Artificial intelligence-based data processing method, device, equipment and readable storage medium

Abstract

The application provides an artificial intelligence-based data processing method and device, equipment and readable storage medium, relating to the technical field of artificial intelligence, which comprises the following steps: collecting log data in a business case execution process in real time to obtain multidimensional log data; performing deviation analysis on the multidimensional log data through a dynamic threshold model to determine at least one deviation case; classifying the at least one deviation case according to a classification rule; performing root cause analysis on the at least one deviation case in combination with the classification of the at least one deviation case to determine the deviation root cause of the at least one deviation case; determining a target case screening strategy based on the at least one deviation case and the corresponding deviation root cause, performing data screening on a to-be-processed business case, and determining a target business case, so as to accurately locate an exception in the case screening strategy, shorten the rectification cycle, reduce the time cost, reduce the mis-screening proportion, and improve the accuracy and automation of business case screening.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a data processing method, apparatus, device, and readable storage medium based on artificial intelligence. Background Technology

[0002] In traditional business scenarios, updating case selection strategies heavily relies on manual intervention and exhibits significant lag in response. When business terms change or intelligent processing deviates, adjustments to case selection strategies depend entirely on technical personnel manually reviewing logs, iterating code, and releasing versions. From problem discovery to the official implementation of rule updates, it often takes several days or even weeks, making it difficult to quickly adapt to various dynamic changes in business. At the same time, existing optimization coverage dimensions are relatively singular, focusing only on the level of intelligent models and basic data verification rules, without building an automated closed-loop optimization system for the selection strategies of core business decisions in various domains. Deviation data generated during intelligent execution cannot be effectively fed back to the rule engine for iteration.

[0003] Therefore, in view of the shortcomings of related technical solutions, the present invention provides a data processing method based on artificial intelligence. Summary of the Invention

[0004] Therefore, it is necessary to provide a data processing method, apparatus, device, and readable storage medium based on artificial intelligence to address the aforementioned technical problems.

[0005] This application provides an artificial intelligence-based data processing method, which includes: real-time collection of log data during the execution of business cases to obtain multi-dimensional log data; deviation analysis of the multi-dimensional log data using a dynamic threshold model to identify at least one deviation case; classification of the at least one deviation case according to classification rules; root cause analysis of the at least one deviation case based on its classification to identify the root cause of the deviation in the at least one deviation case; determination of a target case screening strategy based on the at least one deviation case and its corresponding root cause; and data screening of the business cases to be processed according to the target case screening strategy to identify target business cases.

[0006] This application also provides an artificial intelligence-based data processing device, comprising: a first processing module for real-time acquisition of log data during the execution of business cases to obtain multi-dimensional log data; a second processing module for performing deviation analysis on the multi-dimensional log data using a dynamic threshold model to identify at least one deviation case; a third processing module for classifying the at least one deviation case according to classification rules; a fourth processing module for performing root cause analysis on the at least one deviation case based on its classification to identify the root cause of the deviation in the at least one deviation case; a fifth processing module for determining a target case selection strategy based on the at least one deviation case and its corresponding root cause; and a sixth processing module for performing data screening on the business cases to be processed according to the target case selection strategy to identify target business cases.

[0007] This application also provides an electronic device, including: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of any of the above-described artificial intelligence-based data processing methods.

[0008] This application also provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, it implements the steps of any of the above-described artificial intelligence-based data processing methods.

[0009] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of any of the above-described artificial intelligence-based data processing methods.

[0010] The aforementioned AI-based data processing method, apparatus, device, and readable storage medium include the following steps: Real-time acquisition of log data during business case execution to obtain multi-dimensional log data; deviation analysis of the multi-dimensional log data using a dynamic threshold model to identify at least one deviation case; classification of the at least one deviation case according to classification rules; root cause analysis of the at least one deviation case based on its classification to determine the root cause of the deviation; determination of a target case selection strategy based on the at least one deviation case and its corresponding root cause; and data screening of the business cases to be processed according to the target case selection strategy to identify target business cases. This achieves precise location of anomalies in the case selection strategy without manual intervention, significantly reducing manpower investment in business problem investigation and rectification, shortening the rectification cycle, and reducing time costs. Simultaneously, it can significantly reduce the proportion of false screenings, effectively improve the accuracy of business case selection, and further increase the automation rate of business processes. Attached Figure Description

[0011] To more clearly illustrate the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart illustrating an artificial intelligence-based data processing method in one embodiment; Figure 2 This is a schematic diagram of the overall process of an artificial intelligence-based data processing method in one embodiment; Figure 3 This is a schematic diagram of the sandbox verification process timing of an artificial intelligence-based data processing method in one embodiment. Figure 4 This is a structural block diagram of an artificial intelligence-based data processing device in one embodiment; Figure 5 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] It should be understood that, in the description of this application, unless the context explicitly requires it, words such as "including" or "comprising" throughout the specification should be interpreted as including rather than exclusive or exhaustive; that is, meaning "including but not limited to".

[0015] It should also be understood that the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0016] It should be noted that the terms "S1," "S2," etc., are used only for descriptive purposes and do not specifically refer to the order or sequence, nor are they intended to limit this application. They are merely for the convenience of describing the method of this application and should not be construed as indicating the sequential order of the steps. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0017] In one embodiment, such as Figure 1 As shown, an artificial intelligence-based data processing method is provided, including the following steps: S101: Collect log data in real time during the execution of business cases to obtain multi-dimensional log data.

[0018] Here, the business case execution process refers to the complete lifecycle of a business request from its inception to its completion. For example, the series of operations and internal system processing steps that a user goes through in the order process—browsing products, adding them to cart, submitting the order, making payment, and receiving the goods—constitute a business execution process.

[0019] Here, log data refers to timestamped records automatically generated by the business system during operation, used to record events, states, or operations. Examples include user button clicks, service call requests and responses, and generated error stacks.

[0020] Here, multi-dimensional log data is a collection of logs that covers the entire business process, multiple perspectives, is sufficiently detailed, and has no missing key information. It not only includes error stack traces, but also business identities (such as user ID, order ID), end-to-end tracking IDs, context parameters, environmental information, etc., which can completely reconstruct a business panorama.

[0021] S102: Perform deviation analysis on multi-dimensional log data using a dynamic threshold model to identify at least one deviation case.

[0022] Here, the dynamic threshold model is an algorithm or model that automatically calculates the normal fluctuation range based on the historical trend, periodic pattern and real-time distribution of data. It can automatically adapt to the tidal effect of business (such as morning and evening peak hours) and sudden changes, and avoid missed or false alarms caused by improper threshold settings.

[0023] Among them, the dynamic threshold model can automatically adjust the judgment criteria based on the business runtime, business traffic volume, scenario changes, and historical operating patterns.

[0024] Specifically, the dynamic threshold model is trained in real time and the threshold parameters are automatically updated with business log data. The model update methods can include adding rules, adjusting thresholds, revising logic, and taking invalid rules offline.

[0025] Here, deviation analysis is based on the normal baseline established by the dynamic threshold model. It compares various business indicators in the real-time log data to identify and quantify the degree to which they deviate from the normal range.

[0026] Here, deviation cases are a collection of abnormal events confirmed by deviation analysis and possessing complete contextual information, including multi-dimensional information such as time, deviation indicators, degree of deviation, scope of impact, related links, and log snapshots, which can be directly used for review and attribution.

[0027] Specifically, core indicators for measuring business health are aggregated from stored log data by time window and used as the object of deviation analysis. A dynamic threshold model is trained using historical indicator data with a sufficiently long period of time. The historical indicator data includes trends, seasonality, and noise. The model is deployed into the real-time computing process, which continuously generates adaptive threshold ranges for the current moment. The real-time indicator value of the current time window is input into the model and compared with the dynamically generated threshold. If the indicator is not within the normal range, it is determined that a deviation has occurred, and the source of the deviation is locked. The deviation phenomenon is associated with relevant multi-dimensional logs and aggregated into a structured deviation case, which includes the phenomenon, the scope of impact, and the root cause clues, and is assigned a unique identifier.

[0028] S103: Classify at least one deviation case according to the classification rules.

[0029] Here, classification rules are a set of predefined logic or models used to categorize deviant cases into different categories.

[0030] The classification rules can be jointly formulated by business, technology and operations experts, and the basis can include dimensions such as the root cause type of the deviation, the scope of impact, the urgency, the deviation type, and the business scenario.

[0031] The classification rules can be explicit keyword matching trees or decision tables, or classification algorithms trained based on historical labeled data.

[0032] One deviation case corresponds to one or more categories, and the same deviation case can be classified in different fields using different classification rules.

[0033] S104: Based on the classification of at least one deviation case, perform root cause analysis on at least one deviation case to determine the root cause of deviation in at least one deviation case.

[0034] Here, root cause analysis is a structured problem localization method that aims to identify the fundamental and systemic causes of problems by looking beyond surface anomalies and direct causes.

[0035] Here, the root cause of the deviation is the most fundamental and essential reason that led to the deviation case, as confirmed by root cause analysis.

[0036] For example, the root cause of a deviation may include a line of incorrect code logic, an improper configuration change, a hardware failure, or a violated architectural design principle.

[0037] S105: Determine the target case selection strategy based on at least one deviation case and its corresponding root cause.

[0038] Here, the target case selection strategy refers to the part of the system configuration, strategy, or code logic that is identified as defective, incomplete, or missing after deviation root cause analysis and needs to be modified, added, or deleted.

[0039] S106: Based on the target case screening strategy, perform data screening on the business cases to be processed to determine the target business cases.

[0040] In the above embodiments, this application achieves accurate positioning of abnormal issues in the case screening strategy without the need for manual intervention, significantly reducing the manpower input for business problem screening and rectification, shortening the rectification cycle and reducing time costs. At the same time, it can significantly reduce the proportion of false screenings, effectively improve the accuracy of business case screening, and further improve the automation rate of business processing.

[0041] In some specific implementations, log data is collected in real time during the execution of business cases to obtain multi-dimensional log data, including: Collect at least one of the following during the business execution process in real time: interaction logs, interface call logs, business result logs, and filtering strategy trigger logs, to obtain multi-dimensional log data; Multi-dimensional log data is associated with corresponding case identifiers and stored in the log database.

[0042] Here, the interaction log records the content of each round of dialogue between the AI ​​(Artificial Intelligence) agent and the user, the user's response time, the AI's structured parsing results, and the parsing confidence level; the screening strategy trigger log records each screening strategy triggered by the case screening strategy engine during business execution, including the case screening strategy ID, trigger time, trigger condition matching status, and case screening strategy execution action (pass / intercept / transfer to human); the interface call log records the request parameters, return results, and call time of the AI ​​agent calling the core business system API (Application Programming Interface); and the business result log records the final business processing results (such as underwriting conclusion, claim amount, case status) and subsequent human intervention.

[0043] Here, a case identifier is an ID (identifier) ​​used to uniquely associate a complete business process or a specific deviation case. It links records scattered across different log types to form a traceable chain of evidence. Common examples include business transaction numbers (such as order IDs), end-to-end trace IDs (Trace IDs), or dedicated case numbers.

[0044] The case identifier can be a work order number, session ID, process instance number, etc.

[0045] Specifically, during log collection, each session allows AI to identify the business order number and simultaneously bind the case ID to the session. In insurance business scenarios, a business transaction is typically completed through a single, continuous conversation, with one case ID corresponding to one session ID.

[0046] The data acquisition module is deployed at the business execution layer.

[0047] In this way, a full-view log profile is constructed with business cases as the link, so that any business process can be instantly restored with complete multi-dimensional details by a unique identifier, achieving millimeter-level holographic backtracking.

[0048] In some specific implementations, a dynamic threshold model is used to perform deviation analysis on multi-dimensional log data to identify at least one deviation case, including: Based on the case identifier, obtain the case data corresponding to each case from the multi-dimensional log data; Calculate the business quality indicators for each case data, including at least one of the following: automation completion rate, manual transfer rate, policy interception rate, average confidence of field parsing, frequency of user correction, average number of interaction rounds, and frequency of policy triggering. The business quality indicators are determined using a dynamic threshold model. In response to the existence of business quality indicators that meet the deviation conditions, the case corresponding to the case data is identified as a deviation case.

[0049] Here, the automation completion rate is the proportion of business cases completed independently by the AI ​​agent (without human intervention) out of the total number of cases; the manual transfer rate is the proportion of cases actively transferred by the AI ​​agent to human agents or passively triggered human intervention; the policy interception rate is the proportion of cases where the policy engine triggers interception actions (such as rejection or transfer to human underwriting); the average number of interaction rounds is the average number of dialogue rounds between the AI ​​agent and the user to complete a complete business task; the average confidence score of field parsing is the average confidence score when the AI ​​agent converts the user's natural language into structured fields; the user correction frequency is the average number of times the user corrects the AI ​​agent's understanding during the interaction; and the policy trigger frequency is the frequency statistics of each atomic policy being triggered.

[0050] Here, case data is detailed data extracted from multi-dimensional logs, describing the entire process and results of the case, with the case identifier as the aggregation dimension. It integrates interaction logs, interface call logs, business result logs, and filtering strategy trigger logs to form complete data for a single case.

[0051] For example, deviation conditions include, but are not limited to: the number of interaction rounds for a single case exceeding twice the average number of rounds for similar cases; the confidence level of field parsing being less than 0.5 and the user correcting the deviation; the triggering frequency of a specific case screening strategy fluctuating beyond a preset threshold in a short period of time (e.g., the triggering rate of a certain underwriting interception case screening strategy suddenly increasing from 5% to 20%); after manual transfer, the processing conclusion of the human agent differs significantly from the initial judgment of the AI ​​agent; and the case is ultimately complained about or appealed by the customer.

[0052] Specifically, based on the preset case identifier, the system accurately extracts the unique case data corresponding to each business case from the multi-dimensional log data stored in the log database. For each set of extracted case data, it statistically calculates one or more corresponding business quality indicators, completes the quantitative statistics of business operation status, calls the dynamic threshold model, adaptively adjusts the evaluation threshold according to the current business scenario, and compares and judges each calculated business quality indicator one by one. When any business quality indicator is detected to have reached and met the preset deviation conditions, the business case to which that set of indicators belongs is immediately determined to be a deviation case.

[0053] In this way, by relying on multi-dimensional quantitative indicators combined with adaptive thresholds, abnormal business cases can be accurately identified, thereby improving the comprehensiveness and accuracy of business anomaly identification.

[0054] In some specific implementations, at least one deviation case is classified according to classification rules, including: Based on the first classification rule, the deviation cases are analyzed to determine the deviation type of the deviation cases; Based on the second classification rule, the deviation cases are analyzed to determine the business scenario type of the deviation cases; Based on the third classification rule, the deviation cases are analyzed to determine their severity.

[0055] The first classification rule is usually based on matching features such as error codes, exception stacks, and process breakpoints in the case data.

[0056] Among these, deviation types can include overly strict / loose screening conditions, AI comprehension bias, OCR (Optical Character Recognition) errors, abnormal data interfaces, and non-standard user expressions.

[0057] The second classification rule is usually based on dimensions such as page ID, service name, interface path, or product type in the case data.

[0058] The business scenarios can include order entry scenarios, underwriting scenarios, claims scenarios, after-sales refunds, new customer registration, marketing coupon distribution, and order payment.

[0059] The third classification rule typically takes into account factors such as the number of users affected by the deviation, the amount of money involved, the duration of business interruption, and whether sensitive data is involved.

[0060] The severity level can be categorized as mild, moderate, or severe.

[0061] For example, in a credit review process, if the average confidence level of a field parsing for a credit application is lower than a threshold, causing a delay in the review process, but without resulting in any loss, only affecting the review of that single application, then according to the first classification rule: the deviation is caused by a data parsing error, and the deviation type is determined to be an OCR recognition error; according to the second classification rule: it occurs in the credit review process, and the business scenario type is determined to be a credit review scenario; according to the third classification rule: there is no loss and the impact is minimal, and the severity is determined to be minor.

[0062] For example, in an e-commerce order placement scenario, an IP address uses a malicious script to place bulk orders, causing a surge in system orders, interface response timeouts, and affecting the normal order placement of multiple users, resulting in the loss of a small number of orders. Based on the first classification rule: the deviation is caused by malicious operation, the deviation type is determined to be a data interface anomaly; based on the second classification rule: it occurs in the order placement process, the business scenario type is determined to be an order placement scenario; based on the third classification rule: there is minor loss, affecting multiple business processes, the severity is determined to be moderate.

[0063] In this way, by clearly identifying the core attributes, occurrence scenarios, and priority of handling deviations, a clear classification basis is provided for subsequent root cause analysis and targeted rectification.

[0064] In some specific implementations, based on the classification of at least one deviation case, root cause analysis is performed on at least one deviation case to determine the root cause of the deviation in at least one deviation case, including: Obtain the deviation case classification by deviation type, business scenario, and severity. Based on deviation type classification, business scenario classification, and severity classification, local subgraphs are determined in a pre-constructed strategy graph; the strategy graph contains multiple types of nodes and relational edges connecting the nodes, and the nodes include at least root cause nodes, evidence nodes, business scenario nodes, deviation type nodes, and severity nodes; Based on the deviation type nodes, root cause nodes, and local subgraphs, determine the initial set of candidate root cause hypotheses; Based on the scenario rules associated with business scenario nodes, the initial candidate root cause hypothesis set is supplemented or removed, and the prior probabilities of candidate root causes are adjusted to obtain a scenario-based root cause hypothesis set. Based on the severity classification, the analytical depth and evidence threshold of the root cause analysis are determined, where the analytical depth corresponds to the number of hops in the causal walk in the policy graph, and the evidence threshold corresponds to the minimum requirement of rule strength. Based on the scenario-based root cause hypothesis set and analysis depth, with each candidate root cause as the center, a limited-depth walk is performed along the causal edge and the evidence chain edge to extract reachable evidence nodes and construct a local reasoning subgraph. The causal edge represents the relationship between the root cause and the evidence, and the evidence chain edge connects the evidence nodes to the evidence nodes, representing the prerequisite order for obtaining the evidence. In the local reasoning subgraph, calculate the information gain of the uncollected evidence nodes and collect the evidence with the largest information gain. Based on the collected evidence, backpropagation is performed along the causal edges to update the posterior probability of each candidate root cause, and the posterior probability is adjusted according to the cooperative or inhibitory relationship of the relation edges. Repeat the above collection and update steps until the termination condition is met, and output the candidate root cause that meets the evidence threshold and has the highest posterior probability as the root cause of the biased case.

[0065] Here, a strategy graph is a domain knowledge base stored in a graph structure. Nodes represent entities or concepts, and edges represent deterministic case selection strategies or probabilistic relationships between them, which are used to guide the reasoning process.

[0066] Specifically, the strategy graph is a structured graph pre-built based on business scenarios, deviation types, historical root cause data, etc. It contains multiple types of nodes and relational edges connecting the nodes. The nodes include at least root cause nodes (representing the root cause that may cause the deviation), evidence nodes (representing relevant data / clues that can be used to verify the root cause), business scenario nodes (corresponding one-to-one with the business scenario classification, representing the specific business scenario in which the deviation occurred), deviation type nodes (corresponding one-to-one with the deviation type classification, representing the specific type of deviation), and severity nodes (corresponding one-to-one with the severity classification, representing the severity level of the deviation). Relational edges are used to represent the association logic between the nodes and can include causal edges, correlation edges, mutually exclusive edges, and evidence chain edges.

[0067] Here, the local subplot is a relevant region activated and delineated from the full policy map based on the classification information of the current bias.

[0068] Here, the initial candidate root cause hypothesis set is a collection of all possible root causes initially screened from the graph based solely on the type of deviation, without scenario filtering.

[0069] Specifically, based on deviation type nodes, root cause nodes, and local subgraphs, the initial candidate root cause hypothesis set is determined by: relying on the local subgraph and combining the relationship between deviation type nodes and root cause nodes, screening out all root cause candidates that may cause the current deviation, constructing the initial candidate root cause hypothesis set, and ensuring that the hypothesis set covers all potential root causes corresponding to the deviation type.

[0070] Here, scenario rules are logical relationships defined between business scenarios and root causes, used to dynamically add or delete candidate root causes or adjust their probability of occurrence based on the current scenario.

[0071] For example, in the insurance scenario, the tolerance for OCR recognition errors may be higher, and optimizations may be made in a more lenient direction, such as trying to obtain the license plate number through other means and having the user confirm it before continuing the process. However, in the underwriting scenario, the tolerance is lower, and the process may be switched to manual confirmation or the user may be asked to provide the relevant materials again.

[0072] Here, the resolution depth is the maximum number of hops required to walk along the edges of the strategy graph to collect evidence, which determines the depth of the investigation.

[0073] Here, analysis depth and evidence threshold are the analytical input scales determined by the severity level. Analysis depth refers to the number of levels of evidence traced (e.g., whether to investigate only the service layer or delve deeper into the host and network layers), while the evidence threshold refers to the quantity and strength of evidence required to establish a hypothesis. High severity requires exhaustive tracing and irrefutable evidence, while low severity allows for reasonable inferences and rapid remediation.

[0074] Among them, the higher the severity, the more causal jumps there are, and the deeper the root cause tracing is; the higher the severity, the stricter the evidence threshold.

[0075] Specifically, the severity level may affect the approval path and priority of the optimization strategy; minor severity can be configured to be automatically released after sandbox verification; moderate severity requires manual confirmation of case screening for strategy change proposals, and only after confirmation can they enter the sandbox verification and release process; severe severity needs to limit the processing time based on moderate severity, and if it is not processed within a certain time, an alarm notification will be sent to the operation and maintenance or business personnel.

[0076] Here, the evidence threshold is the minimum strength of evidence required to establish the root cause; the higher the severity, the stricter the threshold setting.

[0077] Here, the local reasoning subgraph is a subnetwork specifically used for this reasoning, which is formed by extracting the related evidence nodes and edges from the candidate root causes in the scenario-based root cause hypothesis set within a limited depth.

[0078] Here, information gain is a measure of the expected amount of information that collecting a previously unobserved evidence node can bring to distinguish between candidate root causes.

[0079] Here, backward probability propagation updates the posterior probability of the root cause occurrence by using Bayes' theorem in reverse along the causal edge based on observed evidence.

[0080] Here, cooperative and inhibitory relationships are adjusted in conjunction with the relationship edges (such as mutually exclusive edges) in the graph during probability updates. An increase in the probability of one root cause will suppress its mutually exclusive root cause.

[0081] Here, the termination condition is a clear criterion for stopping the root cause analysis process, which mainly includes the posterior probability of the candidate root cause reaching a preset threshold, reaching the maximum causal walk hop count, and having collected all high information gain evidence.

[0082] In this way, the core root cause of the deviation can be identified efficiently and accurately.

[0083] In some specific implementations, a target case selection strategy is determined based on at least one deviation case and its corresponding root cause, including: Based on the root causes of the deviation, determine the screening strategy for cases to be updated from the preset strategy library; Obtain the update type corresponding to the historical execution dataset and the root cause of the deviation associated with the case filtering strategy to be updated; Based on the update type, determine the strategy parameters to be updated and generate at least one candidate parameter value; Input at least one candidate parameter value into the historical execution dataset, evaluate the effect of each candidate parameter value, and obtain the evaluation index. Based on evaluation metrics and preset business constraints, the optimal parameter value is selected from the candidate parameter values, and an update suggestion containing the optimal parameter value is generated. Based on the update suggestions, the screening strategy for cases to be updated is updated to obtain the target case screening strategy.

[0084] Here, the historical execution dataset is a complete snapshot of the data filtering strategy that has been triggered and executed over a period of time. Each record contains the input of the case filtering strategy at that time (such as confidence score, user information), the judgment result (pass / block / transfer to human agent), and the final actual business result (such as whether the user complained or whether the refund was successful).

[0085] The update types can include parameter tuning strategies, add / delete strategies, field mapping optimization strategies, interaction strategy optimization strategies, interface call optimization strategies, and logic corrections.

[0086] Each update strategy includes a proposal ID, a list of associated deviation case IDs, the original case selection strategy content, the suggested revised case selection strategy content, the reason for the change (with data support), and the estimated scope of impact.

[0087] Among them, the parameter tuning strategy is to generate threshold adjustment suggestions when the interception rate of a certain data screening strategy is detected to be too high and a large number of cases are released after manual review; for example, the "policyholder age > 60 years old transfer to manual underwriting" is adjusted to "policyholder age > 65 years old transfer to manual underwriting". The addition and deletion strategy can be to generate suggestions for adding exemption conditions when it is detected that a certain case screening strategy has blocked a large number of actually compliant boundary cases, or to abolish the old strategy when a large number of non-compliant boundary cases are missed; for example, to add the exemption condition "If the insured's profession is an athlete, the medical examination can be waived" to the "BMI>30 requires medical examination" strategy. The field mapping optimization strategy is a suggestion generated by the system to optimize the field mapping strategy when it is detected that a certain type of data field is frequently parsed incorrectly by the AI ​​agent or recognized by OCR. For example, adding an alias field mapping for the name of a hospital can improve the recognition accuracy. The interaction strategy optimization strategy is to generate interaction strategy adjustment suggestions when the system detects that the frequency of user correction in a certain scenario is consistently high. For example, the method of collecting a certain type of information may be changed from "open-ended questions and answers" to "option-based guidance". The interface call optimization strategy is to generate interface call optimization suggestions (such as adding a retry mechanism or adjusting the timeout threshold) when the system detects that a core business system interface frequently times out or returns an exception.

[0088] Here, the candidate parameter values ​​are a set of policy parameter values ​​to be evaluated generated within an adjustable range based on the update type.

[0089] Here, the evaluation metric involves substituting each candidate parameter value into the historical execution dataset for replay calculation, and observing how the business metrics change under the new case selection strategy.

[0090] The evaluation metrics typically include automated completion rate, accuracy rate, manual transfer rate, and expected number of complaints, which are used to quantify the overall impact of the new parameters.

[0091] Here, the business constraints are the bottom lines that cannot be crossed when selecting the optimal parameters.

[0092] The update suggestions may include threshold adjustments, addition of exemption conditions, optimization of field mapping, and adjustment of interaction strategies.

[0093] Specifically, taking the underwriting case screening strategy "age > 60 transfer to manual review" as an example, the system extracts the dataset of all blocked cases in the most recent statistical period of this case screening strategy, generates a series of candidate thresholds with the original threshold as the center, evaluates the execution result of the case dataset for each candidate threshold, calculates the expected improvement in pass rate and interception rate, and then selects the one with the highest improvement rate as the new threshold in combination with business constraints.

[0094] Specifically, if a case is blocked by a certain case screening strategy (such as "BMI>30 requires a physical examination"), but is released after manual review and a reason is provided (such as "professional athlete, normal body fat percentage"), the system will collect these cases, analyze the reasons for release provided by the manual reviewer through AI, automatically extract structured exemption conditions (such as "professional = athlete"), and generate a case screening strategy change proposal.

[0095] In this way, by using historical data to simulate and select the best strategy parameters, the strategy update can take into account both the risk control effect and the actual business operation needs, thereby improving the rationality and practicality of strategy optimization.

[0096] In some specific implementations, the method further includes: Deploy the target case selection strategy to the sandbox verification unit and perform replay verification and simulation testing; In response to the successful verification of the target case screening strategy, a corresponding strategy version number is generated and stored in the strategy version repository; The target case selection strategy is published to the strategy version repository, and the circuit breaker rollback unit is activated to continuously monitor the performance metrics of the target case selection strategy. During monitoring, if the target case screening strategy's operational metrics trigger the preset circuit breaker condition, the target case screening strategy will be rolled back to the previous stable version through the circuit breaker rollback unit, and an alarm notification will be sent.

[0097] Here, the sandbox verification unit is a simulation execution environment that is completely isolated from the production environment but has the same data structure and logic. It does not process real traffic, but can replay production logs or execute simulation test cases. It is used to verify the correctness, performance and impact on business results of the case selection strategy before it goes live, and will not pose any risk to real users.

[0098] The sandbox verification unit has the following verification capabilities: Historical case replay: Replaying the original interaction data corresponding to the associated deviation case IDs in the sandbox environment to verify whether the optimized case selection strategy can correctly handle these deviation cases; Batch simulation test: Randomly selecting historical normal cases from the production environment for batch simulation to verify whether the optimized case selection strategy will cause false negatives to normal cases (i.e., whether the new case selection strategy is too lenient, causing risky cases that should have been blocked to be allowed to pass); A / B comparison analysis: Comparing the performance differences between the new and old case selection strategies on the same test dataset to quantify the optimization effect.

[0099] Here, replay verification and simulation testing involves inputting historical real request data or constructed simulation test cases into a sandbox to run the new and old case filtering strategy. By comparing the output, the behavior of the new case filtering strategy is verified to see if it meets expectations, whether there are logical errors or performance degradation.

[0100] Here, the case selection strategy repository serves as the centralized storage and management center for case selection strategies. It is implemented using a database that supports versioning (such as Git, or a relational database with a version field). The stored content includes: the case selection strategy ID and content; the case selection strategy version number and effective time range; dependencies between case selection strategies; and the case selection strategy change history (comparison of content before and after changes, reasons for changes, and related deviation cases), etc.

[0101] All business case selection strategies are managed in a versioned manner. Each change to the case selection strategy generates a new version number and records the change content, change time, person making the change (or automated system), and related deviation analysis reports. It supports viewing, comparing, and rolling back any historical version.

[0102] Here, the circuit breaker rollback unit is deployed in the production environment and is an automated component that continuously monitors key performance indicators after the new case selection strategy is launched. Once the indicators deteriorate to the point of triggering a preset red line, the case selection strategy is automatically switched back to the previous stable version and an alert is issued, without the need for manual intervention.

[0103] Here, the preset circuit breaker condition is a predefined set of indicator thresholds used to determine whether the new case screening strategy has caused unacceptable damage to the production environment.

[0104] The circuit breaker conditions may include: the interception rate of the target case screening strategy exceeds the preset fluctuation range (such as exceeding 200% of the historical average); the manual transfer rate spikes abnormally after the target case screening strategy is triggered; and the failure rate of core business system interface calls increases significantly due to the target case screening strategy.

[0105] In this way, a closed-loop security protection system is constructed across the entire chain.

[0106] In one embodiment, Figure 2 This is a schematic diagram of the overall process in the embodiments of this application, such as... Figure 2 As shown, the overall process in this application includes: log collection, including collecting various business logs such as interaction logs, trigger logs, interface logs, and result logs; multi-dimensional quality indicator calculation, including automated completion rate, manual transfer rate, rule interception count, and interaction rounds; automatic identification of deviation cases based on dynamic threshold judgment conditions; classification and labeling of deviation cases, including scenarios, types, and severity; automatic generation of optimization strategies based on the labeled deviation cases, and sending the optimized rules into a sandbox environment for verification. After sandbox verification is completed, it is determined whether the verification passes: if the verification fails, the process returns to the automatic optimization strategy generation stage, and the strategy is re-optimized; if the verification passes, rule version data is generated, and the process enters the production environment release and monitoring stage; after the rule is released, the running status is continuously monitored, and it is determined whether an anomaly is triggered; if an anomaly is triggered, the rollback to the previous version is immediately executed, and the process ends; if no anomaly is triggered, the current rule remains effective, and the process ends.

[0107] In one embodiment, Figure 3 This is a schematic diagram of the sandbox verification process timing in an embodiment of this application, as shown below. Figure 3 As shown, the sandbox verification processing sequence in this application includes: Phase 1: Traffic recording and Phase 2: Sandbox playback verification.

[0108] Specifically, the core objective of Phase One is to acquire and store historical business data from the real production environment that can be used for subsequent verification. The production environment (business execution layer) sends request and response messages from the actual business processing to the traffic recording agent. After receiving this data, the traffic recording agent performs anonymization processing on the session data to remove sensitive information such as user privacy, and stores the processed deviation case session data in the historical case library to provide real test samples for subsequent verification.

[0109] Specifically, Phase Two is executed in an independent sandbox validation environment to verify the correctness of the new version of rules or models, avoiding the risks of direct deployment. The sandbox validation environment replays historical request data stored in the historical case library in an orderly manner according to timestamps. This request data is forwarded to the shadow AI agent, which then calls the shadow rule engine to process the requests using the new version of rules or models. After processing, the shadow rule engine returns the new decision result to the shadow AI agent, which then outputs the new result to the comparison unit. Simultaneously, the historical case library provides the comparison unit with the historical actual results corresponding to the request. The comparison unit performs a difference analysis between the new version's decision result and the historical actual results, identifying the differences between them. After completing the difference analysis, the comparison unit generates a quantitative verification report and feeds it back to the sandbox validation environment for relevant personnel to evaluate the effectiveness and risks of the new version, providing data support for the decision to deploy the rules or models.

[0110] In one embodiment, an artificial intelligence-based data processing system is provided, comprising: a log collection module, a deviation analysis module, an optimization strategy generation module, a rule verification and release module, and a rule version repository. This system works in conjunction with a pre-configuration module (rule configuration unit) and a business execution module (AI agent execution environment) to form a complete closed-loop optimization chain.

[0111] The log collection module is deployed in the business execution layer and is responsible for collecting multi-dimensional log data of the AI ​​agent in the business execution process in real time. The log collection module uses the session ID as an index to associate and store all log data in the log database.

[0112] The deviation analysis module is the core analysis engine for closed-loop optimization, which includes an indicator calculation unit, a deviation identification unit, and a deviation classification unit. The indicator calculation unit aggregates and calculates the collected log data to generate multi-dimensional business quality indicators. The deviation identification unit automatically identifies deviation cases based on preset deviation judgment conditions. The deviation classification unit classifies and labels the identified deviation cases.

[0113] The optimization strategy generation module automatically generates targeted rule optimization suggestions based on the classification results of deviation cases. The module outputs these suggestions in a structured "rule change proposal" format. Each proposal includes: proposal ID, a list of associated deviation case IDs, the original rule content, the suggested rule content after the change, the reason for the change (with supporting data), and the estimated scope of impact.

[0114] The rule verification and release module is a key security feature of this invention, comprising a sandbox verification unit, a version management unit, and a circuit breaker rollback unit. Before the optimization strategy is officially released, the sandbox verification unit deploys it to a sandbox environment isolated from the production environment. The version management unit maintains the rule version repository. After the optimized rules are released to the production environment, the circuit breaker rollback unit continuously monitors the operational metrics of the new rules. Once the circuit breaker condition is triggered, the circuit breaker rollback unit automatically rolls the rules back to the previous stable version and sends an alarm notification to the operations and maintenance personnel.

[0115] In one embodiment, a complete closed-loop optimization process is provided, including a data acquisition stage, a deviation analysis and identification stage, an optimization strategy generation stage, and a verification and release stage.

[0116] Specifically, the data acquisition phase includes real-time collection of business execution logs. During the execution of business tasks by the AI ​​agent, the log acquisition module collects interaction logs, filtering strategy trigger logs, interface call logs, and business result logs in real time, and stores them in association with the session ID as an index.

[0117] Specifically, the deviation analysis and identification stage includes multi-dimensional quality indicator calculation, automatic identification of deviation cases, and deviation case classification and labeling.

[0118] Specifically, the calculation of multi-dimensional quality indicators includes: the indicator calculation unit of the deviation analysis module aggregates and calculates the collected log data to generate quality indicators such as automated completion rate, manual transfer rate, policy interception rate, average number of interaction rounds, field parsing confidence, user correction frequency, and specific policy triggering frequency.

[0119] Specifically, the automatic identification of deviation cases includes: the deviation identification unit automatically filters out deviation cases based on preset deviation judgment conditions; for example: abnormally high number of interaction rounds, persistently low confidence of field parsing, abnormal fluctuations in the triggering frequency of a certain strategy, and contradictions between manual review conclusions and AI judgments.

[0120] Specifically, the deviation case classification and labeling includes: the deviation classification unit classifies the identified deviation cases and labels the deviation type (rules too strict / too lenient, AI understanding deviation, OCR error, etc.), business scenario, and severity.

[0121] Specifically, the optimization strategy generation stage includes root cause analysis of deviations, automatic generation of optimization strategies, and generation of optimization analysis reports.

[0122] Specifically, the root cause analysis of deviations includes: the optimization strategy generation module performs root cause analysis on at least one deviation case after classification; for example, if the interception rate of a certain rule is too high and the deviation is classified as "the rule is too strict", then it is determined that the threshold or condition of the rule needs to be adjusted.

[0123] Specifically, the automatic generation of optimization strategies includes: automatically generating targeted rule optimization suggestions based on the root cause analysis results, including threshold adjustments, addition of exemption conditions, field mapping optimization, and interaction strategy adjustments, and outputting structured "rule change proposals".

[0124] Specifically, generating an optimization analysis report includes integrating the results of deviation case identification, root cause analysis process, and optimization strategy recommendations into an optimization analysis report for review by business personnel or operations and maintenance personnel (optional manual confirmation step).

[0125] Specifically, the verification and release phase includes sandbox environment verification, rule version management, production environment release and monitoring, and circuit breaker protection.

[0126] Specifically, sandbox environment verification includes: deploying rule change proposals to the sandbox verification unit, performing historical deviation case replay verification, and batch normal case simulation testing. If verification passes, proceed to the next step; if verification fails, return to the automatic optimization strategy generation step to regenerate or adjust the optimization strategy.

[0127] Specifically, rule version management includes: the version management unit generates a new version number for the verified rules, records the comparison of rule content before and after the change, the reason for the change and the related deviation case ID, and stores it in the rule version repository.

[0128] Specifically, production environment deployment and monitoring include: deploying optimized rules to the rule base of the pre-configured module for subsequent loading and use during AI agent instantiation; and the circuit breaker rollback unit continuously monitoring the operational metrics of the new rules.

[0129] Specifically, the circuit breaker protection includes: if a new rule triggers the circuit breaker condition in the production environment, the circuit breaker rollback unit automatically rolls the rule back to the previous stable version and sends an alarm notification.

[0130] Thus, regarding the closed-loop self-optimization link: by establishing a complete automated link from log collection, deviation identification, policy generation to verification and release, insurance business rules can continuously evolve based on real AI agent business execution data. The technical effect is that it changes the traditional operation and maintenance model that relies on manual analysis and rule modification, shortening the rule update cycle from several days to hours. Regarding the automatic identification of multi-dimensional deviation cases: by defining multi-dimensional quantitative indicators and deviation judgment conditions such as interaction rounds, confidence levels, and policy triggering frequencies, the system achieves systematic automatic identification of deviation cases. The technical effect is that deviation discovery is transformed from "passively waiting for complaints" to "proactive quantitative monitoring," significantly improving the timeliness and coverage of optimization. Regarding the automatic generation of optimization strategies: by matching deviation classification results with preset optimization strategy templates, the system can automatically generate targeted rule optimization suggestions. The technical effect is that it lowers the professional threshold for rule optimization, allowing business personnel to participate in rule iteration without in-depth understanding of technical details. Regarding sandbox verification and circuit breaker mechanisms: by conducting sandbox replay verification before rule release and circuit breaker monitoring after release, the security of automated rule updates is ensured. The technical effect is that while enabling rapid rule iteration, it effectively prevents business risks caused by improper rule changes.

[0131] Example 1: Automatic Optimization of Underwriting Rule Thresholds An insurance company uses AI underwriting agents to process critical illness insurance applications. One rule in the underwriting rule base states: "If the applicant's age is >60, then proceed to manual underwriting." The system's workflow is as follows: Log Collection: The log collection module collected interaction logs, filtering strategy trigger logs, and business result logs for all underwriting cases within the past 30 days. Data shows that the rule was triggered a total of 500 times, of which 450 times the conclusion given by the human underwriter after being transferred to a human agent was "standard risk underwriting".

[0132] Deviation Identification: The deviation analysis module's indicator calculation unit calculated that the manual transfer rate for this rule was 90% (450 / 500), while the actual rejection / surcharge rate after manual review was only 10%. Based on this, the deviation identification unit determined that the rule had a "rule that is too strict" deviation, resulting in a large number of cases that could have been automatically underwritten being transferred to manual review, wasting human resources.

[0133] Optimization Strategy Generation: After analyzing the deviation data, the optimization strategy generation module automatically generates rule change proposals. Original rule: IF policyholder age > 60 THEN Action = Transfer to manual underwriting Change suggestion: IF policyholder age > 65 THEN Action = Transfer to manual underwriting Reason for change: For policyholders aged 60-65, the actual rejection rate during this statistical period was only 5%, which is far below the threshold for manual underwriting. It is recommended to relax the upper age limit to improve the automation rate.

[0134] Estimated impact: It is expected to reduce the number of manual underwriting cases of this type by about 40%, and the automation rate is expected to increase by 3 percentage points.

[0135] Sandbox Validation: The sandbox validation unit deployed the optimized rules to the sandbox environment and extracted 100 historical cases from the 60-65 age group for playback validation. The validation results showed that the optimized rules correctly approved 95 of the cases, while the remaining 5 cases were blocked because they were accompanied by other health risks that triggered other underwriting rules. No new risks were missed.

[0136] Release and Monitoring: After successful verification, the version management unit generates a new version v1.1 for the rule, stores it in the rule version repository, and releases it to the production environment. The circuit breaker rollback unit continuously monitors the system, and if the new rule causes an abnormal drop in the interception rate exceeding the threshold, it automatically rolls back to v1.0.

[0137] Example 2: OCR Field Mapping Optimization Scenarios An insurance company uses an AI-powered order processing agent to handle car insurance applications. The system needs to identify the "vehicle brand" field from the photos of the vehicle registration certificates uploaded by users. The system's operation process is as follows: Log collection: The average confidence level of the "Vehicle Brand" field collected by the log collection module is only 0.52, and users frequently correct errors. Many users need to manually correct the vehicle brand name recognized by AI after uploading their vehicle registration certificates.

[0138] Deviation Identification: The deviation analysis module identified that the resolution confidence level of the "Vehicle Brand" field was consistently below the threshold, and the deviation was classified as "OCR recognition error / field mapping missing". Further analysis revealed that the recognition errors were mainly concentrated on a few niche domestic brands, and the text recognized by the OCR differed from the standard brand name used in the insurance company's internal system (for example, the OCR recognized it as "BYD Song PLUS", while the system's standard name was "BYD Song PLUS DM-i").

[0139] Optimization Strategy Generation: The optimization strategy generation module automatically generates field mapping optimization suggestions. Original mapping: mapping without aliases Proposed change: Add an alias field mapping for "BYD Song PLUS" ["BYD Song PLUS DM-i", "Song PLUS DM-i"]. Reason for change: This vehicle model was incorrectly identified 47 times in the past 30 days, accounting for 68% of the total number of identification errors for the BYD brand.

[0140] Sandbox Validation: The sandbox validation unit uses historical images containing the vehicle model for playback validation. The optimized mapping rules successfully matched the OCR-recognized "BYD Song PLUS" to the correct system standard name, improving the resolution confidence to 0.94.

[0141] Release and Results: After the optimized field mapping rules were released and went live, the average parsing confidence of the field increased from 0.52 to 0.81, and the frequency of user corrections decreased to 0.4 times per case.

[0142] Example 3: Dynamic Optimization Scenario of Interaction Strategy An insurance company uses AI-powered claims agents to collect accident information. The system found that during the collection of the "Accident Details" field, the average number of user interactions was as high as 6.2 rounds, far exceeding the 2.1 rounds for other fields.

[0143] Deviation Identification: The deviation analysis module identified an interaction efficiency deviation in the "Accident Details" field. Further analysis revealed that when users described the accident using open-ended questions, their descriptions were often too brief (e.g., "I was hit," "I was rear-ended"), causing the AI ​​agent to repeatedly ask for details (time, location, liability division, etc.).

[0144] Optimization Strategy Generation: The optimization strategy generation module generates interaction strategy adjustment suggestions: Change the collection method for this field from "open-ended questions and answers" to "structured guided questions and answers." This means breaking down an open question into several specific questions, guiding the user to answer them sequentially. Q1: What was the exact time of the accident? Q2: Where exactly did the accident occur? Q3: What responsibility do you bear in this accident? (A. Full responsibility B. Primary responsibility C. Shared responsibility D. Secondary responsibility E. No responsibility) Q4: Please briefly describe the course of the accident.

[0145] Verification and Release: The sandbox verification unit conducted A / B testing verification. After adopting the new strategy, the average number of collection rounds for the "Experience Process" field decreased from 6.2 rounds to 3.8 rounds, and the user completion rate increased from 78% to 92%.

[0146] Release and Results: After the optimized interaction strategy was released and went live, the overall automation completion rate of claims cases increased by approximately 5%.

[0147] It should be understood that, although Figures 1-3 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figures 1-3 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0148] In one embodiment, such as Figure 4 As shown, an artificial intelligence-based data processing device is provided. The device includes: a first processing module 401, used to collect log data in real time during the execution of business cases to obtain multi-dimensional log data; a second processing module 402, used to perform deviation analysis on the multi-dimensional log data through a dynamic threshold model to identify at least one deviation case; a third processing module 403, used to classify at least one deviation case according to classification rules; a fourth processing module 404, used to perform root cause analysis on at least one deviation case based on the classification of at least one deviation case to identify the root cause of deviation in at least one deviation case; a fifth processing module 405, used to determine a target case screening strategy based on at least one deviation case and its corresponding root cause; and a sixth processing module 406, used to screen the business cases to be processed according to the target case screening strategy to identify target business cases.

[0149] As a preferred implementation, in this embodiment of the application, the first processing module 401 is specifically used to: collect at least one of the interaction logs, interface call logs, business result logs and filtering strategy trigger logs during the business execution process in real time, and obtain multi-dimensional log data; associate the multi-dimensional log data with the corresponding case identifier and store it in the log database.

[0150] In a preferred implementation, in this embodiment, the second processing module 402 is specifically used for: obtaining case data corresponding to each case from multi-dimensional log data based on the case identifier; calculating business quality indicators for each case data, wherein the business quality indicators include at least one of automation completion rate, manual transfer rate, policy interception rate, average confidence level of field parsing, user correction frequency, average interaction rounds, and policy triggering frequency; determining the business quality indicators through a dynamic threshold model; and determining the case corresponding to the case data as a deviation case in response to the existence of a business quality indicator that meets the deviation condition.

[0151] In a preferred implementation, in this embodiment of the application, the third processing module 403 is specifically used to: analyze the deviation cases according to the first classification rule to determine the deviation type of the deviation cases; analyze the deviation cases according to the second classification rule to determine the business scenario type of the deviation cases; and analyze the deviation cases according to the third classification rule to determine the severity of the deviation cases.

[0152] In a preferred implementation, in this embodiment, the fourth processing module 404 is specifically used for: obtaining the deviation type classification, business scenario classification, and severity classification of deviation cases; determining a local subgraph in a pre-constructed strategy graph based on the deviation type classification, business scenario classification, and severity classification; wherein the strategy graph contains multiple types of nodes and relational edges connecting nodes, and the nodes include at least root cause nodes, evidence nodes, business scenario nodes, deviation type nodes, and severity nodes; determining an initial candidate root cause hypothesis set based on the deviation type nodes, root cause nodes, and local subgraph; supplementing or removing the initial candidate root cause hypothesis set according to the scenario rules associated with the business scenario nodes, and adjusting the prior probabilities of the candidate root causes to obtain a scenario-based root cause hypothesis set; and determining the analysis depth and evidence threshold of the root cause analysis according to the severity classification, wherein the analysis depth corresponds to... The number of hops in the causal walk is determined within the strategy graph, with the evidence threshold corresponding to the minimum requirement for rule strength. Based on the scenario-based root cause hypothesis set and analysis depth, a limited-depth walk is performed along causal edges and evidence chain edges, centered on each candidate root cause, to extract reachable evidence nodes and construct a local inference subgraph. Here, causal edges represent the relationship between root causes and evidence, and evidence chain edges connect evidence nodes to evidence nodes, representing the prerequisite order for obtaining evidence. In the local inference subgraph, the information gain of uncollected evidence nodes is calculated, and the evidence with the largest information gain is collected. Based on the collected evidence, backpropagation is performed along the causal edges to update the posterior probability of each candidate root cause, and the posterior probability is adjusted according to the cooperative or inhibitory relationship of the relation edges. The above collection and update steps are repeated until the termination condition is met, and the candidate root cause that meets the evidence threshold and has the largest posterior probability is output as the deviation root cause of the deviation case.

[0153] In a preferred implementation, in this embodiment, the fifth processing module 405 is specifically used for: determining the case selection strategy to be updated from a preset strategy library based on the root cause of the deviation; obtaining the historical execution dataset associated with the case selection strategy to be updated and the update type corresponding to the root cause of the deviation; determining the strategy parameters to be updated based on the update type and generating at least one candidate parameter value; inputting at least one candidate parameter value into the historical execution dataset, evaluating the effect of each candidate parameter value, and obtaining an evaluation index; selecting the optimal parameter value from the candidate parameter values ​​based on the evaluation index and preset business constraints, and generating an update suggestion containing the optimal parameter value; updating the case selection strategy to be updated according to the update suggestion to obtain the target case selection strategy.

[0154] In a preferred embodiment of this application, the device further includes a seventh processing module, which is specifically used for: deploying the target case screening strategy to the sandbox verification unit and performing replay verification and simulation testing; generating a corresponding strategy version number and storing it in the strategy version library in response to the target case screening strategy passing verification; publishing the target case screening strategy to the strategy version library and starting the circuit breaker rollback unit to continuously monitor the operating indicators of the target case screening strategy; during the monitoring process, if the operating indicators of the target case screening strategy are detected to trigger the preset circuit breaker condition, the target case screening strategy is rolled back to the previous stable version through the circuit breaker rollback unit, and an alarm notification is sent.

[0155] Specific limitations regarding AI-based data processing devices can be found in the limitations of AI-based data processing methods described above, and will not be repeated here. Each module in the aforementioned AI-based data processing device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0156] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 5As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements an artificial intelligence-based data processing method. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0157] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0158] Embodiments of this application also provide a computer-readable storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above-described embodiments of the artificial intelligence-based data processing method.

[0159] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.

[0160] Embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the artificial intelligence-based data processing method.

[0161] Embodiments of this application also provide another computer program product, including a non-volatile computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in any of the above-described embodiments of the artificial intelligence-based data processing method.

[0162] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0163] The foregoing has provided a detailed description of a data processing method, apparatus, device, and readable storage medium based on artificial intelligence provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only intended to aid in understanding the method and core ideas of this application. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of the claims of this application.

Claims

1. A data processing method based on artificial intelligence, characterized in that, The method includes: Real-time collection of log data during business case execution to obtain multi-dimensional log data; By performing deviation analysis on the multi-dimensional log data using a dynamic threshold model, at least one deviation case is identified. According to the classification rules, the at least one deviation case is classified; Based on the classification of the at least one deviation case, root cause analysis is performed on the at least one deviation case to determine the root cause of the deviation in the at least one deviation case; Based on the at least one deviation case and the corresponding root cause of the deviation, a target case selection strategy is determined; Based on the target case screening strategy, the business cases to be processed are screened to determine the target business cases; The step of determining the target case selection strategy based on the at least one deviation case and the corresponding root cause of deviation includes: Based on the root causes of the deviation, determine the case selection strategy to be updated from the preset strategy library; Obtain the historical execution dataset associated with the case filtering strategy to be updated and the update type corresponding to the root cause of the deviation; Based on the update type, determine the strategy parameters to be updated and generate at least one candidate parameter value; Input the at least one candidate parameter value into the historical execution dataset, evaluate the effect of each candidate parameter value, and obtain the evaluation index; Based on the evaluation indicators and preset business constraints, the optimal parameter value is selected from the candidate parameter values, and an update suggestion containing the optimal parameter value is generated. The update suggestion includes threshold adjustment, addition of exemption conditions, optimization of field mapping, and adjustment of interaction strategy. Based on the update suggestion, the case selection strategy to be updated is updated to obtain the target case selection strategy; The step of performing deviation analysis on the multi-dimensional log data using a dynamic threshold model to identify at least one deviation case includes: Based on the case identifier, obtain the case data corresponding to each case from the multi-dimensional log data; Calculate the business quality indicators for each case data, wherein the business quality indicators include at least one of the following: automation completion rate, manual transfer rate, policy interception rate, average confidence of field parsing, user correction frequency, average number of interaction rounds, and policy triggering frequency. The business quality indicators are determined using the dynamic threshold model. In response to the existence of a business quality indicator that meets the deviation condition, the case corresponding to the case data is determined to be a deviation case.

2. The data processing method based on artificial intelligence according to claim 1, characterized in that, The real-time acquisition of log data during the execution of business cases yields multi-dimensional log data, including: Collect at least one of the following during the execution of business cases: interaction logs, interface call logs, business result logs, and filtering strategy trigger logs, to obtain multi-dimensional log data; The multi-dimensional log data and corresponding case identifiers are associated and stored in the log database.

3. The data processing method based on artificial intelligence according to claim 1, characterized in that, The classification of the at least one deviation case according to the classification rules includes: Based on the first classification rule, the deviation cases are analyzed to determine the deviation type of the deviation cases; Based on the second classification rule, the deviation cases are analyzed to determine the business scenario type of the deviation cases; The deviation cases are analyzed according to the third classification rule to determine the severity of the deviation cases.

4. The data processing method based on artificial intelligence according to claim 3, characterized in that, The step of combining the classification of the at least one deviation case with the root cause analysis of the at least one deviation case to determine the root cause of the deviation in the at least one deviation case includes: Obtain the deviation type classification, business scenario classification, and severity classification of the deviation cases; Based on the deviation type classification, the business scenario classification, and the severity classification, a local subgraph is determined in the pre-constructed strategy graph; wherein the strategy graph contains multiple types of nodes and relational edges connecting the nodes, and the nodes include at least root cause nodes, evidence nodes, business scenario nodes, deviation type nodes, and severity nodes; Based on the deviation type nodes, the root cause nodes, and the local subgraph, an initial set of candidate root cause hypotheses is determined; Based on the scenario rules associated with the business scenario nodes, the initial candidate root cause hypothesis set is supplemented or removed, and the prior probabilities of the candidate root causes are adjusted to obtain a scenario-based root cause hypothesis set. Based on the severity classification, the analytical depth and evidence threshold of the root cause analysis are determined, wherein the analytical depth corresponds to the number of hops in the causal walk in the policy graph, and the evidence threshold corresponds to the minimum requirement for rule strength. Based on the scenario-based root cause hypothesis set and the analysis depth, with each candidate root cause as the center, a limited-depth walk is performed along the causal edge and the evidence chain edge to extract reachable evidence nodes and construct a local reasoning subgraph, wherein the causal edge represents the relationship between the root cause and the evidence, and the evidence chain edge connects the evidence nodes to the evidence nodes, representing the prerequisite order for obtaining the evidence. In the local reasoning subgraph, the information gain of the uncollected evidence nodes is calculated, and the evidence with the largest information gain is collected. Based on the collected evidence, reverse probability propagation is performed along the causal edge to update the posterior probability of each candidate root cause, and the posterior probability is adjusted according to the cooperative or inhibitory relationship of the relation edge. Repeat the above collection and update steps until the termination condition is met, and output the candidate root cause that satisfies the evidence threshold and has the highest posterior probability as the root cause of the biased case.

5. The data processing method based on artificial intelligence according to claim 1, characterized in that, The method further includes: The target case selection strategy is deployed to the sandbox verification unit to perform replay verification and simulation testing. In response to the target case screening strategy being verified, a corresponding strategy version number is generated and stored in the strategy version library; The target case selection strategy is published to the strategy version repository, and the circuit breaker rollback unit is activated to continuously monitor the operation metrics of the target case selection strategy. During the monitoring process, if the operation indicators of the target case screening strategy trigger the preset circuit breaker condition, the target case screening strategy will be rolled back to the previous stable version through the circuit breaker rollback unit, and an alarm notification will be sent.

6. An artificial intelligence-based data processing apparatus for implementing the artificial intelligence-based data processing method as described in any one of claims 1-5, characterized in that, The device includes: The first processing module is used to collect log data in real time during the execution of business cases to obtain multi-dimensional log data; The second processing module is used to perform deviation analysis on the multi-dimensional log data through a dynamic threshold model to determine at least one deviation case. The third processing module is used to classify the at least one deviation case according to the classification rules; The fourth processing module is used to perform root cause analysis on the at least one deviation case based on the classification of the at least one deviation case, and to determine the root cause of the deviation in the at least one deviation case. The fifth processing module is used to determine the target case screening strategy based on the at least one deviation case and the corresponding root cause of deviation. The sixth processing module is used to filter the business cases to be processed according to the target case filtering strategy and determine the target business cases.

7. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images