Insurance sales behavior risk early warning method

Abstract

The present application relates to the technical field of risk early warning, in particular to a risk early warning method for insurance sales behavior, in the present application, by collecting voice interaction and graphical interface operation data, a multi-modal behavior sequence is constructed and joint representation learning is carried out to obtain a unified joint feature representation; based on this feature, a future behavior sequence is predicted, and a trend deviation degree is generated by comparison with a safe sequence to realize risk forward-looking perception; at the same time, the support of voice and operation modal features to the trend deviation degree is respectively evaluated, and a comprehensive confidence is fused to generate to cooperatively verify the credibility of the risk signal; finally, a graded intervention instruction is generated according to the trend deviation degree and the comprehensive confidence to realize accurate and forward-looking intelligent prevention and control of insurance sales risk.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of risk early warning, in particular to an insurance sales behavior risk early warning method. BACKGROUND

[0002] In the process of digitization in the insurance sales field, compliance monitoring and risk early warning of sales behavior have become the focus of the industry. Existing technical solutions usually rely on independent analysis of single modal data (such as voice dialogue content or interface operation logs) to identify potential violations.

[0003] However, this single-path monitoring mode has obvious limitations. On the one hand, the manifestations of sales risks are often complex and dynamically changing, which may be reflected in both the tendency of language communication and the specific patterns of interface operation. It is difficult to fully capture the complete picture of risks from a single data source, which may lead to insufficient early warning sensitivity or misjudgment due to lack of information dimensions.

[0004] On the other hand, even if multiple sources of data are combined, traditional analysis methods are often limited to static matching or rule determination of already occurred behaviors, lacking the ability to predict the dynamic evolution trend of behavior sequences. This leads to risk intervention usually lagging behind the actual risk occurrence, making it difficult to achieve effective prevention in a forward-looking manner. In addition, existing methods often fail to fully consider the contribution difference of different modal features in risk representation and their dynamic correlation when fusing different modal information to jointly support risk decision-making, making the final risk assessment result have limited credibility and difficult to support precise and differentiated risk disposal strategies. SUMMARY

[0005] The present application relates to the technical field of risk early warning, in particular to an insurance sales behavior risk early warning method.

[0006] To achieve the above-mentioned purpose, the present application provides an insurance sales behavior risk early warning method, which specifically includes the following method steps:

[0007] S1, collect multi-modal behavior data in the insurance sales process, the multi-modal behavior data comprising voice interaction data and graphical interface operation data; wherein the voice interaction data is generated by real-time recording of dialogue audio between a salesperson and a customer through a voice collection device deployed in a sales scene; the graphical interface operation data is generated by recording operation events generated when a user interacts with a graphical interface through an operation monitoring module integrated in insurance sales software, the operation events comprising mouse click tracks, keyboard input content and interface element change logs.

[0008] This step S1 provides a comprehensive and synchronous multi-modal behavior data basis for constructing a risk early warning mechanism by collecting voice interaction data and graphical interface operation data in real time; its role is to convert dynamic and heterogeneous behavior information (such as dialogue content and interface operation) in the sales process into structured sequences that can be processed by a computing system, thereby laying a data-level premise for subsequent joint representation of multi-modal behavior sequences and future trend prediction, ensuring the coverage and real-time nature of the source of risk perception.

[0009] S2, joint representation learning of multi-modal behavior sequences is performed on the multi-modal behavior data to obtain a unified joint feature representation; wherein:

[0010] The construction process of the multi-modal behavior sequence specifically comprises:

[0011] The time-synchronized voice interaction data and graphical interface operation data are respectively input into a voice feature extraction network and an operation feature extraction network to obtain an initial feature vector sequence of the voice modality and an initial feature vector sequence of the operation modality;

[0012] The initial feature vector sequence of the voice modality and the initial feature vector sequence of the operation modality are spliced to form a multi-modal behavior sequence.

[0013] The generation process of the joint feature representation specifically comprises:

[0014] The multi-modal behavior sequence is input into a sequence encoding network based on a self-attention mechanism, the correlation weights between different time step features within the multi-modal behavior sequence are calculated to perform deep information fusion and context modeling on the multi-modal behavior sequence, and a unified joint feature representation is output.

[0015] The core role of this step S2 is to perform deep information fusion and context modeling on the spliced multi-modal behavior sequence through the sequence encoding network based on the self-attention mechanism to generate a unified joint feature representation; this process can effectively capture the internal correlation and mutual influence of voice and operation modalities in the time dimension, integrate behavior information from different sources into a high-order feature rich in context semantics, and provide a high-quality and integrated feature input for subsequent precise quantitative prediction of the trajectory and the degree of deviation from the safety benchmark.

[0016] S3, input the joint feature representation to a risk trajectory prediction model to obtain a behavior feature prediction sequence of future time steps, wherein the generation process of the behavior feature prediction sequence specifically comprises:

[0017] The risk trajectory prediction model adopts a sequence-to-sequence structure based on an encoder and a decoder;

[0018] The encoder of the risk trajectory prediction model encodes the input joint feature representation and converts it into an intermediate state vector containing historical behavior context information;

[0019] The decoder of the risk trajectory prediction model takes the intermediate state vector as the initial condition and recursively generates feature vectors of future consecutive time steps in a self-recurrent manner to output the behavior feature prediction sequence.

[0020] Comparing the behavior feature prediction sequence with the preset safe feature sequence generates a trend deviation, specifically comprising:

[0021] The dynamic time warping method is used to calculate the minimum alignment path distance between the behavior feature prediction sequence and the preset safe feature sequence, and the minimum alignment path distance is taken as the initial deviation;

[0022] The cosine similarity between the feature vectors of the behavior feature prediction sequence and the preset safe feature sequence at the corresponding time steps is calculated, and the arithmetic mean of all time step cosine similarities is taken as the average similarity;

[0023] The initial deviation and the difference obtained by subtracting the average similarity are weighted and summed, and the summation result is normalized to a value between zero and one as the trend deviation.

[0024] This step S3 uses the risk trajectory prediction model with a sequence-to-sequence structure to prospectively generate a behavior feature prediction sequence of future time steps based on the joint feature representation of historical behavior, and further generates a trend deviation by combining the comprehensive measurement method of dynamic time warping and cosine similarity; its effect is to realize the leap from static analysis to dynamic prediction, dynamically quantify the deviation degree of the prediction trajectory from the safety benchmark, and move the risk warning forward, providing direct and quantitative trend basis for the proactive intervention in the core target.

[0025] S4, based on the joint feature representation, respectively extract a speech modality feature subset and an operation modality feature subset; calculate a first support degree of the speech modality feature subset on the trend deviation, and calculate a second support degree of the operation modality feature subset on the trend deviation, wherein the calculation process of the first support degree specifically comprises:

[0026] The voice modality feature subset is input into a voice support degree evaluation network, a nonlinear mapping relationship between the voice modality feature subset and the trend deviation degree is calculated through a full connection layer in the voice support degree evaluation network, and a first support degree is output.

[0027] The second support degree calculation process specifically includes:

[0028] The operation modality feature subset is input into an operation support degree evaluation network, a nonlinear mapping relationship between the operation modality feature subset and the trend deviation degree is calculated through a full connection layer in the operation support degree evaluation network, and a second support degree is output.

[0029] According to the first support degree and the second support degree, a comprehensive confidence degree is calculated, specifically including:

[0030] The first support degree and the second support degree are weighted and summed, wherein the sum of the weight coefficients of the first support degree and the weight coefficients of the second support degree is one, and the weight coefficients are pre-set according to the stable contribution proportion of each modality feature subset to risk prediction in historical data;

[0031] The result of the weighted sum is input into an S-shaped function for mapping, and the output value is adjusted to a range of zero to one as the comprehensive confidence degree.

[0032] This step S4 aims to effectively fuse different modal evidences to cooperatively evaluate the risk confidence level mechanism; its role is to calculate the first support degree and the second support degree of the voice and operation modality feature subsets to the trend deviation degree respectively, and to generate a comprehensive confidence degree by weighted fusion according to its historical contribution; this process evaluates the support strength of different modal evidences to the current risk trend prediction, thereby calibrating the reliability of the risk signal indicated by the trend deviation degree, and providing a key reliability criterion for finally realizing more accurate graded early warning and intervention.

[0033] S5, according to the comprehensive confidence degree and the trend deviation degree, a graded intervention instruction is generated and corresponding risk disposal operation is executed, wherein the generation process of the graded intervention instruction specifically includes:

[0034] A grading strategy table is pre-set, which defines the risk levels corresponding to different numerical intervals of the comprehensive confidence degree and the trend deviation degree;

[0035] The currently calculated comprehensive confidence degree and trend deviation degree are input as a joint input, the grading strategy table is queried, and the corresponding risk level is matched;

[0036] According to the matched risk level, a graded intervention instruction bound to the risk level is generated.

[0037] The step S5 is a decision and execution terminal of risk disposal, which is used for matching a risk level and generating a graded intervention instruction by inquiring a preset strategy table based on the two key indexes of the comprehensive confidence and the trend deviation degree; the mechanism combines the trend prediction of the risk and the confidence level of the prediction, and determines the level and mode of the intervention, so that the single threshold alarm is upgraded to the multi-dimensional, accurate and forward graded early warning and intervention, and the pertinence, appropriateness and effectiveness of the risk disposal measures are ensured.

[0038] Compared with the prior art, the beneficial effects of the present application are:

[0039] The present application fuses multi-modal behavior data such as voice and operation to perform joint representation and context modeling, and constructs an analysis basis capable of deeply understanding the dynamic characteristics of the sales behavior; then, the sequence prediction technology is used to prospectively generate a future behavior trajectory, and the trend deviation degree between the predicted trajectory and a safety benchmark is quantified to realize early risk insight; further, the support of different modal evidences to the trend deviation degree is evaluated, and a comprehensive confidence is generated by fusion, so that a risk confidence level evaluation mechanism based on multi-modal evidence cooperation is established, and the reliability and accuracy of the risk judgment are effectively improved; finally, the graded intervention instruction is generated based on the dual decision basis of the trend deviation degree and the comprehensive confidence, so that the insurance sales behavior risk is from passive response to active, accurate and forward graded early warning and intervention, and the intelligent level and disposal efficiency of the risk management are significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 It is a whole method step schematic diagram of the present application.

[0041] Figure 2 It is a core flow chart of steps S3 and S4 of the present application. DETAILED DESCRIPTION

[0042] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0043] Next, please refer to Figure 1 The present embodiment is directed to an insurance sales behavior risk early warning method, which comprises the following steps:

[0044] S1, before the insurance sales process starts, the purpose, scope (including voice interaction data and graphical interface operation data) and use of data collection are clearly shown to the user through a graphical interface, and a data collection authorization agreement is presented, and only after the user's explicit authorization and consent is obtained, the data collection process is started.

[0045] During the collection process, the voice interaction data is generated by real-time recording the dialogue audio between the sales personnel and the customer through the voice collection device deployed in the sales scene; at the same time, the graphical interface operation data is generated by continuously recording the operation events generated by the user when interacting with the graphical interface within the authorized scope of the user through the operation monitoring module integrated in the insurance sales software, including mouse click track, keyboard input content and interface element change log; the voice interaction data and the graphical interface operation data jointly constitute the user-aware multi-modal behavior data obtained in the multi-modal behavior data collection process.

[0046] S2, multi-modal behavior sequence joint representation learning is performed on the multi-modal behavior data to obtain a unified joint feature representation, specifically including:

[0047] First, the time-synchronized voice interaction data and graphical interface operation data are input into the voice feature extraction network and the operation feature extraction network respectively to obtain the initial feature vector sequence of the voice modality and the initial feature vector sequence of the operation modality; wherein the voice feature extraction network is a neural network for processing voice interaction data, which receives voice signals and extracts features from the signals to output the initial feature vector sequence of the voice modality; the operation feature extraction network is a neural network for processing graphical interface operation data, which receives operation event data and extracts features from the operation to output the initial feature vector sequence of the operation modality;

[0048] Then, the initial feature vector sequence of the voice modality and the initial feature vector sequence of the operation modality are spliced, specifically, after alignment according to the time step, for each same time step, the initial feature vector sequence of the voice modality and the initial feature vector sequence of the operation modality are connected in the feature dimension to form a longer joint feature vector, thereby constituting a multi-modal behavior sequence.

[0049] Finally, the multimodal behavior sequence is input into a sequence encoding network based on a self-attention mechanism. By calculating the correlation weights between features at different time steps within the multimodal behavior sequence, deep information fusion and context modeling are performed on the multimodal behavior sequence, and a unified joint feature representation is output. The multimodal behavior sequence is input into a sequence encoding network based on a self-attention mechanism. This sequence encoding network calculates the correlation weights between the feature vector at each time step in the sequence and the feature vectors at all other time steps. Then, it performs a weighted summation of the feature vectors based on these weights, thereby fusing the deep information of the multimodal sequence and modeling the context dependencies, and finally outputting a unified joint feature representation.

[0050] Please see Figure 2 S3. Input the joint feature representation into the risk trajectory prediction model to obtain the behavioral feature prediction sequence for future time steps, specifically including:

[0051] The risk trajectory prediction model adopts a sequence-to-sequence structure based on encoder and decoder. First, the encoder part of the risk trajectory prediction model encodes the joint feature representation of the input and converts it into an intermediate state vector containing historical behavior context information.

[0052] Subsequently, the decoder part of the risk trajectory prediction model uses the intermediate state vector as the initial condition and recursively generates feature vectors for multiple consecutive future time steps in an autoregressive manner. The prediction for each time step is taken as the decoder output of the previous time step as the input for the current step, thereby outputting a future time step behavior feature prediction sequence with the same dimension as the historical behavior sequence.

[0053] The behavioral feature prediction sequence is compared with a preset security feature sequence to generate a trend deviation, which specifically includes:

[0054] The dynamic time warping method is used to calculate the minimum alignment path distance between the predicted behavioral feature sequence and the preset safety feature sequence, and this minimum alignment path distance is used as the initial deviation. Simultaneously, the cosine similarity between the feature vectors of the predicted behavioral feature sequence and the preset safety feature sequence at corresponding time steps is calculated, and the arithmetic mean of the cosine similarities at all time steps is taken as the average similarity. Finally, the initial deviation is weighted and summed with the difference between one and the average similarity, and the sum is normalized to a value between zero and one. This value is the generated trend deviation. The dynamic time warping method involves constructing a distance matrix between the predicted behavioral feature sequence and the preset safety feature sequence, and then finding a path from the upper left corner to the lower right corner of the matrix that minimizes the sum of feature distances between corresponding points on the path. This minimum path distance is used as the minimum alignment path distance.

[0055] S4, based on the joint feature representation, the specific process of extracting the speech modal feature subset and the operation modal feature subset respectively is:

[0056] Through a feature selection network, according to a predefined modal feature index, the dimensions corresponding to the speech modal are segmented from the joint feature representation to form a speech modal feature subset, and the dimensions corresponding to the operation modal are segmented to form an operation modal feature subset.

[0057] The specific process of calculating the first support degree of the speech modal feature subset to the trend deviation degree is:

[0058] The speech modal feature subset is input into a speech support evaluation network, which calculates the nonlinear mapping relationship between the speech modal feature subset and the trend deviation degree through a fully connected layer, and outputs the first support degree representing the contribution weight of the speech modal to the risk trend; wherein the speech support evaluation network is a fully connected neural network, and its function is to quantitatively evaluate the support degree of the speech modal feature to the finally calculated trend deviation degree; the network first receives the speech modal feature subset as input, which is a specific dimension vector segmented from the unified joint feature representation; then, the input feature is nonlinearly transformed through multiple fully connected layers inside the network, each layer introduces nonlinearity through its activation function, gradually learns and captures the complex and high-level correlation patterns between the speech modal feature and the trend deviation degree; finally, the network produces a scalar value through the calculation of the output layer, which is the first support degree, which accurately represents the contribution weight of the information contained in the speech interaction behavior to the formation of the risk trend in the context of the current sales conversation, and the higher the value represents the stronger the support of the speech modal.

[0059] The specific process of calculating the second support degree of the operation modal feature subset to the trend deviation degree is:

[0060] The operation mode feature subset is input to an operation support degree evaluation network, which calculates a nonlinear mapping relationship between the operation mode feature subset and the trend deviation degree through a fully connected layer, and outputs a second support degree representing the contribution weight of the operation mode to the risk trend; wherein the operation support degree evaluation network is a fully connected neural network with a specific structure, and its core task is to evaluate the support degree of the graphical interface operation behavior feature to the trend deviation degree; the network takes the operation mode feature subset as the processing object, and the feature subset is also extracted from the joint feature representation according to the predefined index; the operation support degree evaluation network performs multiple nonlinear mapping and feature integration on the input features through the fully connected layers stacked inside, and the hierarchical structure is designed to deeply analyze the internal causal relationship between the operation event sequence and the potential risk trend. After layer-by-layer transmission and calculation, the network generates a specific numerical output, i.e., the second support degree, at the output layer, which directly reflects the evidence strength or importance ratio provided by the operation mode (such as mouse click, keyboard input, and other interface interaction behaviors) to the judgment of the risk trend in the current sales process.

[0061] The specific process of calculating the comprehensive confidence degree according to the first support degree and the second support degree includes:

[0062] The first support degree and the second support degree are weighted and summed, wherein the sum of the weight coefficients of the first support degree and the weight coefficients of the second support degree is one, and wherein the weight coefficients are pre-set according to the stable contribution proportion of each modal feature subset to risk prediction in historical data; then the result of the weighted sum is input to an S-shaped function for mapping, and the output value is adjusted to a range of zero to one, and the final numerical value is the comprehensive confidence degree; wherein the S-shaped function is a nonlinear activation function that can smoothly and monotonically map input values of any range to the open interval between zero and one; when the weighted sum result passes through the function, the output value is compressed and normalized to a standard scale, and this final output numerical value is defined as the comprehensive confidence degree, which represents the overall confidence degree of the entire multi-modal system on the reliability of the current risk warning result, and the closer the numerical value is to one, the higher the confidence degree.

[0063] The specific process of generating a hierarchical intervention instruction and performing a corresponding risk disposal operation according to the comprehensive confidence degree and the trend deviation degree includes:

[0064] A hierarchical strategy table is pre-set, which defines the risk levels corresponding to different numerical intervals of the comprehensive confidence degree and the trend deviation degree; the current calculated comprehensive confidence degree and trend deviation degree are input as a joint input, and the hierarchical strategy table is queried to match to a unique corresponding risk level;

[0065] The hierarchical strategy table is a pre-defined and stored two-dimensional lookup table, the rows and columns of which correspond to the discretized numerical intervals of the comprehensive confidence and the trend deviation respectively; when constructing the hierarchical strategy table, a series of continuous and mutually non-overlapping numerical ranges are divided for the comprehensive confidence and the trend deviation according to historical risk case analysis and domain knowledge, and each numerical range is mapped to a specific risk level; when it is necessary to generate a hierarchical intervention instruction, the current calculated comprehensive confidence and trend deviation are taken as joint inputs, first, it is determined in which comprehensive confidence interval the current comprehensive confidence value falls in the hierarchical strategy table, and simultaneously, it is determined in which trend deviation interval the current trend deviation value falls in the hierarchical strategy table; by locating to this specific comprehensive confidence interval and trend deviation interval, a unique corresponding risk level is cross-referenced in the two-dimensional structure of the hierarchical strategy table;

[0066] The risk level is bound to a pre-set hierarchical intervention instruction; the specific hierarchical intervention instruction is generated immediately, and the risk disposal operation corresponding to the instruction is executed through the control interface thereof; the risk disposal operation corresponding to the hierarchical intervention instruction is executed through the control interface thereof, and the risk disposal operation includes sending a risk reminder label to the graphical interface of the sales personnel, starting a forced pop-up prompt to require confirmation and locking the sales process, etc.

[0067] The basic principles, main features and advantages of the present application are shown and described above. It should be understood by those skilled in the art that the present application is not limited by the above-mentioned embodiments, the above-mentioned embodiments and descriptions in the specification are only preferred examples of the present application, and are not intended to limit the present application, various changes and improvements of the present application can be made without departing from the spirit and scope of the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A risk warning method for insurance sales behavior, characterized in that, The methods and steps include the following: S1. Collect multimodal behavioral data during the insurance sales process, including voice interaction data and graphical interface operation data; S2. Perform joint representation learning on the multimodal behavior data to obtain a unified joint feature representation; S3. Input the joint feature representation into the risk trajectory prediction model to obtain the behavioral feature prediction sequence for future time steps; and compare the behavioral feature prediction sequence with a preset safety feature sequence to generate a trend deviation. S4. Based on the joint feature representation, extract the speech modal feature subset and the operational modal feature subset respectively; calculate the first support of the speech modal feature subset to the trend deviation, and calculate the second support of the operational modal feature subset to the trend deviation; calculate the comprehensive confidence based on the first support and the second support; S5. Based on the comprehensive confidence level and the trend deviation, generate tiered intervention instructions and execute corresponding risk management operations; The process of generating the trend deviation specifically includes: The minimum alignment path distance between the behavioral feature prediction sequence and the preset security feature sequence is calculated using the dynamic time warping method, and this minimum alignment path distance is used as the initial deviation. Calculate the cosine similarity between the feature vectors of the predicted behavioral feature sequence and the preset security feature sequence at the corresponding time step, and take the arithmetic mean of the cosine similarity of all time steps as the average similarity. The initial deviation is weighted and summed with the difference between one and the average similarity. The sum is then normalized to a value between zero and one, and this sum is taken as the trend deviation. The calculation process for the first support specifically includes: The subset of speech modal features is input into the speech support evaluation network. The nonlinear mapping relationship between the speech modal features and the trend deviation is calculated through the fully connected layer in the speech support evaluation network, and the first support is output. The calculation process for the second support specifically includes: The operational modality feature subset is input into the operational support evaluation network. The nonlinear mapping relationship between the operational modality feature subset and the trend deviation is calculated through the fully connected layer in the operational support evaluation network, and the second support is output.

2. The risk warning method for insurance sales behavior according to claim 1, characterized in that, The voice interaction data is generated by recording the audio of the conversation between sales personnel and customers in real time through voice acquisition devices deployed in the sales scenario; the graphical interface operation data is generated by recording operation events generated when users interact with the graphical interface through the operation monitoring module integrated in the insurance sales software. The operation events include mouse click trajectories, keyboard input content, and interface element change logs.

3. The risk warning method for insurance sales behavior according to claim 1, characterized in that, The process of constructing the multimodal behavior sequence specifically includes: Time-synchronized voice interaction data and graphical interface operation data are input into the voice feature extraction network and the operation feature extraction network, respectively, to obtain the initial feature vector sequence of the voice mode and the initial feature vector sequence of the operation mode. The initial feature vector sequence of the speech modality and the initial feature vector sequence of the operation modality are concatenated to form a multimodal behavior sequence.

4. The risk warning method for insurance sales behavior according to claim 3, characterized in that, The process of generating the joint feature representation specifically includes: Multimodal behavior sequences are input into a sequence encoding network based on a self-attention mechanism. By calculating the correlation weights between features at different time steps within the multimodal behavior sequence, deep information fusion and context modeling are performed on the multimodal behavior sequence, and a unified joint feature representation is output.

5. The risk warning method for insurance sales behavior according to claim 1, characterized in that, The process of generating the behavioral feature prediction sequence specifically includes: The risk trajectory prediction model adopts a sequence-to-sequence structure based on an encoder and decoder; The encoder of the risk trajectory prediction model encodes the joint feature representation of the input and converts it into an intermediate state vector containing historical behavioral context information. The decoder of the risk trajectory prediction model uses the intermediate state vector as an initial condition to recursively generate feature vectors for multiple consecutive future time steps in an autoregressive manner, and outputs a behavioral feature prediction sequence.

6. The risk warning method for insurance sales behavior according to claim 1, characterized in that, The calculation process of the overall confidence level specifically includes: The first support and the second support are weighted and summed, where the weight coefficient of the first support and the weight coefficient of the second support are summed to one. The weight coefficients are preset according to the stable contribution ratio of each modal feature subset to risk prediction in historical data. The weighted summation result is input into a sigmoid function for mapping, and the output value is adjusted to the range between zero and one, which is used as the overall confidence level.

7. The risk warning method for insurance sales behavior according to claim 1, characterized in that, The process of generating the tiered intervention instructions specifically includes: A pre-defined grading strategy table defines the risk level corresponding to the comprehensive confidence level and trend deviation for different numerical ranges. Using the calculated overall confidence level and trend deviation as joint inputs, query the tiered strategy table and match the corresponding risk level. Based on the matched risk level, a tiered intervention instruction is generated that is linked to that risk level.

Images