Image auditing method and system based on machine learning and drools engine

Abstract

The present application relates to the field of insurance claim auditing, and proposes an image auditing method and system based on machine learning and Drools engine, comprising: obtaining an insurance claim image to be audited and extracting an image feature vector; calculating an insurance fraud probability value through a forest algorithm model; dynamically calculating a probability threshold value based on business attribute information; comparing the insurance fraud probability value with the probability threshold value, and selecting an insurance business rule chain according to the comparison result; obtaining the current state of the insurance claim process, dynamically reconstructing the insurance business rule chain, and obtaining a reconstructed insurance business rule chain; outputting a rule decision result through a Drools rule engine; performing decision fusion on the probability value and the rule result, using a decision trajectory tracking mechanism to resolve conflicts, and obtaining a final auditing decision result; and outputting the decision result and recording audit trail information. The present application improves the accuracy and reliability of insurance claim image auditing, and meets regulatory compliance requirements.

Description

Technical Field

[0001] This invention relates to the field of insurance claims review technology, and in particular to an image review method and system based on machine learning and the Drools engine. Background Technology

[0002] Image verification for insurance claims is a core component of the insurance industry's risk control system, involving the verification of the authenticity and fraud detection of images submitted by customers as proof of claims. Insurance claim images typically include various types such as accident scene photos, damage evidence images, repair invoices, and medical bills, characterized by their large quantity, diverse formats, and complex content. With the rapid growth of the insurance business and the continuous innovation of fraud methods, traditional manual verification methods can no longer meet the demands of large-scale, high-efficiency, and high-accuracy operations. There is an urgent need for intelligent image verification technology to improve verification efficiency and reduce fraud risk.

[0003] In existing technologies, insurance claim image verification mainly employs traditional image recognition technology and static rule engine systems to achieve basic fraud risk identification. However, existing methods do not adequately consider the decision fusion mechanism between the probability values ​​output by machine learning models and the execution results of rule engines. This makes it difficult to organically integrate probabilistic predictions with rule-based judgments, resulting in the inability to resolve conflicts between the two decision-making methods and thus affecting the accuracy and reliability of the verification results. Summary of the Invention

[0004] In view of this, the present invention proposes an image review method and system based on machine learning and the Drools engine. This solves the problem that the existing technology does not adequately consider the decision fusion mechanism between the probability values ​​output by the machine learning model and the execution results of the rule engine, making it difficult to organically integrate probabilistic prediction and rule judgment, thus failing to resolve the conflict between the two decision-making methods and affecting the accuracy and reliability of the review results.

[0005] The technical solution of this invention is implemented as follows: On one hand, this invention provides an image review method based on machine learning and the Drools engine, including the following steps:

[0006] Obtain the insurance claim image to be reviewed, and extract the image feature vector of the insurance claim image;

[0007] The image feature vector is input into the forest algorithm model to calculate the insurance fraud probability value;

[0008] Obtain business attribute information of insurance claim cases, and dynamically calculate probability thresholds based on the business attribute information;

[0009] The insurance fraud probability value is compared with the probability threshold, and the corresponding insurance business rule chain in the Drools rule engine is selected based on the comparison result.

[0010] Obtain the current status of the insurance claims process, and dynamically reconstruct the insurance business rule chain based on the current status to obtain the reconstructed insurance business rule chain;

[0011] The reconstructed insurance business rule chain is evaluated using the Drools rule engine, and the rule decision results are output.

[0012] The insurance fraud probability value and the rule decision result are fused together, and a decision trajectory tracing mechanism is used to resolve conflicts to obtain the final review decision result.

[0013] Output the final audit decision result and record the audit trajectory information corresponding to the final audit decision result.

[0014] Based on the above technical solutions, preferably, the step of obtaining business attribute information of insurance claim cases and dynamically calculating probability thresholds based on the business attribute information includes:

[0015] The business attribute information includes the claim amount, policy lifecycle, and time period risk characteristics. Quantitative feature values ​​corresponding to the claim amount, policy lifecycle, and time period risk characteristics of the insurance claim case are extracted.

[0016] Establish a multi-dimensional mapping relationship between business attribute feature values ​​and probability thresholds, and perform fusion calculation on the quantified feature values ​​based on the multi-dimensional mapping relationship to dynamically generate probability thresholds that are suitable for the characteristics of the current insurance claim case.

[0017] Based on the above technical solutions, preferably, the step of establishing a multi-dimensional mapping relationship between business attribute feature values ​​and probability thresholds, and performing fusion calculations on the quantified feature values ​​according to the multi-dimensional mapping relationship, includes:

[0018] Construct a three-dimensional weight matrix of claim amount risk coefficient, policy period decay factor and time period risk volatility index;

[0019] The quantitative feature values ​​are weighted by the corresponding risk coefficient, decay factor and volatility index respectively, and the dynamic probability threshold adjustment parameters are calculated by nonlinear fusion function. A probability threshold is generated based on the dynamic probability threshold adjustment parameters.

[0020] Based on the above technical solution, preferably, the step of comparing the insurance fraud probability value with the probability threshold and selecting the corresponding insurance business rule chain in the Drools rule engine according to the comparison result includes:

[0021] Calculate the confidence index of the insurance fraud probability value, combine it with the feature stability score of the image feature vector, and perform confidence correction on the insurance fraud probability value to obtain the corrected insurance fraud probability value.

[0022] Based on the deviation between the corrected insurance fraud probability value and the probability threshold, a probability-threshold deviation vector is constructed, and the most matching insurance business rule chain is adaptively selected from the Drools rule engine according to the probability-threshold deviation vector.

[0023] Based on the above technical solutions, preferably, the step of constructing a probability-threshold deviation vector based on the deviation between the corrected insurance fraud probability value and the probability threshold, and adaptively selecting the most matching insurance business rule chain according to the probability-threshold deviation vector, includes:

[0024] The numerical deviation, trend deviation, and distribution interval deviation of the corrected insurance fraud probability value from the probability threshold are calculated to form a three-dimensional deviation vector;

[0025] The three-dimensional deviation vector is matched with the pre-stored rule chain feature vector library for vector similarity. The insurance business rule chain corresponding to the rule chain feature vector with the highest similarity is selected, and the execution weight of each rule node in the rule chain is dynamically adjusted according to the similarity value.

[0026] Based on the above technical solutions, preferably, the step of obtaining the current state of the insurance claims process and dynamically reconstructing the insurance business rule chain according to the current state to obtain the reconstructed insurance business rule chain includes:

[0027] Construct a multi-dimensional feature description model of the insurance claims process status, and predict the evolution path of the subsequent process status based on the current status features;

[0028] The insurance business rule chain is reconstructed in a forward-looking manner according to the evolution path. The forward-looking reconstruction includes adjusting the topology of the rule chain and redistributing the execution priorities among the rule nodes to obtain a reconstructed insurance business rule chain that adapts to the state evolution.

[0029] Based on the above technical solutions, preferably, the step of forward-lookingly reconstructing the insurance business rule chain according to the evolution path includes:

[0030] Based on the key state nodes in the evolution path, identify the combination of rule nodes in the insurance business rule chain that generate execution conflicts or efficiency bottlenecks.

[0031] The identified rule node combinations are decoupled and recombined. By constructing a dynamic dependency graph between rule nodes, the execution sequence and branch conditions of the insurance business rule chain are redesigned. Each reconstructed rule branch is assigned an execution weight proportional to the state evolution probability, thus obtaining the reconstructed insurance business rule chain.

[0032] Based on the above technical solutions, preferably, the step of fusing the insurance fraud probability value and the rule decision result, and using a decision trajectory tracing mechanism to resolve conflicts, to obtain the final review decision result includes:

[0033] Construct a cross-validation confidence matrix between the insurance fraud probability value and the rule decision result, and identify decision conflict points based on the confidence difference;

[0034] A full-link decision trajectory map is established, which includes feature extraction, probability calculation, rule triggering, and decision generation. The source of conflict is located by trajectory backtracking, and a multi-level conflict resolution strategy based on evidence strength weight is adopted to generate the final review decision result.

[0035] Based on the above technical solutions, preferably, the establishment of a full-link decision trajectory map, which includes feature extraction, probability calculation, rule triggering, and decision generation, and the location of conflict sources through trajectory backtracking, includes:

[0036] Assign a unique trajectory node identifier to each decision-making stage, record the causal transmission relationship and information flow between trajectory nodes, and construct a directed acyclic full-link decision trajectory map;

[0037] When a decision conflict is detected, the entire decision trajectory map is traversed in reverse from the conflict point. The contribution of each trajectory node to the conflict is calculated, the trajectory node with the highest contribution is identified as the source of the conflict, and the input data and processing parameters of the source of the conflict are extracted for conflict analysis.

[0038] On the other hand, the present invention also provides an image review system based on machine learning and the Drools engine, the system comprising:

[0039] The image feature extraction module is used to acquire the insurance claim image to be reviewed and extract the image feature vector of the insurance claim image;

[0040] The fraud probability calculation module is used to input the image feature vector into the forest algorithm model to calculate the insurance fraud probability value;

[0041] The dynamic threshold calculation module is used to obtain business attribute information of insurance claim cases and dynamically calculate probability thresholds based on the business attribute information.

[0042] The rule chain selection module is used to compare the insurance fraud probability value with the probability threshold, and select the corresponding insurance business rule chain in the Drools rule engine according to the comparison result.

[0043] The rule chain dynamic reconstruction module is used to obtain the current state of the insurance claims process and dynamically reconstruct the insurance business rule chain based on the current state to obtain the reconstructed insurance business rule chain.

[0044] The rule engine execution module is used to perform Drools rule engine judgments on the reconstructed insurance business rule chain and output the rule decision results;

[0045] The decision fusion and conflict resolution module is used to fuse the insurance fraud probability value and the rule decision result, and use a decision trajectory tracing mechanism to resolve conflicts to obtain the final review decision result.

[0046] The decision output and review module is used to output the final review decision result and record the audit trajectory information corresponding to the final review decision result.

[0047] The image review method and system based on machine learning and the Drools engine of this invention have the following advantages over the prior art:

[0048] (1) By integrating machine learning models with the Drools rule engine, using forest algorithm models to perform image feature analysis and fraud probability calculation, dynamically adjusting probability thresholds based on business attribute information, and dynamically reconstructing the rule chain according to the claims process status, the conflict between machine learning results and rule decisions is resolved through the decision trajectory tracing mechanism, thereby improving the accuracy and reliability of insurance claims image review, and meeting regulatory compliance requirements through complete audit trajectory records.

[0049] (2) The credibility of the fraud probability value is corrected by confidence index and feature stability score, and the probability-threshold deviation vector is constructed to quantify the degree of deviation. This realizes the rule chain selection from single numerical comparison to multi-dimensional vector matching, which improves the accuracy and adaptability of rule chain selection and ensures that the selected rule chain is most matched with the risk characteristics of the current claim case.

[0050] (3) By constructing a multi-dimensional feature description model of the insurance claims process status, the evolution path of the subsequent process status is predicted. Based on the key status nodes, the execution conflict and efficiency bottleneck in the rule chain are identified. The execution sequence and branch conditions of the rule chain are redesigned using dynamic dependency graphs. This achieves a forward-looking reconstruction from passive response to active prediction, improving the adaptability and execution efficiency of the rule chain to changes in process status.

[0051] (4) By constructing a cross-validation confidence matrix of insurance fraud probability value and rule decision results, a full-link decision trajectory map is established. The trajectory backtracking mechanism is used to accurately locate the source of conflict and quantify the conflict contribution of each node. Based on the evidence strength weight, a multi-level conflict resolution strategy is implemented, realizing a technological leap from traditional simple decision merging to intelligent conflict analysis and resolution, improving the scientific nature of decision fusion and the reliability of the final review results. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a flowchart of an image review method based on machine learning and the Drools engine according to the present invention. Detailed Implementation

[0054] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0055] Please see Figure 1 This invention provides an image review method based on machine learning and the Drools engine, comprising the following steps:

[0056] Obtain the insurance claim image to be reviewed, and extract the image feature vector of the insurance claim image;

[0057] The image feature vector is input into the forest algorithm model to calculate the insurance fraud probability value;

[0058] Obtain business attribute information of insurance claim cases, and dynamically calculate probability thresholds based on the business attribute information;

[0059] The insurance fraud probability value is compared with the probability threshold, and the corresponding insurance business rule chain in the Drools rule engine is selected based on the comparison result.

[0060] Obtain the current status of the insurance claims process, and dynamically reconstruct the insurance business rule chain based on the current status to obtain the reconstructed insurance business rule chain;

[0061] The reconstructed insurance business rule chain is evaluated using the Drools rule engine, and the rule decision results are output.

[0062] The insurance fraud probability value and the rule decision result are fused together, and a decision trajectory tracing mechanism is used to resolve conflicts to obtain the final review decision result.

[0063] Output the final audit decision result and record the audit trajectory information corresponding to the final audit decision result.

[0064] Specifically, this embodiment integrates machine learning models with the Drools rule engine, uses a forest algorithm model for image feature analysis and fraud probability calculation, dynamically adjusts the probability threshold based on business attribute information, and dynamically reconstructs the rule chain according to the claims process status. The decision trajectory tracing mechanism solves the conflict between machine learning results and rule decisions, improving the accuracy and reliability of insurance claims image review. At the same time, the complete audit trajectory record meets regulatory compliance requirements.

[0065] The step of obtaining the insurance claim image to be reviewed and extracting the image feature vector of the insurance claim image includes:

[0066] The insurance claim image is preprocessed to obtain a standardized insurance claim image.

[0067] In one specific embodiment, the image type identifier of the insurance claim image is obtained, and the image type identifier includes document images, on-site photo images, and receipt images;

[0068] Based on the image type identifier, a corresponding preprocessing strategy is selected to perform size normalization, illumination balancing, and noise removal on the insurance claim image to obtain the standardized insurance claim image.

[0069] Multi-dimensional image features are extracted from the standardized insurance claim image and combined to obtain the image feature vector.

[0070] In one specific embodiment, texture features, edge features, and region features are extracted from the standardized insurance claim image, respectively;

[0071] The texture features, edge features, and region features are weighted and fused according to preset weights to obtain the image feature vector.

[0072] Specifically, this embodiment establishes a mapping relationship between image type identifiers and preprocessing strategies, and adopts customized size normalization, illumination balancing, and noise removal strategies for document images, on-site photos, and ticket images, respectively. It also extracts texture features, edge features, and region features from standardized images and performs weighted fusion, thereby realizing a shift from coarse image processing to refined feature extraction, improving the quality of image feature vectors, and providing a more accurate and reliable data foundation for fraud probability calculation.

[0073] The step of inputting the image feature vector into the forest algorithm model to calculate the insurance fraud probability value includes:

[0074] A primitive forest algorithm model is constructed, and the primitive forest algorithm model is trained based on insurance fraud sample data to obtain a trained forest algorithm model.

[0075] In one specific embodiment, multiple decision tree structures are initialized and combined to form the original forest algorithm model;

[0076] Historical fraud and normal samples of insurance claim images are collected to construct a training dataset. The training dataset is then used to learn the parameters of multiple decision tree structures in the original forest algorithm model, resulting in a trained forest algorithm model.

[0077] The image feature vector is input into the trained forest algorithm model for prediction calculation, and the insurance fraud probability value is output.

[0078] In one specific embodiment, the image feature vector is input into each decision tree in the trained forest algorithm model to obtain multiple decision tree prediction results;

[0079] The prediction results of the multiple decision trees are weighted and voted on to calculate the insurance fraud probability value.

[0080] Specifically, this embodiment constructs a primitive forest algorithm model and trains parameters based on historical fraud samples and normal samples. It adopts a weighted voting processing mechanism with multiple decision tree structures, achieving progress from a single classifier to ensemble learning, improving the accuracy of insurance fraud probability prediction, and solving the problems of overfitting and insufficient generalization ability of a single model.

[0081] The process of obtaining business attribute information of insurance claim cases and dynamically calculating probability thresholds based on the business attribute information includes:

[0082] The business attribute information includes the claim amount, policy lifecycle, and time period risk characteristics. Quantitative feature values ​​corresponding to the claim amount, policy lifecycle, and time period risk characteristics of the insurance claim case are extracted.

[0083] Establish a multi-dimensional mapping relationship between business attribute feature values ​​and probability thresholds, and perform fusion calculation on the quantified feature values ​​based on the multi-dimensional mapping relationship to dynamically generate probability thresholds that are suitable for the characteristics of the current insurance claim case.

[0084] The step of establishing a multi-dimensional mapping relationship between business attribute feature values ​​and probability thresholds, and performing fusion calculations on the quantized feature values ​​based on the multi-dimensional mapping relationship, includes:

[0085] Construct a three-dimensional weight matrix of claim amount risk coefficient, policy period decay factor and time period risk volatility index;

[0086] The quantitative feature values ​​are weighted by the corresponding risk coefficient, decay factor and volatility index respectively, and the dynamic probability threshold adjustment parameters are calculated by nonlinear fusion function. A probability threshold is generated based on the dynamic probability threshold adjustment parameters.

[0087] In one specific embodiment, the formula for calculating the dynamic probability threshold adjustment parameter is:

[0088] ;

[0089] in, Adjust parameters for dynamic probability thresholds; It is the main adjustment coefficient, used to control the overall adjustment range; This is the current claim amount; This is the risk coefficient for claim amount, used to reflect the risk weight of different amount ranges; The largest historical claim amount is used for numerical normalization. The policy period impact coefficient; The time decay constant; The policy lifespan (e.g., in years); This is the time-cycle risk volatility intensity coefficient; The number of days in the year for the current date (e.g., February 2nd is the 33rd day); This is the phase offset parameter for time-cycle risk, used to reflect the time offset of time-cycle risk characteristics in different regions.

[0090] The probability threshold is calculated as follows:

[0091] ;

[0092] in, This is the final dynamic probability threshold; Basic probability threshold; Adjust parameters for dynamic probability thresholds; For the first The fusion weight of business attributes; For the first dimensional quantized eigenvalues ​​( Corresponding to the characteristics of the claim amount, Corresponding to the characteristics of the policy period, Corresponding time cycle risk characteristics); For the first Historical mean of the dimensional feature; For the first Historical standard deviation of dimensional features; For the first The nonlinear power exponent of the dimensional feature is used to adjust the influence strength of each dimensional feature; It is the hyperbolic tangent activation function.

[0093] In one specific embodiment, the dynamic threshold calculation for the fast auto insurance claims scenario includes:

[0094] The following specific parameters and implementation steps are used in a certain auto insurance fast claims review system:

[0095] First, extract business attribute information;

[0096] Claim Amount Feature Extraction: Given a current auto insurance claim amount of 150,000 yuan, the system extracts historical claim amount distribution data from the past year and calculates... , (Highest historical claim amount) (Risk coefficient for medium-sized amounts).

[0097] Policy lifecycle feature extraction: The policy has been in effect for 2.5 years, and the system calculates... Based on the analysis of historical data, the conclusion is as follows (Policy Cycle Impact Coefficient) (Time decay constant).

[0098] Time-cycle risk feature extraction: The case occurred on June 15th (the 166th day of the year). Based on the analysis of historical fraud cases, the following settings were established: , (Time-cycle risk volatility intensity coefficient at this time) (Regional phase shift).

[0099] Secondly, the dynamic probability threshold adjustment parameters are calculated;

[0100] An improved hyperbolic tangent fusion function is used:

[0101] ;

[0102] Calculation results: .

[0103] Finally, the final dynamic probability threshold is generated;

[0104] Base threshold The weights for each dimension are: , , ;

[0105] ;

[0106] The final calculation yielded The value of .

[0107] Specifically, this implementation introduces the hyperbolic tangent activation function as the core of nonlinear transformation, which can effectively handle the saturation effect and boundary constraints of business attributes; by adopting a multiplicative weight fusion mechanism, the features of each dimension can be modulated to avoid the excessive dominance of a single dimension in decision-making; and by using an exponential adjustment mechanism, the importance of each dimension under different business scenarios can be precisely controlled.

[0108] This embodiment constructs a multi-scale time modeling framework based on the temporal evolution of insurance fraud and the characteristics of time cycle risk fluctuations.

[0109] Using the exponential decay model The nonlinear impact of the policy lifecycle on fraud probability is captured using a sinusoidal periodic function. Accurately model time-cycle risk characteristics; through phase offset parameters Consider the differences in risk over time periods across different regions and types of insurance.

[0110] By adopting Dynamic normalization is performed to avoid interference from extreme value data in threshold calculation; historical statistical features are introduced. and To achieve adaptive updates of feature standardization, a combination of Z-score standardization and hyperbolic tangent transform is used to ensure numerical stability.

[0111] The step of comparing the insurance fraud probability value with the probability threshold and selecting the corresponding insurance business rule chain in the Drools rule engine based on the comparison result includes:

[0112] Calculate the confidence index of the insurance fraud probability value, combine it with the feature stability score of the image feature vector, and perform confidence correction on the insurance fraud probability value to obtain the corrected insurance fraud probability value.

[0113] Based on the deviation between the corrected insurance fraud probability value and the probability threshold, a probability-threshold deviation vector is constructed, and the most matching insurance business rule chain is adaptively selected from the Drools rule engine according to the probability-threshold deviation vector.

[0114] The step of constructing a probability-threshold deviation vector based on the deviation between the corrected insurance fraud probability value and the probability threshold, and adaptively selecting the best-matching insurance business rule chain based on the probability-threshold deviation vector, includes:

[0115] The numerical deviation, trend deviation, and distribution interval deviation of the corrected insurance fraud probability value from the probability threshold are calculated to form a three-dimensional deviation vector;

[0116] The three-dimensional deviation vector is matched with the pre-stored rule chain feature vector library for vector similarity. The insurance business rule chain corresponding to the rule chain feature vector with the highest similarity is selected, and the execution weight of each rule node in the rule chain is dynamically adjusted according to the similarity value.

[0117] In one specific embodiment, the formula for calculating the corrected insurance fraud probability value is:

[0118] ;

[0119] in, To correct the probability value of insurance fraud; This represents the original insurance fraud probability value. This is a confidence level index, with a value range of... ; The feature stability score for the image feature vector is given, with a value range of [value range missing]. ; This is the confidence level correction index, used to control the correction strength; This is a correction coefficient for the number of features; The feature quantity decay parameter; The number of valid features.

[0120] The similarity calculation formula for the three-dimensional deviation vector and vector similarity matching is as follows:

[0121] ;

[0122] ;

[0123] in, It is a three-dimensional deviation vector; For the first Feature vectors of a pre-stored rule chain; To correct for the probability value of insurance fraud, This is the final dynamic probability threshold; and These are the 95th and 5th percentiles of the probability value, used to reflect the distribution interval; and These are the 75th and 25th percentiles of the threshold, respectively. This is the distance penalty coefficient, used to balance angular similarity and Euclidean distance; Three-dimensional deviation vector With the Feature vectors of pre-stored rule chains The similarity between them.

[0124] In one specific embodiment, the rule chain for intelligent selection in auto insurance collision claims includes the following case background:

[0125] In a car insurance claim case involving a single-vehicle collision, the claim amount was 120,000 yuan. Evidence materials such as photos of the accident scene, photos of vehicle damage, and repair invoices were uploaded.

[0126] The specific processing steps are as follows:

[0127] The first step is to correct the credibility of the insurance fraud probability value;

[0128] The raw data includes: the original insurance fraud probability value. Confidence index is (The model has a high level of confidence in the prediction result); the feature stability score is... (Image features are relatively stable); the number of effective features is (15 effective features were extracted from the image).

[0129] Parameter settings include: confidence correction index is The feature quantity correction coefficient is: The feature quantity decay parameter is: .

[0130] The correction calculation process includes:

[0131] ;

[0132] The corrected insurance fraud probability value decreased to 0.530, reflecting the corrective effect of confidence level and feature stability on the original probability.

[0133] The second step is to construct a three-dimensional deviation vector;

[0134] Known parameters include: the corrected insurance fraud probability value. Assume the dynamic probability threshold is The probability quantile is , Threshold quantiles are , .

[0135] Perform time series derivative calculations;

[0136] Based on the probability change trend over the past 5 minutes:

[0137] (The probability is decreasing);

[0138] (The threshold increased slightly.)

[0139] Then construct the three-dimensional deviation vector:

[0140] .

[0141] The third step is vector similarity matching and rule chain selection.

[0142] The pre-stored rule chain feature vector library includes:

[0143] Rule Chain 1 (Low-Risk Fast Track): ;

[0144] Rule Chain 2 (Medium Risk Standard Review): ;

[0145] Rule Chain 3 (High Risk, Strict Audit): .

[0146] Similarity calculation (taking rule chain 2 as an example):

[0147] Set distance penalty coefficient ;

[0148]

[0149] ;

[0150] Other rule chain similarities:

[0151] ;

[0152] .

[0153] Rule Chain 2 (medium-risk standard audit) was selected as the execution chain because it had the highest similarity.

[0154] The fourth step is dynamic weight adjustment;

[0155] Based on a similarity score of 0.837, the weights of each rule node in rule chain 2 are adjusted. It is assumed that rule chain 2 includes rules for verifying the authenticity of images, checking the reasonableness of amounts, and reviewing timeliness.

[0156] Image authenticity verification rules: Original weight 0.3 → adjusted to ;

[0157] Reasonableness check rules for amounts: Original weight 0.4 → adjusted to ;

[0158] Timeliness review rules: Original weight 0.3 → adjusted to .

[0159] Existing technologies typically use the probability values ​​output by machine learning models directly, ignoring the uncertainty of model predictions and the differences in the quality of input data.

[0160] This embodiment employs dual correction based on confidence and stability, considering both model confidence and feature stability to avoid over-reliance on a single metric; it also uses adaptive adjustment of the number of features through an exponential decay function. The non-linear impact of the number of modeling features on reliability; utilizing a power function correction mechanism, through... The parameters enable precise control of the correction intensity, adapting to different business scenarios.

[0161] This embodiment transforms the probability-threshold comparison problem into a similarity matching problem in a high-dimensional vector space, upgrading it from a one-dimensional scalar comparison to a multi-dimensional feature matching problem. Numerical deviation directly reflects the relative difference between the probability value and the threshold; trend deviation captures dynamic evolution characteristics and predicts future deviation directions; and distribution interval deviation, based on quantile statistics, reflects differences in data distribution characteristics.

[0162] This embodiment combines cosine similarity (angular matching) and Euclidean distance (numerical matching), balancing the two metrics through an exponential penalty function. Cosine similarity focuses on the consistency of vector directions, while the distance penalty term ensures that numerical differences are not too large. The parameters are balanced by similarity in angle and numerical proximity. Based on the similarity values, the weights of each node in the selected rule chain are fine-tuned to achieve precise control over the strength of rule execution.

[0163] The process of obtaining the current state of the insurance claims process, and dynamically reconstructing the insurance business rule chain based on the current state, to obtain a reconstructed insurance business rule chain, includes:

[0164] Construct a multi-dimensional feature description model of the insurance claims process status, and predict the evolution path of the subsequent process status based on the current status features;

[0165] The insurance business rule chain is reconstructed in a forward-looking manner according to the evolution path. The forward-looking reconstruction includes adjusting the topology of the rule chain and redistributing the execution priorities among the rule nodes to obtain a reconstructed insurance business rule chain that adapts to the state evolution.

[0166] The forward-looking reconstruction of the insurance business rule chain based on the evolution path includes:

[0167] Based on the key state nodes in the evolution path, identify the combination of rule nodes in the insurance business rule chain that generate execution conflicts or efficiency bottlenecks.

[0168] The identified rule node combinations are decoupled and recombined. By constructing a dynamic dependency graph between rule nodes, the execution sequence and branch conditions of the insurance business rule chain are redesigned. Each reconstructed rule branch is assigned an execution weight proportional to the state evolution probability, thus obtaining the reconstructed insurance business rule chain.

[0169] In one specific embodiment, the prediction formula for the evolutionary path is:

[0170] ;

[0171] in, For the future The process state vector of the step; This represents the current process state vector; For the first The state transition matrix of the step; For the first The external input influence matrix of the step; This represents a vector of external environmental factors (such as policy changes, market environment, etc.). This is a time decay factor, reflecting the increased uncertainty in long-term forecasts.

[0172] The formula for allocating the execution weights is as follows:

[0173] ;

[0174] in, The execution weight after the rule node is reconstructed; This represents the original execution weight of the rule node; This is the evolution probability weighting adjustment coefficient; This represents the activation probability of the node during state evolution. For this node and its dependent nodes The coupling strength; This is the coupling penalty coefficient; The dependency complexity index; This is the set of dependencies for the current rule node.

[0175] Specifically, existing insurance claims systems mostly use static rule configurations, which cannot be dynamically adjusted according to changes in the state during process execution, resulting in low rule execution efficiency and an inability to adapt to complex and ever-changing business scenarios; insufficient predictive ability for process state transitions and a lack of forward-looking decision support; and simplistic handling of dependencies between rules, which easily leads to execution conflicts and resource waste.

[0176] This embodiment abstracts the complex state of the insurance claims process into state points in a high-dimensional vector space, with each dimension representing a key attribute of the process (such as case complexity, evidence completeness, risk level, processing time, etc.).

[0177] By constructing state vectors Each component Indicates the first Each state attribute at time... The quantized value. This embodiment realizes the transformation from discrete state to continuous state space, improving the granularity of state description; and supports unified modeling and processing of multi-dimensional state information.

[0178] This embodiment is based on Markov chain theory and incorporates the influence of external environmental factors to construct a multi-step state transition prediction model.

[0179] Through the summation term Capture the cumulative effect of multi-step transfers, rather than a simple one-step transfer; introduce The system takes into account the impact of external factors such as policy changes and market environment on the process status. This reflects the objective law of increasing uncertainty in long-term forecasts, thus improving the reliability of forecasts; The variation in the number of steps reflects the differences in the state transition patterns within different prediction periods.

[0180] This embodiment models the dependencies between rule nodes as a directed acyclic graph (DAG) and analyzes dependency paths and influence propagation mechanisms using graph theory algorithms. Based on state evolution prediction results, the topology of the dependency graph is adjusted in real time; weights are assigned to dependency edges to quantify the strength and importance of dependencies; and a topological sorting algorithm is used to detect and resolve potential dependency cycles.

[0181] This embodiment uses a non-linear weight allocation mechanism. Ensure the relative rationality of weight allocation; through Weights are amplified or reduced based on node activation probabilities. The coupling effects of all dependent nodes are cumulatively calculated; utilizing... Nonlinear penalty for coupling strength is achieved through negative exponent; based on The penalty intensity is dynamically adjusted based on the complexity of the dependency relationship.

[0182] This embodiment, based on state evolution path analysis, identifies rule node combinations that may cause execution conflicts in advance. It traverses all possible execution paths using a depth-first search; identifies rule node combinations that simultaneously access the same resources; and detects rule node pairs with inconsistent execution timing requirements. By decomposing complex rule chains into loosely coupled functional modules, identifying rule nodes that can be executed in parallel, it improves overall execution efficiency; and it selects the optimal execution path based on real-time status.

[0183] The step of performing Drools rule engine judgment on the reconstructed insurance business rule chain and outputting rule decision results includes:

[0184] The reconstructed insurance business rule chain is executed and scheduled, including parallel decomposition based on rule complexity and dynamic execution priority adjustment based on resource load.

[0185] In one specific embodiment, the reconstructed insurance business rule chain is decomposed into multiple parallel executable rule sub-chains according to the dependencies between rules, and an independent Drools engine execution thread is allocated to each rule sub-chain;

[0186] The system monitors the resource consumption and execution progress of each execution thread in real time. When an execution bottleneck is detected, the system dynamically adjusts the execution priority of the rule sub-chain and optimizes the overall execution performance through rule prefetching and result caching mechanisms.

[0187] Based on the intermediate states and confidence indices of rule execution, a multi-level rule decision result is constructed, which includes the main decision, alternative decisions, and decision uncertainty.

[0188] In one specific embodiment, the execution status, rule triggering frequency, and execution time of each rule subchain are collected, and the decision confidence of each rule subchain is calculated.

[0189] Based on the decision confidence level, a hierarchical structure of primary decision and alternative decision is constructed, wherein the primary decision is the rule output with the highest confidence level, and the alternative decision is the rule output with the second highest confidence level. The difference in confidence level between the primary decision and the alternative decision is calculated as a decision uncertainty index to obtain the multi-level rule decision result.

[0190] Specifically, this embodiment constructs a parallel decomposition mechanism based on rule complexity and a dynamic execution priority adjustment strategy based on resource load. It splits the reconstructed rule chain into multiple parallel executable rule sub-chains according to dependencies and assigns independent execution threads. Combined with rule prefetching and result caching mechanisms, it optimizes the overall execution performance. At the same time, it constructs a master decision-alternative decision hierarchy based on decision confidence and quantifies decision uncertainty. This achieves a breakthrough from serial rule execution to parallel scheduling, improving the execution efficiency of the rule engine and the reliability assessment capability of decision results.

[0191] The process of fusing the insurance fraud probability value and the rule-based decision result, and using a decision trajectory tracing mechanism to resolve conflicts, to obtain the final review decision result includes:

[0192] Construct a cross-validation confidence matrix between the insurance fraud probability value and the rule decision result, and identify decision conflict points based on the confidence difference;

[0193] A full-link decision trajectory map is established, which includes feature extraction, probability calculation, rule triggering, and decision generation. The source of conflict is located by trajectory backtracking, and a multi-level conflict resolution strategy based on evidence strength weight is adopted to generate the final review decision result.

[0194] The establishment of a full-link decision trajectory map, which includes feature extraction, probability calculation, rule triggering, and decision generation, and the location of conflict sources through trajectory backtracking, includes:

[0195] Assign a unique trajectory node identifier to each decision-making stage, record the causal transmission relationship and information flow between trajectory nodes, and construct a directed acyclic full-link decision trajectory map;

[0196] When a decision conflict is detected, the entire decision trajectory map is traversed in reverse from the conflict point. The contribution of each trajectory node to the conflict is calculated, the trajectory node with the highest contribution is identified as the source of the conflict, and the input data and processing parameters of the source of the conflict are extracted for conflict analysis.

[0197] In one specific embodiment, the formula for calculating the decision conflict point identification based on confidence difference is:

[0198] ;

[0199] ;

[0200] in, for Cross-validation confidence matrix; The probability-probability autocorrelation confidence level. For probability-rule cross-confidence, For rule-probability cross-confidence, Rule-rule autocorrelation confidence; These are the matrix determinant weight coefficients; These are the Frobenius norm weighting coefficients; It serves as a comprehensive measure of the overall degree of conflict. Let be the trace function of the matrix. Let be the determinant function of the matrix. The Frobenius norm of the matrix;

[0201] For the cross-validation confidence matrix The elements in ;

[0202] in, ; This is the variance amplification factor, used to highlight instability; The number of samples used to calculate the elements of the confidence matrix is ​​the total number of decision instances participating in cross-validation. For the first The confidence score from the first decision source (e.g., a probabilistic model) in each sample. For the first The confidence value from a second decision source (such as a rule engine) in each sample; It is the variance function;

[0203] The formulas for calculating the contribution of each trajectory node to the conflict and the multi-level conflict resolution strategy are as follows:

[0204] ;

[0205] ;

[0206] in, Contribution to the collision of trajectory nodes; The output value of the trajectory node; The traffic weight for the information flow path; The depth of the trajectory node in the decision trajectory; This represents the depth of influence coefficient. For the first Layer resolution weights; The strength of evidence for trajectory nodes; It is the power exponent of the strength of evidence, used to nonlinearly enhance the influence of high-confidence evidence; This refers to a specific information flow path from the input to the current trajectory node, including the propagation route of information in the decision trajectory graph, such as "input → feature extraction → probability calculation → current node"; It is the set of all possible information flow paths from the input to the current node in the full-link decision trajectory graph, used to calculate the integrity of the information flow distribution; This is the sum of traffic weights for all information flow paths, used to normalize the information traffic of a single path and ensure the rationality of path weight allocation. This is the final result of resolving multi-level conflicts; The total number of elimination layers; It is the set of all trajectory nodes involved in conflict resolution, including feature extraction nodes, probability calculation nodes, and rule triggering nodes; This refers to the specific decision output values ​​generated by the node during the conflict resolution process, such as the risk level and handling recommendations, which are quantitative results.

[0207] Specifically, this embodiment constructs a cross-validation confidence matrix and a full-link decision trajectory map, thereby achieving intelligent fusion and conflict resolution of the decision results of the probability model and the rule engine, and improving the accuracy and reliability of the decision.

[0208] By quantifying the conflict contribution of trajectory nodes and employing a multi-level resolution strategy, the precise location and automated resolution of conflict sources were achieved, significantly reducing the need for manual intervention.

[0209] The established end-to-end traceability mechanism ensures complete transparency and explainability of the decision-making process, meeting regulatory compliance requirements;

[0210] Meanwhile, the hierarchical resolution mechanism based on evidence strength weights improves decision-making efficiency in complex scenarios.

[0211] The step of outputting the final audit decision result and recording the audit trajectory information corresponding to the final audit decision result includes:

[0212] Based on the insurance business scenario and the characteristics of the receiving end, the final review decision result is encapsulated in multiple dimensions and distributed across platforms through adaptive protocol conversion.

[0213] In one specific embodiment, the business roles, permission levels, and data format requirements of the receiving end are analyzed to generate a personalized decision result display template, which includes risk level visualization, key evidence summary, and follow-up processing suggestions.

[0214] Based on the technical architecture and communication protocol of the receiving end, the data transmission format and encryption method are dynamically selected, the transmission path is determined through intelligent routing algorithms, and the decision results are distributed.

[0215] A multi-level audit trajectory information model is constructed, which includes decision-making basis, confidence level changes and conflict resolution process, and uses a blockchain-style storage structure to record audit trajectory information.

[0216] In one specific embodiment, a unique identifier based on timestamps and hash values ​​is assigned to each audit trajectory node, and the data source of the decision basis, the confidence calculation process, and the detailed steps of conflict resolution are recorded to form a structured trajectory information block;

[0217] The trajectory information blocks are linked together in chronological order using a hash chain, with each block containing the hash digest of the previous block, to obtain the audit trajectory chain. A multi-layered digital signature verification mechanism is then set up to verify the audit trajectory information.

[0218] In one specific implementation, after obtaining the final review decision, the following operations are further performed:

[0219] First, perform multi-dimensional encapsulation and cross-platform distribution;

[0220] Based on insurance business scenarios, such as fast claims settlement for auto insurance, hospitalization claims settlement for critical illness insurance, and disaster claims settlement for property insurance, as well as the characteristics of the receiving end, including the mobile terminal of the claims adjuster, the web terminal of the claims reviewer, the risk control and audit BI terminal, and the customer service IVR / SMS terminal, the final review decision results are structured and encapsulated to obtain a unified data object including risk level, key evidence, suggested actions, and explanatory metadata.

[0221] The adaptive protocol conversion module automatically selects protocols and formats such as REST / JSON, gRPC / Protobuf, MQ messages, or HL7 / FHIR based on the interface capabilities of the receiving end to complete cross-platform distribution.

[0222] Secondly, personalized display templates are generated;

[0223] Analyze the business roles, permission levels, and data format requirements of the receiving end, and automatically output different display templates for different roles:

[0224] For mobile devices used by surveyors, the focus is on visualizing risk levels (e.g., using color-coded indicators and radar charts), simplified thumbnails of key evidence, and suggestions for on-site re-photography / re-collection.

[0225] For claims auditors on the web platform, it provides a rule hit details table, Top-K model feature importance, a conflict resolution process summary, and a downloadable evidence package.

[0226] For the risk control and audit BI end, the output includes the risk distribution of batch cases, threshold offset trends, rule chain execution delay statistics, and model confidence time series curves.

[0227] The templates uniformly include suggestions for subsequent processing, such as "transfer to manual review", "supplementary information", "direct approval", and "trigger anti-fraud audit".

[0228] Next, we will optimize transmission and ensure security.

[0229] Based on the receiving end's technical architecture and communication protocol, the transmission format (e.g., gzip-JSON, binary-Protobuf) and encryption method (TLS 1.3 / national cryptographic SM2 / SM4) are dynamically selected, and an intelligent routing algorithm is used to select the optimal route (with latency, packet loss rate, and bandwidth as weighted indicators) to complete the distribution.

[0230] Enable retry and idempotency control on critical links to ensure consistent results across multiple devices.

[0231] Then, a multi-level audit trajectory information model and blockchain-style storage are implemented;

[0232] Construct a three-layer audit trajectory model, including a decision-making basis layer, a confidence level change layer, and a conflict resolution layer:

[0233] The decision-making basis layer is used to record feature sources, version numbers, feature selections, and standardization parameters.

[0234] The confidence level variation layer is used to record the confidence level, deviation vector, and fusion weight of each stage.

[0235] The conflict resolution layer is used to record conflict location nodes, contribution rankings, resolution strategies, and threshold parameters.

[0236] Each audit trace node is assigned a unique identifier based on a timestamp and hash, resulting in a structured trace information block. The trace information blocks are linked together in chronological order using a hash chain, and a multi-layered digital signature (system signature + auditor signature) verification mechanism is used at key nodes to achieve a "traceable and tamper-proof" audit trace chain.

[0237] Finally, end-side feedback and closed-loop management are implemented;

[0238] After processing, the receiving end sends back a receipt (such as read, accepted / rejected suggestion, supplementary material upload, etc.). The system will bidirectionally associate the receipt with the original audit trail chain to form closed-loop data.

[0239] The closed-loop data can be used periodically for model retraining and rule optimization.

[0240] Specifically, this embodiment uses multi-dimensional encapsulation and adaptive distribution to ensure that the same decision result can be accurately and cost-effectively received and understood across different business scenarios and terminal forms, improving cross-platform availability and delivery efficiency. Personalized display templates based on role-based permissions provide customized risk visualization and evidence summaries, reducing the cognitive burden of manual review and improving processing efficiency and consistency.

[0241] By dynamically selecting transmission formats, encryption methods, and optimal routing, latency and bandwidth consumption are reduced while ensuring data security and compliance, improving the stability and performance of large-scale distribution. The introduction of a blockchain-style storage structure and a multi-signature audit trail chain ensures audit data is traceable, non-repudiable, and tamper-proof, meeting the stringent requirements of the insurance industry for interpretability, compliance auditing, and regulatory reporting. A closed-loop governance channel is formed through edge-side receipts, providing high-quality supervisory data.

[0242] This invention also provides an image review system based on machine learning and the Drools engine, the system comprising:

[0243] The image feature extraction module is used to acquire the insurance claim image to be reviewed and extract the image feature vector of the insurance claim image;

[0244] The fraud probability calculation module is used to input the image feature vector into the forest algorithm model to calculate the insurance fraud probability value;

[0245] The dynamic threshold calculation module is used to obtain business attribute information of insurance claim cases and dynamically calculate probability thresholds based on the business attribute information.

[0246] The rule chain selection module is used to compare the insurance fraud probability value with the probability threshold, and select the corresponding insurance business rule chain in the Drools rule engine according to the comparison result.

[0247] The rule chain dynamic reconstruction module is used to obtain the current state of the insurance claims process and dynamically reconstruct the insurance business rule chain based on the current state to obtain the reconstructed insurance business rule chain.

[0248] The rule engine execution module is used to perform Drools rule engine judgments on the reconstructed insurance business rule chain and output the rule decision results;

[0249] The decision fusion and conflict resolution module is used to fuse the insurance fraud probability value and the rule decision result, and use a decision trajectory tracing mechanism to resolve conflicts to obtain the final review decision result.

[0250] The decision output and review module is used to output the final review decision result and record the audit trajectory information corresponding to the final review decision result.

[0251] Specifically, this embodiment of an image review system based on machine learning and the Drools engine achieves end-to-end intelligent processing of insurance claim image review through the collaborative work of eight functional modules, thereby improving the accuracy of review decisions and processing efficiency.

[0252] The system employs a dual-judgment mechanism combining a forest algorithm model and dynamic threshold calculation, which effectively reduces the false positive rate and false negative rate.

[0253] The established end-to-end tracking and auditing mechanism ensures complete transparency and auditability of the decision-making process, meeting regulatory compliance requirements;

[0254] The entire system boasts excellent scalability and a modular architecture, providing the insurance industry with a complete solution integrating intelligent identification, dynamic decision-making, conflict resolution, and compliance auditing, thus driving the development of insurance claims review towards intelligence, automation, and standardization.

[0255] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An image moderation method based on machine learning and the Drools engine, characterized in that, Includes the following steps: Obtain the insurance claim image to be reviewed, and extract the image feature vector of the insurance claim image; The image feature vector is input into the forest algorithm model to calculate the insurance fraud probability value; Obtain business attribute information of insurance claim cases, and dynamically calculate probability thresholds based on the business attribute information; The insurance fraud probability value is compared with the probability threshold, and the corresponding insurance business rule chain in the Drools rule engine is selected based on the comparison result. Obtain the current status of the insurance claims process, and dynamically reconstruct the insurance business rule chain based on the current status to obtain the reconstructed insurance business rule chain; The reconstructed insurance business rule chain is evaluated using the Drools rule engine, and the rule decision results are output. The insurance fraud probability value and the rule decision result are fused together, and a decision trajectory tracing mechanism is used to resolve conflicts to obtain the final review decision result. The process of fusing the insurance fraud probability value and the rule-based decision result, and using a decision trajectory tracing mechanism to resolve conflicts, to obtain the final review decision result includes: Construct a cross-validation confidence matrix between the insurance fraud probability value and the rule decision result, and identify decision conflict points based on the confidence difference; A full-link decision trajectory map is established, which includes feature extraction, probability calculation, rule triggering and decision generation. The source of conflict is located by trajectory backtracking, and a multi-level conflict resolution strategy based on evidence strength weight is adopted to generate the final review decision result. Output the final audit decision result and record the audit trajectory information corresponding to the final audit decision result.

2. The image review method based on machine learning and the Drools engine as described in claim 1, characterized in that, The process of obtaining business attribute information of insurance claim cases and dynamically calculating probability thresholds based on the business attribute information includes: The business attribute information includes the claim amount, policy lifecycle, and time period risk characteristics. Quantitative feature values ​​corresponding to the claim amount, policy lifecycle, and time period risk characteristics of the insurance claim case are extracted. Establish a multi-dimensional mapping relationship between business attribute feature values ​​and probability thresholds, and perform fusion calculation on the quantified feature values ​​based on the multi-dimensional mapping relationship to dynamically generate probability thresholds that are suitable for the characteristics of the current insurance claim case.

3. The image review method based on machine learning and the Drools engine as described in claim 2, characterized in that, The step of establishing a multi-dimensional mapping relationship between business attribute feature values ​​and probability thresholds, and performing fusion calculations on the quantized feature values ​​based on the multi-dimensional mapping relationship, includes: Construct a three-dimensional weight matrix of claim amount risk coefficient, policy period decay factor and time period risk volatility index; The quantitative feature values ​​are weighted by the corresponding risk coefficient, decay factor and volatility index respectively, and the dynamic probability threshold adjustment parameters are calculated by nonlinear fusion function. A probability threshold is generated based on the dynamic probability threshold adjustment parameters.

4. The image review method based on machine learning and the Drools engine as described in claim 1, characterized in that, The step of comparing the insurance fraud probability value with the probability threshold and selecting the corresponding insurance business rule chain in the Drools rule engine based on the comparison result includes: Calculate the confidence index of the insurance fraud probability value, combine it with the feature stability score of the image feature vector, and perform confidence correction on the insurance fraud probability value to obtain the corrected insurance fraud probability value; Based on the deviation between the corrected insurance fraud probability value and the probability threshold, a probability-threshold deviation vector is constructed, and the most matching insurance business rule chain is adaptively selected from the Drools rule engine according to the probability-threshold deviation vector.

5. The image review method based on machine learning and the Drools engine as described in claim 4, characterized in that, The step of constructing a probability-threshold deviation vector based on the deviation between the corrected insurance fraud probability value and the probability threshold, and adaptively selecting the best-matching insurance business rule chain based on the probability-threshold deviation vector, includes: The numerical deviation, trend deviation, and distribution interval deviation of the corrected insurance fraud probability value from the probability threshold are calculated to form a three-dimensional deviation vector; The three-dimensional deviation vector is matched with the pre-stored rule chain feature vector library for vector similarity. The insurance business rule chain corresponding to the rule chain feature vector with the highest similarity is selected, and the execution weight of each rule node in the rule chain is dynamically adjusted according to the similarity value.

6. The image review method based on machine learning and the Drools engine as described in claim 1, characterized in that, The process of obtaining the current state of the insurance claims process and dynamically reconstructing the insurance business rule chain based on the current state to obtain a reconstructed insurance business rule chain includes: Construct a multi-dimensional feature description model of the insurance claims process status, and predict the evolution path of the subsequent process status based on the current status features; The insurance business rule chain is reconstructed in a forward-looking manner according to the evolution path. The forward-looking reconstruction includes adjusting the topology of the rule chain and redistributing the execution priorities among the rule nodes to obtain a reconstructed insurance business rule chain that adapts to the state evolution.

7. The image review method based on machine learning and the Drools engine as described in claim 6, characterized in that, The forward-looking reconstruction of the insurance business rule chain based on the evolution path includes: Based on the key state nodes in the evolution path, identify the combination of rule nodes in the insurance business rule chain that generate execution conflicts or efficiency bottlenecks. The identified rule node combinations are decoupled and recombined. By constructing a dynamic dependency graph between rule nodes, the execution sequence and branch conditions of the insurance business rule chain are redesigned. Each reconstructed rule branch is assigned an execution weight proportional to the state evolution probability, thus obtaining the reconstructed insurance business rule chain.

8. The image moderation method based on machine learning and the Drools engine as described in claim 1, characterized in that, The establishment of a full-link decision trajectory map, which includes feature extraction, probability calculation, rule triggering, and decision generation, and the location of conflict sources through trajectory backtracking, includes: Assign a unique trajectory node identifier to each decision-making stage, record the causal transmission relationship and information flow between trajectory nodes, and construct a directed acyclic full-link decision trajectory map; When a decision conflict is detected, the entire decision trajectory map is traversed in reverse from the conflict point. The contribution of each trajectory node to the conflict is calculated, the trajectory node with the highest contribution is identified as the source of the conflict, and the input data and processing parameters of the source of the conflict are extracted for conflict analysis.

9. An image moderation system based on machine learning and the Drools engine, used to execute the image moderation method based on machine learning and the Drools engine as described in any one of claims 1-8, characterized in that, The system includes: The image feature extraction module is used to acquire the insurance claim image to be reviewed and extract the image feature vector of the insurance claim image; The fraud probability calculation module is used to input the image feature vector into the forest algorithm model to calculate the insurance fraud probability value; The dynamic threshold calculation module is used to obtain business attribute information of insurance claim cases and dynamically calculate probability thresholds based on the business attribute information. The rule chain selection module is used to compare the insurance fraud probability value with the probability threshold, and select the corresponding insurance business rule chain in the Drools rule engine according to the comparison result. The rule chain dynamic reconstruction module is used to obtain the current state of the insurance claims process and dynamically reconstruct the insurance business rule chain based on the current state to obtain the reconstructed insurance business rule chain. The rule engine execution module is used to perform Drools rule engine judgments on the reconstructed insurance business rule chain and output the rule decision results; The decision fusion and conflict resolution module is used to fuse the insurance fraud probability value and the rule decision result, and use a decision trajectory tracing mechanism to resolve conflicts to obtain the final review decision result. The decision output and review module is used to output the final review decision result and record the audit trajectory information corresponding to the final review decision result.

Images