Real-time risk early warning method and system for insurance business under dynamic risk assessment

Abstract

The application provides an insurance business real-time risk early warning method and system under dynamic risk assessment, comprising: acquiring multi-modal risk data containing time, space and industry dimensions in the insurance business; separating a core risk element area from the multi-modal risk data through a mask mechanism to generate an initial risk feature map with a mask label; inputting the initial risk feature map into a dynamic risk assessment network of an encoder-decoder architecture; wherein the encoder and the decoder both contain multiple levels of cascaded risk processing modules, each risk processing module integrates a gating cycle unit and is configured with a binary risk gate for dynamically controlling the update path of the risk features; and generating a real-time early warning signal based on the target risk features output by the decoder and in combination with a preset risk level threshold. In the application, the processing path of the risk features can be dynamically adjusted, thereby improving the timeliness and accuracy of the risk early warning.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for real-time risk early warning of insurance business under dynamic risk assessment. Background Technology

[0002] In modern insurance operations, risk assessment and early warning are crucial for ensuring the industry's stable operation. As the scale of insurance business continues to expand and business scenarios become increasingly complex, traditional risk assessment methods are gradually revealing numerous problems. Currently, most insurance business risk assessments adopt a static assessment model, relying solely on historical data and a fixed risk indicator system for risk judgment.

[0003] On the one hand, it is difficult to effectively handle the dynamic risk data in insurance business that changes across multiple dimensions such as time, space, and industry. For example, information such as the timing and geographical distribution of natural disasters and the timing and coverage of industry policy adjustments are difficult to integrate and analyze in depth under traditional assessment models, resulting in insufficient comprehensiveness of risk assessment.

[0004] On the other hand, traditional risk assessment networks have simple structures and often use fixed feature extraction and processing procedures, making it difficult to adjust the processing path of risk features in real time according to different business scenarios and risk changes. This results in poor timeliness and accuracy of risk warnings, failing to meet the urgent needs of real-time risk management in the insurance business. Summary of the Invention

[0005] The main objective of this invention is to provide a method and system for real-time risk warning of insurance business under dynamic risk assessment, aiming to overcome the shortcomings of current risk warning methods in terms of timeliness and accuracy.

[0006] To achieve the above objectives, this invention provides a real-time risk warning method for insurance business under dynamic risk assessment, comprising the following steps:

[0007] Acquire multimodal risk data from insurance business, encompassing time, space, and industry dimensions;

[0008] The core risk element regions are separated from multimodal risk data through a masking mechanism to generate an initial risk feature map with masking labels;

[0009] The initial risk feature map is input into a dynamic risk assessment network with an encoder-decoder architecture; both the encoder and the decoder contain multi-level cascaded risk processing modules, and each risk processing module integrates a gated loop unit and is configured with a binary risk gate to dynamically control the update path of the risk features.

[0010] Based on the target risk characteristics output by the decoder, and combined with the preset risk level threshold, a real-time early warning signal is generated.

[0011] Furthermore, the encoder is used to perform a time-dimensional hybrid mapping on the initial risk feature map, including:

[0012] After normalizing the initial risk feature map, the target update probability of the current window is calculated based on the state update probability and change value of the previous window through the multi-stage time window iteration mechanism of the gated loop unit.

[0013] The target update probability is rounded to generate the gating parameters of the binary risk gate. The hidden state features are updated or kept at their original values ​​based on the gating parameters. After full connection processing, the temporal risk features are output.

[0014] The encoder is also used for spatial dimension mixing mapping, including:

[0015] The time-series risk characteristics are segmented into regional risk sub-characteristics and industry risk sub-characteristics along the risk dimension channel;

[0016] The regional risk sub-features are convolved along the time axis in the sub-dimensions to generate regional risk propagation features.

[0017] By fusing regional propagation characteristics with industry risk sub-characteristics through element-wise multiplication, a risk feature matrix representing spatiotemporal correlation is obtained;

[0018] The decoder performs multi-scale feature reconstruction on the risk feature matrix representing spatiotemporal correlation to obtain the target risk feature.

[0019] Furthermore, the training process of the dynamic risk assessment network includes:

[0020] Construct training samples;

[0021] The training samples are used to extract features through a dynamic risk assessment network to generate a risk feature prediction distribution.

[0022] The variance matrix of the predicted distribution is calculated, and the objective loss function is constructed by combining the difference between the predicted distribution and the risk label with the variance matrix. The dynamic risk assessment network is then trained end-to-end.

[0023] Furthermore, the construction of the target loss function includes:

[0024] In the first stage, the accuracy of the model for risk variance is constrained based on the spatiotemporal coordinate differences and variance matrix between the predicted distribution and the risk label.

[0025] In the second stage, weights are assigned to features of different dimensions based on risk variance to enhance feature learning in high uncertainty regions.

[0026] The dynamic risk assessment network is trained sequentially based on a two-stage loss function.

[0027] Furthermore, the variance matrix includes time dimension variance, spatial dimension variance, and spatiotemporal coupling variance, and the accuracy of the constraint model for the risk variance includes:

[0028] Calculate the characteristic differences between the predicted distribution and the risk label in the time, space and spatiotemporal coupling dimensions;

[0029] The first risk loss term is obtained by weighting the feature differences with a function that uses the variance of each dimension as an exponent and summing the results.

[0030] The second risk loss term is determined based on the differences in risk level distribution and the variance matrix.

[0031] The first risk loss term and the second risk loss term are weighted and summed to obtain the target loss function for the first stage.

[0032] Furthermore, the dynamic control mechanism of the binary risk gate includes:

[0033] For the current time window, based on the state update probability sequence and change gradient of the previous window, the target update probability matrix is ​​predicted by the logistic regression model.

[0034] The target update probability matrix is ​​binarized to generate a gated mask. The attention head of the gated recurrent unit is selectively activated based on the gated mask to achieve adaptive weighted aggregation of temporal risk features.

[0035] Furthermore, the multimodal risk data includes policyholder behavior sequences, geographical area risk indicators, and industry claims rates.

[0036] This invention also provides a real-time risk early warning system for insurance business under dynamic risk assessment, comprising:

[0037] The acquisition module is used to acquire multimodal risk data in insurance business that includes time, space, and industry dimensions;

[0038] The generation module is used to separate the core risk element regions from multimodal risk data through a masking mechanism and generate an initial risk feature map with masking labels.

[0039] The input module is used to input the initial risk feature map into the dynamic risk assessment network of the encoder-decoder architecture; the encoder and decoder both contain multi-level cascaded risk processing modules, each risk processing module integrates a gated loop unit and is configured with a binary risk gate to dynamically control the update path of the risk features;

[0040] The early warning module is used to generate real-time early warning signals based on the target risk characteristics output by the decoder and in combination with preset risk level thresholds.

[0041] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods described above.

[0042] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the methods described above.

[0043] The present invention provides a method and system for real-time risk early warning of insurance business under dynamic risk assessment, comprising: acquiring multimodal risk data of insurance business including time, space, and industry dimensions; separating core risk element regions from the multimodal risk data through a masking mechanism to generate an initial risk feature map with mask markings; inputting the initial risk feature map into a dynamic risk assessment network with an encoder-decoder architecture; wherein both the encoder and decoder contain multi-level cascaded risk processing modules, each risk processing module integrating a gated loop unit and configuring a binary risk gate for dynamically controlling the update path of risk features; and generating a real-time early warning signal based on the target risk features output by the decoder and combined with a preset risk level threshold. In this invention, multimodal risk data of insurance business including time, space, and industry dimensions is acquired, enabling real-time risk early warning of insurance business from multiple dimensions; simultaneously, the special network structure of the dynamic risk assessment network allows for dynamic adjustment of the risk feature processing path, improving the timeliness and accuracy of risk early warning. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the steps of a real-time risk warning method for insurance business under dynamic risk assessment in one embodiment of the present invention;

[0045] Figure 2 This is a structural block diagram of a real-time risk warning system for insurance business under dynamic risk assessment in one embodiment of the present invention;

[0046] Figure 3 This is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0047] The implementation, functional features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0049] Reference Figure 1One embodiment of the present invention provides a method for real-time risk warning of insurance business under dynamic risk assessment, comprising the following steps:

[0050] Step S1: Obtain multimodal risk data from the insurance business, including time, space, and industry dimensions;

[0051] Step S2: Separate the core risk element regions from the multimodal risk data using a masking mechanism to generate an initial risk feature map with masked labels;

[0052] Step S3: Input the initial risk feature map into the dynamic risk assessment network of the encoder-decoder architecture; wherein the encoder and decoder both contain multi-level cascaded risk processing modules, and each risk processing module integrates a gated loop unit and is configured with a binary risk gate to dynamically control the update path of the risk features;

[0053] Step S4: Based on the target risk characteristics output by the decoder, a real-time early warning signal is generated in combination with a preset risk level threshold.

[0054] In this embodiment, as described in step S1 above, the risk situation is affected by various factors during the operation of insurance business. These factors are presented in different dimensions and modalities of data. Time-dimensional data includes the time series of historical risk events and the risk change trends at each stage of the business cycle; spatial-dimensional data covers the geographical location information of risk events and the differences in risk characteristics in different regions; industry-dimensional data includes changes in industry policies, industry development trends, and risk cases within the same industry. Through multi-source data acquisition technologies, such as from insurance business databases, external industry data platforms, and geographic information systems, these different dimensions and modalities of data are collected and integrated to form a comprehensive and information-rich multimodal risk dataset, providing basic data support for subsequent risk assessment and early warning.

[0055] As described in step S2 above, a masking mechanism is used to separate core risk element regions from multimodal risk data, generating an initial risk feature map with masked labels. Since multimodal risk data contains a large amount of redundant information, not all data is equally important for risk assessment. The masking mechanism filters and analyzes multimodal risk data by setting preset rules and algorithms. Its principle is similar to adding a "filter" to the data, identifying core risk element regions that play a crucial role in risk assessment, such as high-risk areas within a specific time period or business segments significantly affected by industry policies. These core risk element regions are marked and combined with the original data to transform it into a visualized initial risk feature map. This feature map highlights the core risk elements through masking, allowing subsequent processing to focus on key risk information and improving the efficiency and accuracy of risk assessment.

[0056] As described in step S3 above, the initial risk feature map is input into the dynamic risk assessment network of the encoder-decoder architecture. Both the encoder and decoder contain multi-level cascaded risk processing modules. Each risk processing module integrates a gated recurrent unit (GRU) and is configured with a binary risk gate to dynamically control the update path of the risk features. The initial risk feature map, as visualized data containing core risk elements, first enters the encoder. The multi-level cascaded risk processing modules in the encoder perform progressive feature extraction and abstraction on the input feature map. The gated recurrent unit (GRU) in each risk processing module is a recurrent neural network unit that can capture the dynamic information of risk features changing over time or other dimensions in the feature map. It controls the inflow and outflow of information through a gating mechanism, selectively retaining and updating risk features. For example, when processing time-series related risk features, the GRU can remember past risk states and, combined with the current input, determine how to update the current risk feature representation.

[0057] The configured binary risk gate further enhances the flexibility and dynamism of risk feature processing. The binary risk gate outputs either 0 or 1. When the risk gate outputs 1, the current risk processing module is allowed to update and transform the input risk features; when the output is 0, the current module is blocked from updating the risk features, allowing the risk features to be directly passed to the next-level module. This mechanism can dynamically select the update path for risk features based on different business scenarios and risk changes, avoiding unnecessary feature transformations while retaining key risk information.

[0058] After processing by the encoder, the risk features are abstracted into higher-level feature representations, which are then passed to the decoder. The decoder, also composed of multi-level cascaded risk processing modules, decodes and reconstructs the abstract features output by the encoder, restoring them to target risk features that more closely resemble the original data features and contain deeper risk information. During the decoder's processing, the GRU and binary risk gates within the risk processing modules continue to function, further optimizing and adjusting the risk features so that the final output target risk features accurately reflect the current risk status of the insurance business.

[0059] As described in step S4 above, a real-time early warning signal is generated based on the target risk characteristics output by the decoder, combined with a preset risk level threshold. The preset risk level threshold is pre-set based on factors such as historical risk data of the insurance business, industry standards, and the company's own risk tolerance, serving as a boundary value for classifying different risk levels. When the decoder outputs the target risk characteristics, it compares and analyzes them with the preset risk level threshold. If the risk level reflected by the target risk characteristics exceeds a certain threshold, a corresponding real-time early warning signal is generated based on the corresponding risk level. This early warning signal can promptly notify relevant personnel in the insurance business so that they can take targeted risk response measures, achieving real-time monitoring and effective management of insurance business risks.

[0060] In this invention, multimodal risk data including time, space, and industry dimensions are acquired from insurance business to provide real-time risk warnings for insurance business from multiple dimensions; at the same time, through the special network structure of the dynamic risk assessment network, the processing path of risk characteristics can be dynamically adjusted, thereby improving the timeliness and accuracy of risk warnings.

[0061] In one embodiment, the encoder is used to perform a time-dimensional hybrid mapping on the initial risk feature map, including:

[0062] After normalizing the initial risk feature map, the target update probability of the current window is calculated based on the state update probability and change value of the previous window through the multi-stage time window iteration mechanism of the gated loop unit.

[0063] The target update probability is rounded to generate the gating parameters of the binary risk gate. The hidden state features are updated or kept at their original values ​​based on the gating parameters. After full connection processing, the temporal risk features are output.

[0064] The encoder is also used for spatial dimension mixing mapping, including:

[0065] The time-series risk characteristics are segmented into regional risk sub-characteristics and industry risk sub-characteristics along the risk dimension channel;

[0066] The regional risk sub-features are convolved along the time axis in the sub-dimensions to generate regional risk propagation features.

[0067] By fusing regional propagation characteristics with industry risk sub-characteristics through element-wise multiplication, a risk feature matrix representing spatiotemporal correlation is obtained;

[0068] The decoder performs multi-scale feature reconstruction on the risk feature matrix representing spatiotemporal correlation to obtain the target risk feature.

[0069] In this embodiment, during the time-dimensional hybrid mapping process of the encoder:

[0070] First, the initial risk feature map is normalized. This step ensures that the data in the initial risk feature map is consistent in format and range, eliminating the interference of data differences on subsequent analysis and laying the foundation for accurate risk feature extraction.

[0071] Next, a multi-stage time window iteration mechanism using gated loop units is employed. This mechanism divides time into multiple windows, and when calculating the target update probability for the current time window, it references the state update probability and change value of the previous window. This is similar to planning the adjustment direction of current work based on the work results and trends over a period of time. In this way, it is possible to capture the dynamic changes in risk characteristics over time and determine the degree to which the current window retains historical information and introduces new information.

[0072] Then, the calculated target update probability is rounded to generate the gating parameters for the binary risk gate. The binary risk gate acts like a smart switch, with only two states: 0 and 1. When the gating parameter is 1, the switch is open, allowing the current risk processing module to update its hidden state features and incorporate new risk information; when the gating parameter is 0, the switch is closed, and the hidden state features are directly passed to the next-level module, maintaining the original state.

[0073] Finally, after fully connected processing, the updated hidden state features are integrated and transformed to output time-series risk features containing temporal dependencies. These time-series risk features now contain information on the changes in risk over time, providing a time-series-level risk basis for subsequent analysis.

[0074] In the spatial dimension hybrid mapping process of the encoder:

[0075] The first step is to segment the obtained time-series risk characteristics into regional risk sub-characteristics and industry risk sub-characteristics according to the risk dimension channels. This is equivalent to breaking down a complex risk entity into a geographically related regional risk component and an industry-related risk component, which facilitates in-depth analysis of different types of risks separately.

[0076] The second step involves performing convolution operations along the time axis on the sub-dimensions of regional risk. Convolution acts like an "information detector," sliding along the time dimension to capture the patterns of regional risk propagation and change at different points in time, uncovering the diffusion trends and evolution patterns of regional risks over time, and ultimately generating regional risk propagation characteristics.

[0077] The third step involves fusing regional risk propagation characteristics with industry risk sub-features through element-wise multiplication. This operation multiplies corresponding elements of the two features, combining the temporal evolution information of regional risk with industry risk information to obtain a risk feature matrix representing spatiotemporal correlations. Each element in this matrix reflects the interaction between regional and industry risks at a specific spatiotemporal location, achieving a comprehensive mapping of risks across spatial and temporal dimensions.

[0078] When the decoder performs multi-scale feature reconstruction on the risk feature matrix representing spatiotemporal correlation, it relies on its multi-level cascaded risk processing module architecture, combined with gated loop units and binary risk gates, to achieve dynamic processing of risk features. The specific process is as follows:

[0079] First, the risk feature matrix representing spatiotemporal correlation is input into the decoder and processed by the first-level risk processing module. The gated loop unit within this module analyzes the currently input risk features and, combined with historically transmitted risk state information, selectively retains or updates risk features through a gating mechanism, capturing the dynamic changes of risk features over time. Simultaneously, the binary risk gate calculates a 0 or 1 state value based on a preset algorithm and the input features. When the value is 1, the current module is allowed to update the risk features, enabling the module to transform and extract the input features. When the value is 0, the update operation is blocked, allowing the risk features to be directly transmitted to the next-level module, thus initially screening and controlling the processing path of the risk features.

[0080] Subsequently, the risk features processed at the first level enter the next level of risk processing module, repeating the collaborative processing process of the gating loop unit and the binary risk gate. Different cascaded risk processing modules, configured with different parameters, mine risk features from multiple scales; for example, some modules focus on extracting local, detailed features of the risk, while others focus on obtaining global, macroscopic features. During each level of processing, the gating loop unit continuously updates the temporal dimension information of the risk features, while the binary risk gate dynamically adjusts the processing path according to changes in the risk features, ensuring that key risk features at different scales are preserved and enhanced.

[0081] After risk features have been processed sequentially by multiple risk processing modules, a feature set containing multi-scale information has been formed. At this point, the decoder integrates the risk features obtained from each processing stage through a specific fusion mechanism. During the fusion process, the gated loop unit further refines the fused features, optimizing the correlations between them; the binary risk gate then performs a final update path control based on the overall risk feature state after fusion, removing redundant information and strengthening the representation of core risk features.

[0082] Finally, the fused risk features undergo standardization and other post-processing operations to output the final target risk features. These target risk features not only fully preserve the spatiotemporal correlation characteristics of risks but also integrate multi-scale feature representations and temporal evolution information, providing a high-quality data foundation for subsequent accurate risk level assessment and real-time early warning based on risk features.

[0083] In one embodiment, the training process of the dynamic risk assessment network includes:

[0084] Construct training samples;

[0085] The training samples are used to extract features through a dynamic risk assessment network to generate a risk feature prediction distribution.

[0086] The variance matrix of the predicted distribution is calculated, and the objective loss function is constructed by combining the difference between the predicted distribution and the risk label with the variance matrix. The dynamic risk assessment network is then trained end-to-end.

[0087] In this embodiment, the first step is to construct a training sample set containing multimodal risk data and corresponding risk labels. The module collects raw risk data from the insurance business across time, space, and industry dimensions, such as historical claims records, geographical area risk indices, and industry policy changes. The raw data undergoes preprocessing, including data cleaning to remove noise and outliers, feature standardization to unify dimensions, and time series alignment to ensure consistency across the data's time dimension. Then, based on factors such as the severity and scope of impact of historical risk events, each sample is labeled with a corresponding risk level label, forming training sample pairs containing input features and target labels. To enhance the model's generalization ability, data augmentation processing is also performed on the training samples, such as time series perturbation, spatial region masking, and slight feature value shifts, expanding the diversity of the training samples.

[0088] Then, the constructed training samples are input into the dynamic risk assessment network for feature extraction. The encoder part of the network first performs spatiotemporal hybrid mapping processing on the multimodal risk data: through the multi-stage time window iteration mechanism of the gated recurrent unit, combined with the dynamic control of the feature update path by the binary risk gate, temporal risk features are extracted; at the same time, features are segmented along the risk dimension channel, regional risks and industry risks are processed separately and fused to generate a risk feature matrix representing spatiotemporal correlation. The decoder performs multi-scale feature reconstruction on this matrix, also using a multi-level cascaded risk processing module and a binary risk gate to mine risk features from different scales, and finally outputs the target risk features.

[0089] Unlike traditional deterministic outputs, this network employs a probabilistic modeling approach, outputting a predicted distribution of risk characteristics. Specifically, the network not only predicts the mean of the risk characteristics but also their variance, forming a complete probability distribution representation that reflects the model's uncertainty in predicting the risk characteristics.

[0090] Finally, the variance matrix of the predicted distribution is calculated. This matrix contains covariance information between the dimensions of the risk characteristics, reflecting the structure of prediction uncertainty. Combining the difference between the predicted distribution and the risk labels, along with the variance matrix, a target loss function is constructed. The loss function is designed as a sum of two parts: the first part is the negative log-likelihood loss between the predicted distribution and the true risk labels, measuring the deviation of the center of the predicted distribution from the true labels, thus encouraging the model to accurately predict risk levels. The second part is a regularization term based on the variance matrix, penalizing the uncertainty of the model's predictions. By minimizing this loss, the model learns to reduce variance in high-confidence regions and increase variance in low-confidence regions, thereby better capturing risk uncertainty. In particular, the variance matrix plays a dual role in the loss function: on the one hand, it adjusts the weights of different risk dimensions through covariance information, making the model focus more on key risk factors; on the other hand, it uses variance information to quantify the model's confidence in the predictions, achieving effective modeling of risk uncertainty.

[0091] In this embodiment, an end-to-end training approach is adopted, using the backpropagation algorithm to simultaneously optimize the parameters of the encoder and decoder, minimizing the target loss function. During training, the learning rate is dynamically adjusted to ensure convergence stability, and an early stopping strategy is used to prevent overfitting. The final model can not only accurately predict risk levels but also quantify the uncertainty of the prediction.

[0092] In one embodiment, constructing the target loss function includes:

[0093] In the first stage, the accuracy of the model for risk variance is constrained based on the spatiotemporal coordinate differences and variance matrix between the predicted distribution and the risk label.

[0094] In the second stage, weights are assigned to features of different dimensions based on risk variance to enhance feature learning in high uncertainty regions.

[0095] The dynamic risk assessment network is trained sequentially based on a two-stage loss function.

[0096] In this embodiment, the first stage, in the initial phase of model training, focuses on constructing an accurate constraint mechanism for the risk prediction variance. First, the difference between the predicted distribution and the actual risk label in spatiotemporal coordinates is calculated; this difference reflects the degree of bias in the model's prediction. Simultaneously, the variance matrix of the predicted distribution is extracted, containing uncertainty information for each risk feature dimension. By designing a specific loss function term, the spatiotemporal coordinate difference is combined with the variance matrix to form a constraint on the risk variance. Specifically, the loss function penalizes the deviation between the predicted variance and the actual uncertainty, prompting the model to learn a variance representation that accurately reflects risk fluctuations. For example, for spatiotemporal regions with drastic risk changes, the model should output a larger variance value; while for stable regions, the variance value should decrease accordingly. This constraint mechanism ensures that the model can not only predict the average level of risk but also accurately quantify the uncertainty of the prediction, providing a reliable foundation for feature weight allocation in subsequent stages.

[0097] In the second stage, after the model has initially grasped the accurate representation of risk variance, it enters the dynamic feature weight allocation stage. Based on the risk variance learned in the first stage, adaptive weights are assigned to features of different dimensions. Specifically, for high-variance regions (i.e., regions with high model prediction uncertainty), the weights of the corresponding feature dimensions are increased, allowing the model to invest more learning resources in these regions and strengthen its ability to capture complex risk patterns; for low-variance regions (i.e., regions with relatively certain model predictions), the feature weights are appropriately reduced to avoid overfitting. This risk variance-based weight allocation mechanism is essentially a dynamic attention mechanism, enabling the model to focus on high-risk, high-uncertainty regions and features, improving its sensitivity to key risk factors. In this way, the model can more effectively learn feature representations under different risk scenarios, enhancing its ability to identify abnormal risks.

[0098] A two-stage sequential training approach is adopted to fully leverage the advantages of different loss functions. In the first stage, the focus is on optimizing the risk variance-constrained loss function, enabling the model to accurately represent risk uncertainty. This stage provides a reliable variance estimation foundation for subsequent feature weight allocation. In the second stage, the variance estimation parameters obtained in the first stage are fixed, and the focus shifts to optimizing the feature weight allocation loss function based on risk variance. Through this phased training, the model gradually transitions from learning the basic representation of risk to fine-grained modeling of high-uncertainty regions, achieving hierarchical learning of risk features. This strategy effectively solves the convergence difficulty problem caused by simultaneously optimizing multiple objectives in traditional training methods, enabling the model to more stably learn comprehensive and in-depth risk feature representations, ultimately improving the early warning accuracy and reliability of the dynamic risk assessment network.

[0099] In one embodiment, the variance matrix includes time dimension variance, spatial dimension variance, and spatiotemporal coupling variance, and the accuracy of the constraint model for the risk variance includes:

[0100] Calculate the characteristic differences between the predicted distribution and the risk label in the time, space and spatiotemporal coupling dimensions;

[0101] The first risk loss term is obtained by weighting the feature differences with a function that uses the variance of each dimension as an exponent and summing the results.

[0102] The second risk loss term is determined based on the differences in risk level distribution and the variance matrix.

[0103] The first risk loss term and the second risk loss term are weighted and summed to obtain the target loss function for the first stage.

[0104] In this embodiment, the characteristic differences between the predicted distribution and the actual risk labels in the time, space, and spatiotemporal coupling dimensions are calculated first. In the time dimension, the difference between the predicted risk trend over time and the actual risk in the corresponding time series is compared; for example, the degree of agreement between the model's predicted high-risk periods and the actual occurrence periods is assessed. The spatial dimension focuses on the deviation between the predicted risk distribution in geographical areas and the actual spatial distribution of risk; for example, the accuracy of the model's predictions of risk levels in different regions is checked. The spatiotemporal coupling dimension focuses on the risk differences under the combined influence of time and space factors, analyzing whether the model accurately captures the comprehensive risk characteristics of a specific time point and specific region. By quantifying the differences in these dimensions, the degree of deviation between the predicted results and the actual situation in multiple dimensions is obtained, providing basic data for subsequent weighted processing.

[0105] Then, using the time dimension variance, spatial dimension variance, and spatiotemporal coupling variance in the variance matrix, a function with the variance of each dimension as an exponent is constructed to weight the feature differences calculated in step one. Variance reflects the uncertainty of risk in the corresponding dimension; the larger the variance, the higher the volatility and uncertainty of the risk in that dimension. Therefore, a function with variance as an exponent will give greater weight to feature differences in high-variance dimensions during the weighting process, meaning the model is more sensitive to prediction biases in high-uncertainty dimensions during training. The weighted feature differences of the time, space, and spatiotemporal coupling dimensions are summed to obtain the first risk loss term. This process allows the model to dynamically adjust the focus on feature differences in different dimensions according to the degree of uncertainty of risk during training, strengthening the learning of high-risk, high-uncertainty regions.

[0106] Furthermore, in addition to considering the differences in features across various dimensions, the model needs to be further constrained from the perspective of risk level distribution. This involves analyzing the difference between the predicted risk level distribution and the actual risk level distribution, such as comparing whether the proportions of low, medium, and high risk level samples predicted by the model match the actual situation. Simultaneously, the variance matrix is ​​used to consider the uncertainty of risk level distribution differences across different time and spatial dimensions. Through a specific calculation method, the differences in risk level distribution and the uncertainty reflected in the variance matrix are integrated to determine the second risk loss term. This second risk loss term ensures that the model not only approximates the true values ​​in specific feature dimensions but also accurately reflects the actual situation from the perspective of the overall risk level distribution, avoiding systematic misjudgments of risk levels by the model.

[0107] Finally, the first and second risk loss terms are summed according to pre-defined weights to form the target loss function for the first stage. The weights are set to comprehensively consider the importance of different loss terms in constraining the accuracy of the model's risk variance and can be adjusted through experiments or based on domain knowledge. By minimizing this target loss function, the dynamic risk assessment network is driven to optimize its parameters, enabling the model to continuously correct the differences between the predicted distribution and the true risk label across multiple dimensions during training, thereby improving the accuracy of risk variance prediction.

[0108] In one embodiment, the dynamic control mechanism of the binary risk gate includes:

[0109] For the current time window, based on the state update probability sequence and change gradient of the previous window, the target update probability matrix is ​​predicted by the logistic regression model.

[0110] The target update probability matrix is ​​binarized to generate a gated mask. The attention head of the gated recurrent unit is selectively activated based on the gated mask to achieve adaptive weighted aggregation of temporal risk features.

[0111] In this embodiment, when processing the risk features of the current time window, the binary risk gate first uses the state update probability sequence and change gradient of the previous window as input to build the prediction basis. The state update probability sequence records the update probability of each risk feature dimension in the previous window, reflecting the model's confidence in historical risk states; the change gradient describes the rate of change of these probabilities over time, capturing the dynamic evolution trend of risk features.

[0112] These two sets of data are processed using a logistic regression model. Based on the input features, the model learns the potential patterns of risk state changes and outputs a target update probability matrix. Each element in this matrix represents the update probability of the corresponding risk feature dimension within the current window, with values ​​ranging from [0, 1]. This design enables the model to dynamically predict the risk feature dimensions that currently require focus, based on historical risk evolution patterns, thus achieving a proactive response to risk changes.

[0113] Then, the target update probability matrix is ​​converted into a gating mask through threshold binarization, typically using 0.5 as the threshold. Elements greater than this value are converted to 1, and those less than 0 are converted to 0. The generated gating mask is a binary matrix, where a value of 1 indicates that the corresponding risk feature dimension needs to be given priority in the current window, and a value of 0 indicates that it can be temporarily ignored.

[0114] In the gated recurrent unit, each attention head is responsible for capturing different types of risk feature association patterns. The gating mask selectively activates these attention heads, achieving adaptive weighted aggregation of temporal risk features. Specifically, when a position in the gating mask is 1, the corresponding attention head is activated, and its output risk feature weight increases, indicating that this feature dimension is more important for the current risk assessment; conversely, when the gating mask is 0, the output weight of the corresponding attention head is suppressed, reducing interference from irrelevant features. This dynamic control mechanism allows the model to adaptively adjust its attention to different risk feature dimensions based on real-time changes in risk. During periods of high risk volatility, the model automatically strengthens its focus on high-uncertainty features, improving early warning sensitivity; while during periods of relatively stable risk, it reduces the processing of redundant features, improving computational efficiency. In this way, binary risk gating effectively enhances the model's adaptability to complex risk scenarios and its predictive accuracy.

[0115] In one embodiment, the multimodal risk data includes policyholder behavior sequences, geographical area risk indicators, and industry claims rates.

[0116] In one embodiment, when acquiring multimodal risk data in insurance business that includes time, space, and industry dimensions, a hierarchical data transmission network is constructed; dedicated communication queue pairs are allocated for risk data from different sources and of different importance, and meta-information tags containing data type, update time, and risk level identifier are generated for each data block;

[0117] During data transmission, the transmission priority of communication queue pairs is dynamically adjusted based on the risk level identifier in the metadata tag. At the same time, the data integrity is monitored in real time through the metadata verification mechanism. Once a data anomaly is detected, the corresponding data block is immediately requested to be retransmitted from the original data source.

[0118] In this embodiment, within the insurance business scenario, multimodal risk data includes various types such as policyholder behavior sequences, geographic area risk indicators, and industry claim rates. Policyholder behavior sequences record the chronological order and specific operations of policyholders' actions such as purchasing, canceling, and making claims; geographic area risk indicators cover spatially related risk factors such as the frequency of natural disasters and the level of economic development in different regions; and industry claim rates reflect the proportion of insurance payouts in various industries. These data come from a wide range of sources and vary in importance. To ensure efficient data transmission, a hierarchical data transmission network is constructed, assigning dedicated communication queue pairs to risk data from different sources and of different importance. These dedicated communication queue pairs act as "dedicated channels" for data transmission, ensuring that different types of data transmission do not interfere with each other. Simultaneously, a metadata tag is generated for each data block, containing the data type (clearly specifying whether it is a policyholder behavior sequence, geographic area risk indicator, or industry claim rate), update time (recording the moment the data was last updated), and risk level identifier (classifying the risk level based on the data's reflection of risk). This operation lays the foundation for subsequent data transmission management and monitoring.

[0119] During data transmission, the risk level identifier in the metadata tag plays a crucial role. The system dynamically adjusts the transmission priority of communication queue pairs based on this identifier, assigning higher priority to data with higher risk levels, ensuring that it is transmitted through the communication queue pair first. This guarantees that important risk information is delivered to the processing end in a timely manner, avoiding delays in risk response due to transmission delays. Simultaneously, a metadata verification mechanism is used to monitor data integrity in real time. This mechanism checks for data loss, errors, or tampering during transmission. Once an anomaly is detected, the system immediately requests a retransmission of the corresponding data block from the original data source, ensuring the accuracy and completeness of the acquired multimodal risk data. This provides reliable data support for subsequent dynamic risk assessments and real-time risk warnings based on this data.

[0120] In one embodiment, after generating a real-time early warning signal based on the target risk characteristics output by the decoder and combined with a preset risk level threshold, the process includes:

[0121] For the multimodal risk data of each dimension, key characters are extracted and added sequentially to the nodes of the corresponding undirected graphs to obtain undirected graphs of time dimension, spatial dimension, and industry dimension; one character is added to each node.

[0122] Based on the characteristics of the time-dimensional undirected graph, the spatial-dimensional undirected graph is mutated to obtain a mutated spatial-dimensional undirected graph; based on the common properties of the time-dimensional and spatial-dimensional undirected graphs, the industry-dimensional undirected graph is structurally adjusted to obtain an adjusted industry-dimensional undirected graph.

[0123] Based on the real-time early warning information, an early warning level is generated, and a first visual information carrier is generated according to the early warning level;

[0124] A first string is generated based on the first visual information carrier and the time-dimensional undirected graph, and a second visual information carrier is generated based on the first string.

[0125] A second string is generated based on the second visual information carrier and the undirected graph of the variation space dimension, and a third visual information carrier is generated based on the second string.

[0126] Based on the third visual information carrier and the adjusted industry dimension undirected graph, an encrypted password string is generated to encrypt the multimodal risk data and store it in the insurance business database; wherein, the first visual information carrier, the second visual information carrier and the third visual information carrier are different forms of visual information carriers.

[0127] In this embodiment, in insurance business, multimodal risk data covers multiple dimensions such as time, space, and industry. The time dimension includes time-series data such as insurance application time and claim settlement time; the spatial dimension involves spatial information such as the policyholder's location and the geographical location of the insured object; the industry dimension includes the policyholder's industry and the industry attributes corresponding to the insurance product. For these different dimensions of data, key characters are first extracted. Key characters can be numbers in timestamps, keywords in region names, core words in industry names, etc. Then, these key characters are added to the corresponding undirected graph nodes in sequence. The nodes of the undirected graph represent the smallest information unit in the data, and only one character is added to each node, thereby constructing undirected graphs for the time dimension, spatial dimension, and industry dimension. This step transforms the abstract multimodal risk data into a structured graphical representation, facilitating subsequent analysis and processing.

[0128] The time-dimensional undirected graph records the characteristics of risk data changes over time. Based on these characteristics, mutation operations are performed on the spatial-dimensional undirected graph. Adjustments include mutation operations such as node splitting or merging, resulting in a mutated spatial-dimensional undirected graph. The mutation rules can be customized. Next, the common attributes of the time-dimensional and spatial-dimensional undirected graphs are analyzed. These common attributes can be shared structural properties of the undirected graphs. Based on these common attributes, the industry-dimensional undirected graph undergoes structural adjustments, and the adjustment rules can also be customized. Through mutation and structural adjustments, the uniqueness and decipherability of the undirected graph are enhanced.

[0129] Furthermore, based on the target risk characteristics output by the decoder and the real-time early warning signal generated by combining it with a preset risk level threshold, the early warning signal is interpreted and analyzed to determine the specific early warning level, which can be categorized as high risk, medium risk, and low risk. Based on the determined early warning level, the relevant risk information is visualized using charts, heatmaps, line graphs, etc., to intuitively present key information such as the risk level, the time range of risk occurrence, spatial distribution, and involved industries, generating the first visual information carrier. This visual information carrier can be a graph, curve, matrix, array, data table, or other similar format.

[0130] Then, key content from the first visual information carrier, such as risk level and the time points of major risks, is combined with node character information in the undirected time-dimension graph and arranged in a certain logical order to generate the first string. For example, the time node characters corresponding to high-risk levels are sequentially connected with the characters of relevant time nodes in the undirected time-dimension graph. Next, the first string undergoes further visualization processing, using formats different from the first visual information carrier, such as text lists or tree diagrams, to clearly display the information in the first string, generating the second visual information carrier. This process achieves the transformation from graphical information to character information, and then to new visual information, presenting risk information from another perspective.

[0131] Furthermore, the key data in the second visualization information carrier is combined with the node character information in the undirected graph of the variation space dimension, and arranged and combined according to specific rules to generate a second string. For example, the risk area information involved in the second visualization information carrier is combined with the characters of the corresponding area nodes in the undirected graph of the variation space dimension. Then, different visualization methods such as map annotations and matrix diagrams can be used to visualize the second string, resulting in a third visualization information carrier. The third visualization information carrier further enriches the display dimensions of risk information, presenting risks in a more detailed spatial perspective.

[0132] Finally, the key information in the third-dimensional visualization carrier is integrated with the node character information in the adjusted industry-dimensional undirected graph, and arranged according to a predetermined encoding rule to generate an encrypted password string. This encoding rule can be customized, including character replacement and position transformation. The generated encrypted password string is used to encrypt the original multimodal risk data, ensuring data security. After encryption, the encrypted data is stored in the insurance business database for subsequent querying, analysis, and backtracking. The entire process, through multi-dimensional data processing, visualization transformation, and encrypted storage, achieves comprehensive management and security protection of insurance business risk data.

[0133] In one embodiment, after generating a real-time early warning signal based on the target risk characteristics output by the decoder and combined with a preset risk level threshold, the process includes:

[0134] The multi-modal risk data is mapped to a three-dimensional index table, with each index item corresponding to a feature value. The time dimension is indexed, the spatial dimension is indexed, and the industry dimension is indexed. Each index item contains three attributes: feature value, confidence score, and timestamp, forming a three-dimensional index table.

[0135] Based on the temporal changes of the time-dimensional index, the spatial-dimensional index is clustered into regions to generate a spatiotemporal correlation matrix;

[0136] By indexing business types by industry dimension, the spatiotemporal correlation matrix is ​​expanded into a three-dimensional risk propagation network;

[0137] The real-time early warning signal is encoded into a heat value matrix; through color gradient mapping, the heat value matrix is ​​converted into a risk heat distribution map to generate a visual information carrier.

[0138] Based on the visual information carrier and the time dimension index, an initial feature code is generated;

[0139] The initial feature code is transformed with the spatiotemporal correlation matrix to generate a second feature code with spatiotemporal characteristics.

[0140] By using the business type code indexed by the industry dimension to verify the second feature code, the final risk feature key is generated. After encrypting the multimodal risk data, it is stored in the insurance business database.

[0141] In this embodiment, the features of the multimodal risk data are first structured and mapped to construct a three-dimensional index table. In the time dimension, a time-dimensional index is established according to chronological order, accurately recording the timestamp of each risk data point, such as the insurance application date and claim settlement date. In the spatial dimension, a spatial dimension index is constructed based on geographical coordinates or administrative divisions, marking the specific location of the data. In the industry dimension, an industry-dimensional index is established according to different insurance business types, distinguishing between health insurance, auto insurance, and other business categories. Each index item contains three attributes: feature value, confidence score, and timestamp. The feature value quantifies the risk data, the confidence score assesses the data reliability, and the timestamp clarifies the time the data was generated. In this way, the originally scattered multimodal data is integrated into an ordered three-dimensional index table, laying the foundation for subsequent analysis.

[0142] Next, based on the temporal variation patterns presented by the time-dimensional index, regional clustering is performed on the spatial-dimensional index. This automatically identifies the fluctuation trends and periodic characteristics of risk data in the time dimension, such as discovering a continuous rise in claims rates in a specific region over a certain period. Based on this, regions with similar risk characteristics are grouped together in the spatial-dimensional index. By calculating the similarity of risk characteristics between regions, a spatiotemporal correlation matrix is ​​generated. This matrix uses time as the horizontal axis and spatial region as the vertical axis; the matrix elements reflect the degree of correlation of risk data under different spatiotemporal combinations, achieving a preliminary fusion of time and spatial dimension risk information.

[0143] Subsequently, business type information from the industry-dimensional index was incorporated to expand the spatiotemporal correlation matrix. For each spatiotemporal combination in the matrix, its corresponding insurance business type was associated, and the risk performance of different business types under those spatiotemporal conditions was analyzed. Combining historical data and industry risk transmission patterns, the risk propagation coefficient between each business type was calculated, expanding the original two-dimensional spatiotemporal correlation matrix into a three-dimensional risk propagation network that includes the industry dimension. This network clearly demonstrates how risk spreads and diffuses over time, through spatial transformations, and across industry intersections, providing a comprehensive, multi-dimensional perspective for understanding the risk landscape.

[0144] Next, the real-time early warning signals are numerically encoded and converted into a heat map matrix. Based on a pre-defined rule of matching risk levels with numerical values, each spatiotemporal region is assigned a corresponding heat map value, with the magnitude of the value representing the level of risk. Through a pre-defined color gradient mapping relationship, the heat map matrix is ​​transformed into an intuitive risk heat map, using different colors to distinguish risk levels, such as red for high-risk areas and green for low-risk areas, generating a visual information carrier. This allows insurance professionals to quickly and intuitively understand the spatiotemporal distribution of risks.

[0145] Based on the generated visual information carrier and time dimension index, key risk feature information is extracted, and an initial feature code is generated according to specific coding rules. The initial feature code contains key location information, risk level, and key time points in the time dimension of the risk heat map, thus initially integrating the visual information and time dimension data.

[0146] Next, matrix transformation operations are performed on the initial feature code and the spatiotemporal correlation matrix. Through matrix multiplication, addition, and other operations, the risk correlation information in the spatiotemporal dimension is integrated into the initial feature code, further strengthening the expression of the spatiotemporal characteristics in the feature code and generating a second feature code with spatiotemporal characteristics. This feature code not only retains the core information of the initial feature code but also incorporates the risk correlation characteristics in the spatiotemporal dimension, reflecting the spatiotemporal attributes of risk more comprehensively.

[0147] Finally, the second feature code is validated using the business type coding rules indexed by the industry dimension. A validation algorithm matches the business type code with the second feature code to ensure the accuracy and completeness of the feature code. After successful validation, a final risk feature key is generated. This key is used to encrypt the original multimodal risk data, and the encrypted risk data is stored in the insurance business database. This completes the entire process of risk data processing, analysis, and secure storage, ensuring data security and subsequent traceability.

[0148] Reference Figure 2 Another embodiment of the present invention also provides a real-time risk early warning system for insurance business under dynamic risk assessment, comprising:

[0149] The acquisition module is used to acquire multimodal risk data in insurance business that includes time, space, and industry dimensions;

[0150] The generation module is used to separate the core risk element regions from multimodal risk data through a masking mechanism and generate an initial risk feature map with masking labels.

[0151] The input module is used to input the initial risk feature map into the dynamic risk assessment network of the encoder-decoder architecture; the encoder and decoder both contain multi-level cascaded risk processing modules, each risk processing module integrates a gated loop unit and is configured with a binary risk gate to dynamically control the update path of the risk features;

[0152] The early warning module is used to generate real-time early warning signals based on the target risk characteristics output by the decoder and in combination with preset risk level thresholds.

[0153] In this embodiment, the specific implementation of each module in the above system embodiment is described in the above method embodiment, and will not be repeated here.

[0154] Reference Figure 3 This invention also provides a computer device, which can be a server, and its internal structure can be as follows: Figure 3 As shown, the computer device includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores the data corresponding to this embodiment. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements the above-described method.

[0155] Those skilled in the art will understand that Figure 3 The structures shown are merely block diagrams of some structures related to the present invention and do not constitute a limitation on the computer devices on which the present invention is applied.

[0156] An embodiment of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method. It is understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0157] In summary, the real-time risk warning method and system for insurance business under dynamic risk assessment provided in this embodiment of the invention includes: acquiring multimodal risk data of insurance business containing time, space, and industry dimensions; separating core risk element regions from the multimodal risk data through a masking mechanism to generate an initial risk feature map with mask markings; inputting the initial risk feature map into a dynamic risk assessment network with an encoder-decoder architecture; wherein both the encoder and decoder contain multi-level cascaded risk processing modules, each risk processing module integrating a gated loop unit and configuring a binary risk gate for dynamically controlling the update path of risk features; and generating a real-time warning signal based on the target risk features output by the decoder and combined with a preset risk level threshold. In this invention, multimodal risk data of insurance business containing time, space, and industry dimensions is acquired, enabling real-time risk warning of insurance business from multiple dimensions; simultaneously, through the special network structure of the dynamic risk assessment network, the processing path of risk features can be dynamically adjusted, improving the timeliness and accuracy of risk warning.

[0158] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the present invention and embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM, etc.

[0159] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0160] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for real-time risk early warning of insurance business under dynamic risk assessment, characterized in that, Includes the following steps: Acquire multimodal risk data from insurance business, encompassing time, space, and industry dimensions; The core risk element regions are separated from multimodal risk data through a masking mechanism to generate an initial risk feature map with masking labels; The masking mechanism filters and analyzes multimodal risk data by setting preset rules and algorithms, identifies the core risk element regions in the data that play a key role in risk assessment, marks the core risk element regions, and combines them with the original data to transform them into a visualized initial risk feature map. The initial risk feature map is input into the dynamic risk assessment network of the encoder-decoder architecture; Both the encoder and decoder contain multi-level cascaded risk processing modules. Each risk processing module integrates a gated loop unit and is configured with a binary risk gate to dynamically control the update path of risk features. Based on the target risk features output by the decoder, and combined with a preset risk level threshold, a real-time early warning signal is generated. For the multimodal risk data across various dimensions, key characters are extracted and sequentially added to the nodes of the corresponding undirected graphs, resulting in time-dimension, spatial-dimension, and industry-dimension undirected graphs. One character is added to each node. Based on the characteristics of the time-dimension undirected graph, the spatial-dimension undirected graph is mutated to obtain a mutated spatial-dimension undirected graph. Based on the common attributes of the time-dimension and spatial-dimension undirected graphs, the industry-dimension undirected graph is structurally adjusted to obtain an adjusted industry-dimension undirected graph. An early warning level is generated based on the real-time early warning signal, and a warning level is determined according to the preset risk level threshold. The system generates a first visual information carrier based on the alert level; a first string is generated based on the first visual information carrier and the undirected graph of the time dimension, and a second visual information carrier is generated based on the first string; a second string is generated based on the second visual information carrier and the undirected graph of the variation space dimension, and a third visual information carrier is generated based on the second string; an encrypted password string is generated based on the third visual information carrier and the undirected graph of the adjustment industry dimension, and the multimodal risk data is encrypted and stored in the insurance business database for subsequent query, analysis and retrospection; wherein, the first visual information carrier, the second visual information carrier and the third visual information carrier are different forms of visual information carriers.

2. The method for real-time risk early warning of insurance business under dynamic risk assessment according to claim 1, characterized in that, The encoder is used to perform a time-dimensional hybrid mapping on the initial risk feature map, including: After normalizing the initial risk feature map, the target update probability of the current window is calculated based on the state update probability and change value of the previous window through the multi-stage time window iteration mechanism of the gated loop unit. The target update probability is rounded to generate the gating parameters of the binary risk gate. The hidden state features are updated or kept at their original values ​​based on the gating parameters. After full connection processing, the temporal risk features are output. The encoder is also used for spatial dimension mixing mapping, including: The time-series risk characteristics are segmented into regional risk sub-characteristics and industry risk sub-characteristics along the risk dimension channel; The regional risk sub-features are convolved along the time axis in the sub-dimensions to generate regional risk propagation features. By fusing regional propagation characteristics with industry risk sub-characteristics through element-wise multiplication, a risk feature matrix representing spatiotemporal correlation is obtained; The decoder performs multi-scale feature reconstruction on the risk feature matrix representing spatiotemporal correlation to obtain the target risk feature.

3. The method for real-time risk early warning of insurance business under dynamic risk assessment according to claim 1, characterized in that, The training process of the dynamic risk assessment network includes: Construct training samples; The training samples are used to extract features through a dynamic risk assessment network to generate a risk feature prediction distribution. The variance matrix of the predicted distribution is calculated, and the objective loss function is constructed by combining the difference between the predicted distribution and the risk label with the variance matrix. The dynamic risk assessment network is then trained end-to-end.

4. The real-time risk early warning method for insurance business under dynamic risk assessment according to claim 3, characterized in that, The construction of the target loss function includes: In the first stage, the accuracy of the model for risk variance is constrained based on the spatiotemporal coordinate differences and variance matrix between the predicted distribution and the risk label. In the second stage, weights are assigned to features of different dimensions based on risk variance to enhance feature learning in high uncertainty regions. The dynamic risk assessment network is trained sequentially based on a two-stage loss function.

5. The real-time risk early warning method for insurance business under dynamic risk assessment according to claim 4, characterized in that, The variance matrix includes time dimension variance, spatial dimension variance, and spatiotemporal coupling variance. The accuracy of the constraint model for risk variance includes: Calculate the characteristic differences between the predicted distribution and the risk label in the time, space and spatiotemporal coupling dimensions; The first risk loss term is obtained by weighting the feature differences with a function that uses the variance of each dimension as an exponent and summing the results. The second risk loss term is determined based on the differences in risk level distribution and the variance matrix. The first risk loss term and the second risk loss term are weighted and summed to obtain the target loss function for the first stage.

6. The method for real-time risk early warning of insurance business under dynamic risk assessment according to claim 2, characterized in that, The dynamic control mechanism of the binary risk gate includes: For the current time window, based on the state update probability sequence and change gradient of the previous window, the target update probability matrix is ​​predicted by the logistic regression model. The target update probability matrix is ​​binarized to generate a gated mask. The attention head of the gated recurrent unit is selectively activated based on the gated mask to achieve adaptive weighted aggregation of temporal risk features.

7. The method for real-time risk early warning of insurance business under dynamic risk assessment according to claim 1, characterized in that, The multimodal risk data includes policyholder behavior sequences, geographical area risk indicators, and industry claims rates.

8. A real-time risk early warning system for insurance business under dynamic risk assessment, characterized in that, include: The acquisition module is used to acquire multimodal risk data in insurance business that includes time, space, and industry dimensions; The generation module is used to separate the core risk element regions from multimodal risk data through a masking mechanism and generate an initial risk feature map with masking labels. The masking mechanism filters and analyzes multimodal risk data by setting preset rules and algorithms, identifies the core risk element regions in the data that play a key role in risk assessment, marks the core risk element regions, and combines them with the original data to transform them into a visualized initial risk feature map. The input module is used to input the initial risk feature map into the dynamic risk assessment network of the encoder-decoder architecture; Both the encoder and decoder contain multi-level cascaded risk processing modules. Each risk processing module integrates a gated loop unit and is configured with a binary risk gate to dynamically control the update path of risk features. The early warning module is used to generate real-time early warning signals based on the target risk characteristics output by the decoder and in combination with preset risk level thresholds. For the multimodal risk data across various dimensions, key characters are extracted and sequentially added to the nodes of the corresponding undirected graphs, resulting in undirected graphs of time, space, and industry dimensions; one character is added to each node; based on the characteristics of the time-dimension undirected graph, the spatial-dimension undirected graph is mutated to obtain a mutated spatial-dimension undirected graph; based on the common attributes of the time-dimension and spatial-dimension undirected graphs, the structure of the industry-dimension undirected graph is adjusted to obtain an adjusted industry-dimension undirected graph; a warning level is generated based on the real-time warning signal, and a first visual information carrier is generated based on the warning level; based on... The first visual information carrier and the time-dimensional undirected graph generate a first string, and a second visual information carrier is generated based on the first string; a second string is generated based on the second visual information carrier and the variation space-dimensional undirected graph, and a third visual information carrier is generated based on the second string; an encrypted password string is generated based on the third visual information carrier and the adjustment industry-dimensional undirected graph, and the multimodal risk data is encrypted and stored in the insurance business database for subsequent query, analysis and retrospection; wherein, the first visual information carrier, the second visual information carrier and the third visual information carrier are different forms of visual information carriers.

Images