preprocessing result;

[0039] The preprocessing result is vectorized to obtain a text vector, and text features are extracted from the text vector.

[0040] Another embodiment of the present invention provides an underwriting risk assessment report generation system, including: an acquisition module, a fusion module, and a generation module;

[0041] The acquisition module is used to acquire the insured's text data, the insured vehicle's video stream, and external environment data in real time, perform time-series segmentation processing on the insured vehicle's video stream to obtain several video segments, and extract the physical defect feature set of each video segment, wherein the physical defect feature set includes the initial surface damage quantification value and the initial component corrosion area percentage.

[0042] The fusion module is used to perform cross-modal fusion of text features, the physical defect feature set, and the initial environmental risk factor score through an attention mechanism layer to generate a fused feature vector, and to decode the fused feature vector to obtain the target surface damage quantification value, the target component corrosion area ratio, and the target environmental risk factor score. The text features are obtained by feature extraction from the text data, and the initial environmental risk factor score is obtained based on the analysis of the external environmental data.

[0043] The generation module is used to input the target surface damage quantification value, the target component corrosion area ratio and the target environmental risk factor score into a preset target large language model, and generate structural integrity assessment items, mechanical failure probability prediction items and environmental damage risk trend items in an autoregressive manner until all assessment items are generated. The obtained report elements are combined according to the template order to obtain the underwriting risk assessment report.

[0044] This application's embodiments, through time-series segmentation processing, can promptly capture changes in the vehicle's physical state across different time periods, thereby extracting physical defect feature sets for each time period. This facilitates subsequent risk assessments based on rich and detailed dynamic data. Cross-modal fusion fully considers the combined impact of vehicle physical state changes, related textual information, and environmental information on risk, eliminating semantic barriers between multi-source data and achieving effective integration of multi-source information. This allows for a more comprehensive assessment of underwriting risks, laying the foundation for improved interpretability of risk assessments. Decoding into numerical values ​​enables risk assessors to intuitively understand the vehicle's risk status in different aspects, thus improving the interpretability of the risk assessment results. Autoregressive generation and item-by-item refinement allow for the gradual improvement and refinement of each part of the assessment report based on the generated content and core assessment parameters, making the report content more complete and accurate in reflecting the vehicle's dynamic risk status. Combining according to template order makes the report content clear and organized, facilitating rapid understanding and use of the assessment results by risk assessors and relevant decision-makers, thereby enhancing the interpretability and practicality of the risk assessment report. Compared with existing technologies, this application can improve the timeliness and interpretability of risk assessment.

[0045] Furthermore, the generation module includes an input unit, a first generation unit, a second generation unit, and a third generation unit:

[0046] The input unit is used to input the target surface damage quantification value, the target component corrosion area ratio and the target environmental risk factor score into a preset target large language model;

[0047] The first generation unit is used to combine the target surface damage quantification value, call the preset structural defect mapping rule library to calculate the risk level, and the large language model generates a structural integrity assessment item containing structural damage description text and risk level according to the first prompt word template.

[0048] The second generation unit is used to determine the mechanical failure probability based on the risk level and the proportion of rust area of ​​the target component, and, in combination with the mechanical failure probability, the large language model generates a mechanical failure probability prediction item based on the second prompt word template.

[0049] The third generation unit is used to determine the environmental risk value and risk trend curve based on the mechanical failure probability and the target environmental risk factor score through a dynamic weighting formula, and to generate an environmental damage risk trend item by the large language model based on the third prompt word template, in combination with the target environmental risk factor score and the risk trend curve. Attached Figure Description

[0050] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0051] Figure 1 This is a flowchart illustrating one embodiment of the method for generating an underwriting risk assessment report provided in this application;

[0052] Figure 2 This is a flowchart illustrating steps S201 to S203 provided in this application;

[0053] Figure 3 This is a flowchart illustrating steps S301 to S304 provided in this application;

[0054] Figure 4 This is a flowchart illustrating steps S401 to S404 provided in this application;

[0055] Figure 5 This is a schematic diagram of an embodiment of the underwriting risk assessment report generation system provided in this application. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0057] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0058] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0059] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0060] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0061] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0062] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0063] In the insurance business, changes in the physical condition of a vehicle are crucial for risk assessment, encompassing factors such as the integrity of the vehicle's structure, the wear and tear of parts, and mileage. Dynamically monitoring these changes provides insurance companies with real-time information about the vehicle's condition at different times and under different usage scenarios, aiding in accurate risk assessment and the reasonable determination of insurance premiums, thus ensuring the stable development of the insurance business. Currently, monitoring the physical condition of vehicles mainly relies on periodic inspections and onboard sensors, but neither method can comprehensively and promptly reflect changes in the overall physical condition of the vehicle, hindering insurance companies from making accurate and rapid underwriting decisions.

[0064] See Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the underwriting risk assessment report generation method provided in this application. In order to dynamically capture changes in the physical state of the insured object, an embodiment of the present invention provides an underwriting risk assessment report generation method, including steps S101 to S103.

[0065] Step S101: Real-time acquisition of the insured's text data, the insured vehicle's video stream, and external environment data; time-series segmentation processing of the insured vehicle's video stream to obtain several video segments; and extraction of the physical defect feature set of each video segment, wherein the physical defect feature set includes the initial surface damage quantification value and the initial component corrosion area percentage.

[0066] In some embodiments, the insured's text data is obtained in real time, specifically by calling the insurance company's database in real time through an API interface to obtain the insured's relevant text data, wherein the text data may include, but is not limited to, credit data, historical insurance data, and historical claims data.

[0067] It should be noted that credit data refers to a collection of information that reflects the credit status of the insured, including basic information such as the insured's name, gender, age, ID number, and contact information, as well as credit score and credit record details; historical insurance data refers to various information and records submitted by the insured when insuring the same vehicle with the insurance company in the past, including but not limited to insurance period, type of insurance, insured amount, and vehicle information; historical claims data refers to a series of records generated when the insured vehicle was involved in an insured accident during the past insurance period and filed a claim with the insurance company, including but not limited to the claim time, reason for claim, and claim processing result.

[0068] In some embodiments, real-time acquisition of the insured's insured vehicle video stream specifically involves capturing real-time video streams of the insured vehicle, including the vehicle's driving environment, road conditions, and the vehicle's operating status, using in-vehicle cameras, mobile device cameras, or traffic monitoring cameras installed on the insured vehicle.

[0069] In some embodiments, real-time acquisition of vehicle external environment data specifically involves: using environmental monitoring equipment or integrating meteorological and geographic information data to acquire vehicle external environment data, wherein the external environment data may include, but is not limited to, data such as temperature, humidity, rainfall, wind speed, and water quality pH.

[0070] Please refer to Figure 2 In some embodiments, the step of performing time-series segmentation processing on the video stream of the insured vehicle to obtain several video segments includes steps S201 to S203.

[0071] Step S201: Calculate the pixel gradient change of consecutive frames in the video stream of the insured vehicle using a three-frame differential gradient algorithm;

[0072] In some embodiments, the input video stream is decoded into a continuous frame sequence, and each frame is grayscaled and denoised using a Gaussian filter function to obtain a preprocessed video. Three consecutive frames are selected from the preprocessed video in sequence, denoted as frame t-1, frame t, and frame t+1. The difference image G1 between frame t and frame t-1, and the difference image G2 between frame t+1 and frame t are calculated respectively. The gradient of the difference images G1 and G2 is calculated using a gradient operator to obtain the pixel gradient change.

[0073] Step S202: When the pixel gradient change exceeds a first preset threshold and the overlap between the change area and the pre-labeled vehicle component mask is greater than a second preset threshold, it is marked as a target keyframe representing a sudden change in physical state.

[0074] In some embodiments, a first preset threshold T1 (e.g., 25-30) and a second preset threshold T2 (e.g., >0.7) are preset. When the pixel gradient change exceeds the first preset threshold T1, i.e., the gradient change > T1, and the overlap between the change area and the pre-labeled vehicle component mask is greater than the second preset threshold T2, i.e., the overlap > T2, the current frame is marked as the target keyframe.

[0075] It's important to note that when using the three-frame differential gradient algorithm to calculate pixel gradient changes, a change region is obtained. This region shows a significant change in pixel intensity, typically indicating a change in the appearance of that area in the video stream. Examples include new scratches or dents on a car's surface, or displacement of components. The pre-annotated vehicle component mask is a predefined mask used to identify the positions of different vehicle components in the image. For example, the positions of the hood, doors, and bumpers are marked. By calculating the overlap between the change region and the mask, it can be determined whether the change occurred on a vehicle component.

[0076] It should be noted that when the pixel gradient change exceeds a first preset threshold, it indicates a significant change in the appearance of the corresponding area in the image. This could be due to vehicle collision causing deformation, new damage to the vehicle surface, or other changes in physical state. However, relying solely on pixel gradient changes is inaccurate. Therefore, it is also necessary to determine if the changed area has a high degree of overlap with the vehicle component mask. If so, it can be confirmed that the vehicle's physical state has changed. For example, if the changed area is mainly in the swaying area of ​​trees in the background of the vehicle's movement, even if the pixel gradient change is large, it will not be marked as a keyframe because of its low overlap with the vehicle component mask. This avoids misjudging changes in the background or other non-vehicle components as changes in the vehicle's physical state. Therefore, only when both of the above conditions are met simultaneously is this frame marked as the target keyframe representing a sudden change in physical state.

[0077] Step S203: Establish a dynamic time window with adjacent target keyframes as boundaries, and use the window start keyframe to end keyframe as segmentation unit to generate video segments, wherein the length of the time window is negatively correlated with the amount of pixel gradient change.

[0078] In some embodiments, a dynamic time window is established using adjacent target keyframes as boundaries, and the length of the time window is determined according to the magnitude of pixel gradient changes and a pre-defined proportional or functional relationship. For example, if frames t1 and t2 are two adjacent target keyframes, a time window is established with t1 as the starting keyframe and t2 as the ending keyframe. Then, based on the determined time window length, video segments are extracted from the starting keyframe to the ending keyframe as segmentation units. Each video segment generated in this way contains a significant change in the vehicle's physical state, facilitating subsequent feature extraction and analysis of the video segments.

[0079] It should be noted that the time window length is negatively correlated with the amount of pixel gradient change. That is, the larger the pixel gradient change, the more drastic the change in the vehicle's physical state, and the shorter the time window length should be to capture the change more precisely; conversely, the window length can be appropriately extended.

[0080] By using time-series segmentation, we can capture the changes in the physical state of the vehicle at different time periods in a timely manner, thereby enabling the extraction of physical defect feature sets for each time period. This facilitates subsequent risk assessment based on rich and detailed dynamic data.

[0081] In some embodiments, extracting the physical defect feature set of each video segment specifically involves: extracting the initial local texture feature vector of each frame in the video segment, aligning the same local region of consecutive frames in the same video segment to obtain a local texture feature vector; performing max pooling along the time dimension on the local texture feature vectors of the same scale within the same video segment to generate aggregated feature vectors of the corresponding scale, and concatenating the aggregated feature vectors of all scales to form a global aggregated feature matrix; using the global aggregated feature matrix as the initial feature input of nodes to a pre-trained graph convolutional defect recognition model, obtaining several component defect probability vectors by iteratively updating the node features, and determining the physical defect feature set based on each component defect probability vector. Specifically, firstly, each frame in the video segment is preprocessed, including image normalization, resizing, etc., to ensure that the image data is suitable for input into the feature extraction model. Secondly, using convolutional neural networks (CNNs) or other feature extraction algorithms, such as SIFT and ORB, initial local texture feature vectors are extracted from each frame of the video segment. Taking CNNs as an example, a pre-trained model (such as ResNet or VGG) can be selected, its final fully connected layer removed, and the remaining part used as a feature extractor. Then, within the same video segment, alignment is performed on the same local region in consecutive frames through feature matching. Subsequently, for local texture feature vectors of the same scale within the same video segment, max pooling is performed along the time dimension. For each local texture feature, the maximum value is taken in the time series to generate an aggregated feature vector at the corresponding scale. All aggregated feature vectors at all scales are then concatenated to form a global aggregated feature matrix. Finally, the global aggregated feature matrix is ​​used as the initial feature input of the nodes to pre-train the graph convolutional defect recognition model. Each node represents a key region in the video segment. In the model, the relationship between nodes is captured by iteratively updating the node features. After multiple iterations, several component defect probability vectors are obtained. Based on the component defect probability vectors, the initial surface damage quantification value and the initial component corrosion area ratio are calculated as the physical defect feature set.

[0082] It should be noted that feature matching methods can be based on feature point detection algorithms such as SIFT and ORB to detect feature points in an image, and then determine the position of the same local region in different frames by calculating the matching relationship between feature points.

[0083] It should be noted that the calculation of the initial surface damage quantification value and the initial component corrosion area ratio based on the component defect probability vector can be achieved by setting a probability threshold. For example, if the probability value of a certain category in the defect probability vector of a certain component is greater than the set threshold, then the component is considered to have a corresponding physical defect. Based on these judgment results, the specific values ​​of the physical defect feature set, such as the vehicle surface damage quantification value and the corrosion area ratio of key components, are obtained. These values ​​will be used in the subsequent risk assessment process.

[0084] It should be noted that the surface damage quantification value (damage quantification value = defect area / total component area) is an indicator used to quantitatively describe the degree of damage to the vehicle surface. The value ranges from 0 to 1, where 0 indicates that the vehicle surface is completely undamaged and in good condition; 1 indicates that the vehicle surface is extremely damaged, possibly almost completely destroyed or covered with a large area of ​​severe damage. The rust area percentage reflects the proportion of the vehicle component's surface area that is rusted. 0% means that the component surface has no rust at all and is in a good, unrusted condition; 100% means that the component surface is completely covered by rust.

[0085] It should be noted that specific feature extraction is not the focus of this application, so it will not be discussed further here.

[0086] Extracting physical defect feature sets in this way facilitates subsequent risk assessments based on rich and detailed dynamic data.

[0087] Step S102: The text features, the physical defect feature set, and the initial environmental risk factor score are fused across modally through an attention mechanism layer to generate a fused feature vector. The fused feature vector is then decoded to obtain the target surface damage quantification value, the target component corrosion area ratio, and the target environmental risk factor score. The text features are obtained by feature extraction from the text data, and the initial environmental risk factor score is obtained based on the analysis of the external environmental data.

[0088] In some embodiments, the text data includes credit data, historical insurance data, and historical claims data. The text features are the results obtained after feature extraction from the text data. Specifically, the process involves: acquiring the text data and cleaning and preprocessing it using natural language processing (NLP) techniques to obtain a preprocessed result; vectorizing the preprocessed result to obtain a text vector, and extracting text features from the text vector. Specifically, firstly, after obtaining the insured's credit data, historical insurance data, historical claims data, and other text data, the text data needs to be denoised and standardized to obtain a text processing result; then, a word segmenter in a NLP tool is used to segment the text processing result, dividing the continuous text into independent words or phrases for easier subsequent text analysis, and removing stop words to obtain a preprocessed result; finally, the frequency of each word in the text is counted using a bag-of-words model, representing the text as a vector, where each element of the vector corresponds to the number of times a word appears or a TF-1 expression. The IDF value is used to generate bag-of-words vectors based on the occurrence of words in each text. A pre-trained word embedding model (such as Word2Vec, GloVe, etc.) is then trained on a specific corpus to map the bag-of-words vectors to a high-dimensional vector space, so that words with similar meanings are closer together in the vector space, thus obtaining text vectors. Finally, some simple statistical features are extracted from the text vectors, such as lexical richness (the number of different words), average word length, and text length, thus obtaining text features that can reflect the basic statistical characteristics of the text.

[0089] This text data processing method can extract key features from multi-dimensional text information, providing a basis for dynamically assessing the physical state of the insured object, enhancing the timeliness and interpretability of risk assessment, and thus helping to dynamically capture changes in the physical state of the insured object.

[0090] In some embodiments, the initial environmental risk factor score is obtained based on the analysis of the external environmental data. Specifically, the process involves: first, identifying environmental risk factors and preprocessing the external environmental data (cleaning and normalization); second, collecting historical environmental data and corresponding insurance risk data as training samples, dividing the dataset into a training set and a test set, training the model using the training set, adjusting parameters to optimize performance, and then using the test set to evaluate indicators such as accuracy, precision, and recall, iteratively optimizing to obtain the environmental risk factor scoring model; subsequently, inputting the external environmental data into the environmental risk factor scoring model to obtain the model output results, and determining the initial environmental risk factor score according to preset scoring rules, such as dividing the model output value into multiple intervals, each interval corresponding to a scoring level, with higher risk resulting in a higher score.

[0091] It should be noted that environmental risk factors can be determined based on the needs of insurance business and the characteristics of the insured object, such as meteorological conditions (temperature, humidity, rainfall, wind speed, etc.), air quality (concentrations of pollutants such as PM2.5, PM10, and sulfur dioxide), geographical factors (altitude, topography, distance from rivers, etc.), water quality (pH, dissolved oxygen, chemical oxygen demand, etc.), and noise levels. Subsequently, by analyzing historical insurance data and industry research literature, the relationship between these environmental factors and the potential risks to the insured object is studied.

[0092] It should be noted that the environmental risk score is a rating index that measures the degree of risk impact that the environment in which a vehicle is located may have on it. The score ranges from 0 to 10. A score of 0 indicates that the environment poses virtually no risk to the vehicle and is a very safe environment. A score of 10 indicates that there are extremely high risk factors in the environment, which are very likely to cause damage to the vehicle.

[0093] Please refer to Figure 3 The step of fusing text features, the physical defect feature set, and the initial environmental risk factor score across modalities through an attention mechanism layer to generate a fused feature vector includes steps S301 to S304.

[0094] Step S301: Perform word embedding encoding on the text features to obtain a text feature vector;

[0095] In some embodiments, the text features are first divided into words and special markers (such as start and end markers) are added. Then, a pre-trained language model (such as BERT or XLNet) is used to map the word segments into high-dimensional semantic vectors to obtain text feature vectors, where each word is represented as a fixed-length vector (such as 768-dimensional) to capture contextual semantics.

[0096] Step S302: Normalize each defect feature in the physical defect feature set to obtain normalized feature values, and concatenate the normalized feature values ​​to obtain a physical feature vector.

[0097] In some embodiments, each defect feature (such as the surface damage quantification value as a percentage of the component's rust area) is normalized using methods such as max-min normalization or Z-score normalization. This transforms feature values ​​with different dimensions and ranges into a unified scale (such as [0, 1] or [-1, 1]) to obtain normalized feature values. Then, the normalized feature values ​​are arranged in a specific order and concatenated into a complete vector, namely the physical feature vector (such as [vehicle surface damage quantification value, component rust area percentage, environmental risk factor score]). This vector comprehensively and compactly represents the physical defect characteristics of the insured vehicle.

[0098] Step S303: The initial environmental risk factor score is numerically vectorized and encoded to obtain an environmental feature vector;

[0099] In some embodiments, the initial environmental risk factor score is converted into a numerical vector, and normalization or other numerical encoding methods are used to ensure that it has the same scale and format as other feature vectors, thus obtaining an environmental feature vector.

[0100] Step S304: Input the text feature vector, the physical feature vector, and the environment feature vector into a preset attention mechanism layer to calculate the attention weights among the three modalities, and perform weighted fusion based on the attention weights to obtain the fused feature vector;

[0101] In some embodiments, firstly, an attention mechanism layer is designed and constructed, which can receive text feature vectors, physical feature vectors, and environmental feature vectors as inputs, and calculate the attention weights between them to obtain the attention weights among the three modalities; then, the text feature vectors, physical feature vectors, and environmental feature vectors are weighted using cross-attention weights. The resulting fusion vector not only contains the original information of text features, physical features, and environmental features, but also integrates the correlation information between the three, which can more comprehensively and deeply represent the comprehensive features of the insured vehicle.

[0102] It should be noted that the initial surface damage quantification value, the initial component corrosion area ratio, and the initial environmental risk factor score can be understood as values ​​initially determined based on relevant data, while the target surface damage quantification value, the target component corrosion area ratio, and the target environmental risk factor score are values ​​determined after fusion and adjustment.

[0103] By performing cross-modal fusion, the combined impact of changes in vehicle physical state, relevant textual information, and environmental information on risk can be fully considered, eliminating semantic barriers between multi-source data and achieving effective integration of multi-source information. This lays the foundation for improving the interpretability of risk assessment in the future.

[0104] In some embodiments, the fused feature vector is decoded to obtain the target surface damage quantification value, the target component corrosion area ratio, and the target environmental risk factor score. Specifically, the new fused feature vector is input into the trained decoding model, and the model outputs the corresponding target surface damage quantification value, target component corrosion area ratio, and target environmental risk factor score according to the learned mapping relationship.

[0105] It should be noted that the specific decoding process is not the focus of this application, so it will not be elaborated here.

[0106] Step S103: Input the target surface damage quantification value, the target component corrosion area ratio and the target environmental risk factor score into the preset target large language model, and generate structural integrity assessment items, mechanical failure probability prediction items and environmental damage risk trend items in an autoregressive manner until all assessment items are generated. Combine the obtained report elements according to the template order to obtain the underwriting risk assessment report.

[0107] Please refer to Figure 4 In some embodiments, the step of inputting the target surface damage quantification value, the target component corrosion area ratio and the target environmental risk factor score into a preset target large language model to generate structural integrity assessment items, mechanical failure probability prediction items and environmental damage risk trend items in an autoregressive manner includes steps S401 to S404.

[0108] Step S401: Input the target surface damage quantification value, the target component corrosion area ratio, and the target environmental risk factor score into the preset target large language model;

[0109] Step S402: Combining the target surface damage quantification value, the preset structural defect mapping rule library is called to calculate the risk level, and the large language model generates a structural integrity assessment item containing structural damage description text and risk level based on the first prompt word template;

[0110] In some embodiments, a preset structural defect mapping rule library is invoked to find the corresponding mapping rule in the rule library based on the target surface damage quantification value, calculate the risk level of vehicle structural integrity, and then use the first prompt word template to guide the large language model to generate a structural integrity assessment item containing structural damage description text and risk level.

[0111] It should be noted that the structural defect mapping rule base is generated by training an XGBoost model with historical vehicle damage data. The input is the target surface damage quantification value, and the output is the risk level (low / medium / high). For example, damage quantification value > 0.4 → high risk. The specific training process is not the focus of this application, so it will not be elaborated here.

[0112] It should be noted that the first prompt word template is: "Based on the quantitative value of vehicle surface damage (volume / surface area ratio = 0.35, confidence interval [0.32, 0.38], damage location: front bumper, left door), fill the specific values ​​in the quantitative value of target surface damage into the prompt word template to form a complete prompt word, input it into the large language model, and generate detailed and professional structural damage description text."

[0113] It should be noted that the structural defect mapping rule base establishes a mapping relationship between vehicle structural parameters and risk levels based on a large amount of historical data and expert experience.

[0114] Step S403: Determine the mechanical failure probability based on the risk level and the proportion of rust area of ​​the target component, and combine the mechanical failure probability with the large language model to generate a mechanical failure probability prediction item based on the second prompt word template;

[0115] In some embodiments, firstly, based on the vehicle's risk level and the percentage of rusted area of ​​the target component, the probability of mechanical failure of the vehicle is calculated using a preset failure probability model or formula. For example, the higher the risk level and the larger the percentage of rusted area, the higher the probability of mechanical failure. Then, a prediction text is generated using a second prompt word template; finally, the calculated failure probability (e.g., 35%) is combined with the generated prediction text to form a mechanical failure probability prediction item.

[0116] It should be noted that the second prompt template is: "Based on the current corrosion parameters (corrosion area of ​​engine block is 12.5%, corrosion severity is 2.3, growth rate is 3% / month) and risk level (high risk), predict the probability of mechanical failure in the next 6 months and provide preventive maintenance recommendations."

[0117] Step S404: Based on the mechanical failure probability and the target environmental risk factor score, the environmental risk value and risk trend curve are determined by a dynamic weighting formula. Combining the target environmental risk factor score and the risk trend curve, the large language model generates an environmental damage risk trend item based on the third prompt word template.

[0118] In some embodiments, firstly, considering factors in the mechanical failure probability and the target environmental risk factor score, such as the impact of environmental humidity on corrosion and the impact of air quality on vehicle component aging, an environmental risk assessment model is established to calculate the environmental risk value faced by the vehicle (environmental risk value = α × mechanical failure probability + β × environmental score, α + β = 1). Then, using a dynamic weighting formula, based on historical environmental parameter data and the current environmental risk value, the trend of environmental risk changes over a future period is predicted, generating a risk trend curve. Finally, based on a third prompt word template, such as "Based on the humidity index of the vehicle's environment being {C}, the air quality score being {D}, the environmental risk value being {E}, and the risk trend curve, describe the future trend of environmental damage risk to the vehicle," the specific values ​​in the target environmental risk factor score, the current environmental risk value, and the key features of the risk trend curve are filled into the prompt word template and input into the large language model to generate an environmental damage risk trend item, helping insurance companies understand the dynamic changes in vehicle risk under different environmental conditions.

[0119] It should be noted that the dynamic weighting formula will weight different environmental parameters based on factors such as time and the importance of the parameters, in order to more accurately reflect the changing trend of risk.

[0120] By generating and refining the data through autoregression, the various parts of the assessment report can be gradually improved and refined based on the generated content and core assessment parameters, making the report more complete and accurate in reflecting the dynamic risk status of the vehicle.

[0121] In some embodiments, the preset target large language model is a pre-trained language model based on the Transformer architecture, including GPT or T5, and the target large language model is obtained after fine-tuning training with insurance industry knowledge.

[0122] In some embodiments, the training process of the target large language model is as follows: Underwriting report samples are obtained and manually labeled to obtain labeling results. These labeling results are then used to perform preliminary training of the large language model through supervised learning to obtain an initial large language model. The target surface damage quantification value, the target component corrosion area ratio, and the target environmental risk factor score are input into the initial large language model to output a simulated report. A reward value is calculated based on the simulated report (reward value R = similarity weight × Sim(simulated report, underwriting report sample) + preset risk prediction weight × (1 - |predicted probability - actual probability|), where Sim is cosine similarity). Based on the reward value, the parameters of the initial large language model are updated using the PPO strategy gradient algorithm. When the iteration termination condition is met, the target large language model is obtained. Specifically, firstly, a large number of representative underwriting reports are collected as samples. These samples should cover various vehicle conditions, risk scenarios, and insurance assessment results. Professional insurance assessors and domain experts are organized to perform detailed manual labeling on these samples. Next, the labeled underwriting report samples are input into a pre-trained large language model, such as GPT, BERT, or T5 based on the Transformer architecture. Using supervised learning, the model parameters are adjusted based on the labeling results, enabling the model to learn the text patterns, semantic features, risk assessment knowledge, and text structure characteristics of the underwriting report. Through multiple rounds of iterative training, the model's performance is gradually optimized, resulting in an initial large language model with a certain ability to recognize and generate underwriting report features. Subsequently, the target surface damage quantification value, the target component corrosion area ratio, and the target environmental risk factor score are input into the initial large language model. Based on the knowledge and patterns learned in the initial training, the model generates corresponding simulated underwriting risk assessment reports. These simulated reports include assessments such as vehicle structural integrity assessment, mechanical failure probability prediction, and environmental damage risk trends. Then, based on the differences between the simulated report and the actual labeled underwriting report samples, as well as the accuracy and reasonableness of the various assessment results in the simulated report, corresponding reward values ​​are designed and calculated. Using the Proximal Policy Optimization (PPO) gradient algorithm, the parameters of the initial large language model are updated according to the calculated reward values. This process of generating simulated reports from input numerical risk description statements, calculating reward values, and updating model parameters is repeated for multiple rounds of iterative training. This continues until preset iteration termination conditions are met, such as reaching a certain number of iterations, the model performance no longer significantly improving on the validation set, or the reward values ​​stabilizing. Finally, the optimized target large language model is obtained.

[0123] It should be noted that the annotation content includes, but is not limited to, the vehicle's specific physical defects, risk level, the rationality of the assessment parameters, and the accuracy and standardization of the text description, forming annotation results that provide supervision signals for model training.

[0124] This application's embodiments, through time-series segmentation processing, can promptly capture changes in the vehicle's physical state across different time periods, thereby extracting physical defect feature sets for each time period. This facilitates subsequent risk assessments based on rich and detailed dynamic data. Cross-modal fusion fully considers the combined impact of vehicle physical state changes, related textual information, and environmental information on risk, eliminating semantic barriers between multi-source data and achieving effective integration of multi-source information. This allows for a more comprehensive assessment of underwriting risks, laying the foundation for improved interpretability of risk assessments. Decoding into numerical values ​​enables risk assessors to intuitively understand the vehicle's risk status in different aspects, thus improving the interpretability of the risk assessment results. Autoregressive generation and item-by-item refinement allow for the gradual improvement and refinement of each part of the assessment report based on the generated content and core assessment parameters, making the report content more complete and accurate in reflecting the vehicle's dynamic risk status. Combining according to template order makes the report content clear and organized, facilitating rapid understanding and use of the assessment results by risk assessors and relevant decision-makers, thereby enhancing the interpretability and practicality of the risk assessment report. Compared with existing technologies, this application can improve the timeliness and interpretability of risk assessment.

[0125] like Figure 5 As shown, based on the above method embodiments, corresponding apparatus embodiments are provided;

[0126] One embodiment of the present invention provides an underwriting risk assessment report generation system, including: an acquisition module 100, a fusion module 200, and a generation module 300;

[0127] The acquisition module 100 is used to acquire the insured's text data, the insured vehicle video stream, and external environment data in real time, perform time-series segmentation processing on the insured vehicle video stream to obtain several video segments, and extract the physical defect feature set of each video segment, wherein the physical defect feature set includes the initial surface damage quantification value and the initial component corrosion area ratio.

[0128] The fusion module 200 is used to perform cross-modal fusion of text features, the physical defect feature set, and the initial environmental risk factor score through an attention mechanism layer to generate a fused feature vector, and to decode the fused feature vector to obtain the target surface damage quantification value, the target component corrosion area ratio, and the target environmental risk factor score. The text features are obtained by feature extraction from the text data, and the initial environmental risk factor score is obtained based on the analysis of the external environmental data.

[0129] The generation module 300 is used to input the target surface damage quantification value, the target component corrosion area ratio and the target environmental risk factor score into a preset target large language model, and generate structural integrity assessment items, mechanical failure probability prediction items and environmental damage risk trend items in an autoregressive manner until all assessment items are generated. The obtained report elements are combined according to the template order to obtain an underwriting risk assessment report.

[0130] In some embodiments, the generation module 300 includes an input unit, a first generation unit, a second generation unit, and a third generation unit:

[0131] The input unit is used to input the target surface damage quantification value, the target component corrosion area ratio and the target environmental risk factor score into a preset target large language model;

[0132] The first generation unit is used to combine the target surface damage quantification value, call the preset structural defect mapping rule library to calculate the risk level, and the large language model generates a structural integrity assessment item containing structural damage description text and risk level according to the first prompt word template.

[0133] The second generation unit is used to determine the mechanical failure probability based on the risk level and the proportion of rust area of ​​the target component, and, in combination with the mechanical failure probability, the large language model generates a mechanical failure probability prediction item based on the second prompt word template.

[0134] The third generation unit is used to determine the environmental risk value and risk trend curve based on the mechanical failure probability and the target environmental risk factor score through a dynamic weighting formula, and to generate an environmental damage risk trend item by the large language model based on the third prompt word template, in combination with the target environmental risk factor score and the risk trend curve.

[0135] It is understood that the above-described device embodiments correspond to the method embodiments of the present invention, and can implement the method for generating an underwriting risk assessment report provided by any of the above-described method embodiments of the present invention.

[0136] It should be noted that the device embodiments described above are merely illustrative, and some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can specifically be implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0137] Based on the above embodiments of the underwriting risk assessment report generation method, another embodiment of the present invention provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the underwriting risk assessment report generation method of any embodiment of the present invention.

[0138] For example, in this embodiment, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the terminal device.

[0139] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0140] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0141] Based on the above-described method embodiments, another embodiment of the present invention provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the underwriting risk assessment report generation method described in any of the above-described method embodiments of the present invention.

[0142] The modules / units integrated in the device / terminal equipment, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0143] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for generating an underwriting risk assessment report, characterized in that, The method comprises the following steps: real-time acquisition of text data of an insured person, vehicle insurance video stream and external environment data, time sequence segmentation processing of the vehicle insurance video stream to obtain a plurality of video segments, and extraction of a physical defect feature set of each video segment, wherein the physical defect feature set comprises an initial surface damage quantitative value and an initial component rust area proportion; cross-modal fusion of text features, the physical defect feature set and an initial environmental risk factor score through an attention mechanism layer to generate a fusion feature vector, and decoding of the fusion feature vector to obtain a target surface damage quantitative value, a target component rust area proportion and a target environmental risk factor score, wherein the text features are obtained by feature extraction on the text data, and the initial environmental risk factor score is obtained based on analysis of the external environment data; input of the target surface damage quantitative value, the target component rust area proportion and the target environmental risk factor score into a preset target large language model to generate a structural integrity evaluation item, a mechanical failure probability prediction item and an environmental damage risk trend item in a self-recurrent manner until all evaluation items are generated, combination of obtained report elements according to a template order to obtain a risk assessment report. Wherein, the extraction of the physical defect feature set of each video segment is specifically: extracting an initial local texture feature vector of each frame of image in the video segment, and aligning the same local area of consecutive frames in the same video segment to obtain a local texture feature vector; performing a maximum pooling operation on the local texture feature vectors of the same scale within the same video segment along the time dimension to generate an aggregated feature vector of the corresponding scale, and splicing the aggregated feature vectors of all scales to form a global aggregated feature matrix; inputting the global aggregated feature matrix as node initial features into a pre-trained graph convolution defect recognition model to obtain a plurality of component defect probability vectors after iteratively updating node features, and determining the physical defect feature set based on the component defect probability vectors.

2. The underwriting risk assessment report generation method of claim 1, wherein, The time sequence segmentation processing of the vehicle insurance video stream to obtain a plurality of video segments is specifically: calculating the pixel gradient change amount of consecutive frames in the vehicle insurance video stream through a three-frame difference gradient algorithm; when the pixel gradient change amount exceeds a first preset threshold and the change area has a coincidence degree with a pre-labeled vehicle component mask greater than a second preset threshold, marking as a target key frame representing a physical state mutation; establishing a dynamic time window with adjacent target key frames as boundaries, and taking the window start key frame to the end key frame as a segmentation unit to generate a video segment, wherein the time window length is negatively correlated with the pixel gradient change amount.

3. The underwriting risk assessment report generation method of claim 1, wherein, The cross-modal fusion of text features, the physical defect feature set and an initial environmental risk factor score through an attention mechanism layer to generate a fusion feature vector is specifically: performing word embedding coding on the text features to obtain a text feature vector; The physical defect features in the physical defect feature set are normalized to obtain normalized feature values, and the normalized feature values are spliced to obtain a physical feature vector; The initial environmental risk factor score is numerically vectorized to obtain an environmental feature vector; The text feature vector, the physical feature vector, and the environmental feature vector are input into a preset attention mechanism layer to calculate attention weights between three modalities, and the fusion feature vector is obtained by weighted fusion based on the attention weights.

4. The underwriting risk assessment report generation method of claim 1, wherein, The target surface damage quantification value, the target component rust area proportion, and the target environmental risk factor score are input into a preset target large language model to generate a structural integrity evaluation item, a mechanical failure probability prediction item, and an environmental damage risk trend item in a self-recurrent manner, specifically as follows: The target surface damage quantification value, the target component rust area proportion, and the target environmental risk factor score are input into a preset target large language model; The target surface damage quantification value is combined to call a preset structural defect mapping rule library to calculate a risk level, and the large language model generates a structural integrity evaluation item containing a structural damage description text and a risk level according to a first prompt word template; Based on the risk level and the target component rust area proportion, the mechanical failure probability is determined, and the large language model generates a mechanical failure probability prediction item according to a second prompt word template in combination with the mechanical failure probability; Based on the mechanical failure probability and the target environmental risk factor score, an environmental risk value and a risk trend curve are determined through a dynamic weighting formula, and the large language model generates an environmental damage risk trend item according to a third prompt word template in combination with the target environmental risk factor score and the risk trend curve.

5. The underwriting risk assessment report generation method of claim 1, wherein, The preset target large language model is a pre-training language model based on the Transformer architecture, including GPT or T5, and the target large language model is obtained after fine-tuning training with insurance industry knowledge.

6. The underwriting risk assessment report generation method of claim 1, wherein, The training process of the target large language model is as follows: Obtain an underwriting report sample, artificially label the underwriting report sample to obtain a labeled result, and use the labeled result to preliminarily train a large language model through supervised learning to obtain an initial large language model; The target surface damage quantification value, the target component rust area proportion, and the target environmental risk factor score are input into the initial large language model to output a simulation report; Based on the simulation report, a reward value is calculated, and based on the reward value, the parameters of the initial large language model are updated through a PPO policy gradient algorithm to obtain a target large language model when the iteration termination condition is met.

7. The underwriting risk assessment report generation method of claim 1, wherein, The text data includes credit investigation data, historical insurance data, and historical claim data, and the text features are obtained by feature extraction on the text data, specifically as follows: The text data is obtained, and the text data is cleaned and preprocessed through natural language processing technology to obtain a preprocessing result; The preprocessing result is vectorized to obtain a text vector, and text features are extracted from the text vector.

8. A risk underwriting assessment report generation system, characterized by, Comprise: The acquisition module, the fusion module and the generation module; The acquisition module is used for acquiring text data of an insured person, a vehicle video stream of an insurance vehicle and external environment data in real time, performing time sequence segmentation processing on the vehicle video stream of the insurance vehicle to obtain a plurality of video segments, and extracting a physical defect feature set of each video segment, wherein the physical defect feature set comprises an initial surface damage quantitative value and an initial component rust area proportion; Wherein, the extraction of the physical defect feature set of each video segment is specifically: extracting an initial local texture feature vector of each frame of image in the video segment, and aligning the same local area of consecutive frames in the same video segment to obtain a local texture feature vector; performing a maximum pooling operation on the local texture feature vectors of the same scale within the same video segment along the time dimension to generate an aggregated feature vector of the corresponding scale, and concatenating the aggregated feature vectors of all scales to form a global aggregated feature matrix; inputting the global aggregated feature matrix as a node initial feature into a pre-trained graph convolution defect recognition model, obtaining a plurality of component defect probability vectors after iteratively updating node features, and determining the physical defect feature set based on the component defect probability vectors; The fusion module is used for performing cross-modal fusion on text features, the physical defect feature set and an initial environment risk factor score through an attention mechanism layer to generate a fusion feature vector, and decoding the fusion feature vector to obtain a target surface damage quantitative value, a target component rust area proportion and a target environment risk factor score, wherein the text features are obtained by feature extraction on the text data, and the initial environment risk factor score is obtained based on analysis of the external environment data. The generation module is used for inputting the target surface damage quantitative value, the target component rust area proportion and the target environment risk factor score into a pre-set target large language model to generate a structure integrity evaluation item, a mechanical failure probability prediction item and an environmental damage risk trend item in a self-recurrent manner, until all evaluation items are generated, combining the obtained report elements according to the template order to obtain a risk assessment report.

9. The underwriting risk assessment report generation system of claim 8, wherein, The generation module comprises an input unit, a first generation unit, a second generation unit and a third generation unit: The input unit is used for inputting the target surface damage quantitative value, the target component rust area proportion and the target environment risk factor score into a pre-set target large language model; The first generation unit is used for combining the target surface damage quantitative value, calling a pre-set structure defect mapping rule library to calculate a risk level, and generating a structure integrity evaluation item comprising a structure damage description text and a risk level by the large language model according to a first prompt word template; The second generation unit is configured to determine a mechanical failure probability based on the risk level and the target component corrosion area proportion, and generate, by the large language model, a mechanical failure probability prediction item according to a second prompt word template in combination with the mechanical failure probability. The third generation unit is configured to determine an environmental risk value and a risk trend curve based on the mechanical failure probability and the target environmental risk factor score by a dynamic weighting formula, and generate, by the large language model, an environmental damage risk trend item according to a third prompt word template in combination with the target environmental risk factor score and the risk trend curve.

Images