Service policy generation and optimization method and device, equipment and medium

Abstract

The invention relates to the technical field of artificial intelligence, can be applied to business scenes such as financial science and technology and medical health, and discloses a service strategy generation and optimization method, device, equipment and medium, and the method comprises the steps: collecting multi-source interaction data to generate a multi-dimensional dynamic portrait, and obtaining an initial service strategy based on preset strategy set matching, and generating a to-be-executed service strategy in combination with an action combination output by the value network, constructing abnormal scene test data by using the generative adversarial network to complete verification and adjustment, obtaining a final service strategy, recording portrait change data and service indexes by executing the final service strategy, and feeding back the portrait change data and the service indexes to the value network and the portrait construction model. Through data-driven dynamic portrait modeling and strategy closed-loop optimization, the defects that in the prior art, a static rule is relied on, and different requirements of clients are difficult to adapt are overcome, dynamic evolution of the client service strategy is achieved, the real-time performance of the service and the resource configuration efficiency are improved, and the client experience is remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a service strategy generation and optimization method, device, equipment and medium. BACKGROUND

[0002] In the field of financial technology business, the business scenarios represented by insurance customer service are facing multiple challenges. The traditional insurance service system mainly relies on static rule base and manual experience to handle claims, insurance and customer consultation and other transactions, lacking dynamic recognition ability for the differentiated characteristics of customer groups. In the face of different customer groups, such as elderly customers with insufficient digital capabilities and high net worth customers with strong demand for customized protection, existing technologies often use uniform service strategies, which are difficult to adapt to individual needs. In addition, there is a fragmentation phenomenon in insurance service channels, and there is a lack of linkage between online insurance, telephone customer service and offline counter service, and customers may experience inconsistent problems during cross-channel service. Although existing intelligent customer service systems have been applied in some insurance companies, most of them are limited to simple question and answer in a single channel, lacking real-time comprehensive analysis of risk preference, emotional state and channel habit, resulting in lack of flexibility in service strategy, inability to dynamically optimize and evolve, and ultimately leading to low service efficiency, uneven resource allocation and declining customer satisfaction.

[0003] In the field of medical and health business, the dynamic nature of user demand is more prominent, for example, the health indicators of patients with chronic diseases fluctuate continuously during long-term management, and patients with acute conditions need quick response and emergency intervention. Most existing medical service support systems are based on fixed processes and manual intervention, lacking dynamic updating ability for user profiling, making it difficult to capture risk changes and psychological state in a timely manner. In multi-channel health services such as online consultation, remote monitoring, telephone follow-up, etc., existing technologies lack a unified coordination mechanism, often leading to service fragmentation and response delay. At the same time, traditional systems cannot effectively combine historical service records and real-time interaction data, lack comprehensive analysis and dynamic adaptation of patient needs and risks, and are difficult to form continuous and accurate service decisions. These deficiencies make medical and health services have obvious defects in personalization and real-time, making it difficult to meet the multi-dimensional needs of users in complex health scenarios. SUMMARY

[0004] The main purpose of the present application is to provide a service strategy generation and optimization method, device, equipment and storage medium, aiming to solve the technical problems that the existing technology cannot construct a dynamically updated customer portrait based on multi-source interaction data and form a closed-loop optimization with the service strategy, resulting in static service strategy, cross-channel fragmentation, and difficulty in real-time adaptation to customer multi-dimensional needs and state changes.

[0005] To achieve the above purpose, the present application provides a service strategy generation and optimization method, comprising:

[0006] collecting raw interaction data from multiple interaction channels, generating a multi-dimensional dynamic portrait based on the raw interaction data through a portrait construction model;

[0007] loading a preset strategy set and generating a pre-grouping strategy set through a clustering module, matching the multi-dimensional dynamic portrait with the pre-grouping strategy set to obtain an initial service strategy;

[0008] combining the multi-dimensional dynamic portrait with service stage identification information to generate a composite state vector, inputting the composite state vector into a pre-trained value network to obtain an action combination;

[0009] optimizing the initial service strategy through the action combination to generate a to-be-executed service strategy;

[0010] constructing abnormal scenario test data through a generative adversarial network, executing the to-be-executed service strategy to process the abnormal scenario test data to generate a verification result;

[0011] adjusting the to-be-executed service strategy based on the verification result to obtain a final service strategy;

[0012] executing the final service strategy and recording portrait change data and service indicators, and feeding back the portrait change data and service indicators to the value network and the portrait construction model.

[0013] Further, to achieve the above-mentioned purpose, the present application provides a service strategy generation and optimization device, comprising:

[0014] a portrait construction module for collecting raw interaction data from multiple interaction channels, generating a multi-dimensional dynamic portrait based on the raw interaction data through a portrait construction model;

[0015] a strategy clustering module for loading a preset strategy set and generating a pre-grouping strategy set through a clustering module, matching the multi-dimensional dynamic portrait with the pre-grouping strategy set to obtain an initial service strategy;

[0016] a value network module for combining the multi-dimensional dynamic portrait with service stage identification information to generate a composite state vector, inputting the composite state vector into a pre-trained value network to obtain an action combination;

[0017] a strategy optimization module for optimizing the initial service strategy through the action combination to generate a to-be-executed service strategy;

[0018] an abnormal verification module for constructing abnormal scenario test data through a generative adversarial network, executing the to-be-executed service strategy to process the abnormal scenario test data to generate a verification result;

[0019] A defense enhancement module is configured to adjust the service policy to be executed based on the verification result to obtain a final service policy.

[0020] A feedback updating module is configured to execute the final service policy and record portrait change data and service indicators, and feed back the portrait change data and service indicators to the value network and the portrait construction model.

[0021] Further, to achieve the above object, the present application also provides a computer device, which comprises a memory, a processor, and a service policy generation and optimization program stored in the memory and executable on the processor, and the service policy generation and optimization program, when executed by the processor, implements the steps of the service policy generation and optimization method as described above.

[0022] Further, to achieve the above object, the present application also provides a computer readable storage medium, which stores a service policy generation and optimization program, and the service policy generation and optimization program, when executed by a processor, implements the steps of the service policy generation and optimization method as described above.

[0023] Beneficial effects: The present application relates to the field of artificial intelligence technology, and can be applied to business scenarios such as financial technology and medical health, and discloses a service policy generation and optimization method, device, equipment and medium, which comprises the following steps: collecting original interaction data from multiple interaction channels, generating a multi-dimensional dynamic portrait based on the original interaction data through a portrait construction model; loading a preset policy set and generating a pre-grouping policy set through a clustering module, matching the multi-dimensional dynamic portrait with the pre-grouping policy set to obtain an initial service policy; combining the multi-dimensional dynamic portrait with service stage identification information to generate a composite state vector, inputting the composite state vector into a pre-trained value network to obtain an action combination; optimizing the initial service policy through the action combination to generate a service policy to be executed; constructing abnormal scenario test data through a generative adversarial network, executing the service policy to be executed to process the abnormal scenario test data to generate a verification result; adjusting the service policy to be executed based on the verification result to obtain a final service policy; executing the final service policy and recording portrait change data and service indicators, and feeding back the portrait change data and service indicators to the value network and the portrait construction model. The present application generates a multi-dimensional dynamic portrait by fusing multi-source interaction data, dynamically optimizes the initial service policy in combination with the action combination output by the value network, constructs test data through a generative adversarial network in an abnormal scenario to complete policy verification and adjustment, and finally realizes closed-loop evolution of the service policy. The method can overcome the deficiencies of the prior art that the service policy relies on static rules and lacks cross-channel adaptation and dynamic updating, thereby improving the real-time adaptability and stability of customer service, significantly improving customer experience and improving service resource utilization efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0024] The application will be further described below in conjunction with the accompanying drawings and embodiments. In the drawings:

[0025] Figure 1 A schematic diagram of an application environment of the service policy generation and optimization method in an embodiment of the application;

[0026] Figure 2 A schematic diagram of a flow of the service policy generation and optimization method in an embodiment of the application;

[0027] Figure 3 A schematic diagram of functional modules of the service policy generation and optimization device in a preferred embodiment of the application;

[0028] Figure 4 A schematic diagram of a structure of a computer device in an embodiment of the application;

[0029] Figure 5 A schematic diagram of another structure of a computer device in an embodiment of the application. DETAILED DESCRIPTION

[0030] It should be understood that the specific embodiments described herein are merely illustrative of the application and do not limit the application.

[0031] The service policy generation and optimization method provided by the embodiments of the application can be applied in, for example, Figure 1In an application environment of the present application, a client communicates with a server through a network. The server can collect original interaction data from multiple interaction channels through the client, generate a multi-dimensional dynamic portrait based on the original interaction data through a portrait construction model, load a preset strategy set and generate a pre-grouping strategy set through a clustering module, match the multi-dimensional dynamic portrait with the pre-grouping strategy set to obtain an initial service strategy, combine the multi-dimensional dynamic portrait with service stage identification information to generate a composite state vector, input the composite state vector into a pre-trained value network to obtain an action combination, optimize the initial service strategy through the action combination to generate a to-be-executed service strategy, construct abnormal scene test data through a generative adversarial network, execute the to-be-executed service strategy to process the abnormal scene test data to generate a verification result, adjust the to-be-executed service strategy based on the verification result to obtain a final service strategy, execute the final service strategy and record portrait change data and service indicators, and feed back the portrait change data and the service indicators to the value network and the portrait construction model. The present application generates a multi-dimensional dynamic portrait by fusing multi-source interaction data, dynamically optimizes the initial service strategy in combination with the action combination output by the value network, constructs test data in an abnormal scene through the generative adversarial network to complete strategy verification and adjustment, and finally realizes closed-loop evolution of the service strategy. This method can overcome the deficiencies of the prior art that the service strategy relies on static rules and lacks cross-channel adaptation and dynamic updating, thereby improving the real-time adaptability and stability of customer service, significantly improving customer experience and improving service resource utilization efficiency. The client can be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers and portable wearable devices. The server can be implemented by an independent server or a server cluster composed of multiple servers. The present application will be described in detail below through specific embodiments.

[0032] Please refer to Figure 2 , Figure 2 The flowchart of an embodiment of the service strategy generation and optimization method provided by the present application is shown in the figure. It should be noted that although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0033] As Figure 2 shown, the service strategy generation and optimization method provided by the present application includes the following steps:

[0034] S10, collecting original interaction data from multiple interaction channels, generating a multi-dimensional dynamic portrait based on the original interaction data through a portrait construction model;

[0035] In this embodiment, the step of collecting raw interaction data from multiple interaction channels involves parallel acquisition and synchronous processing of multi-source data. The types of interaction channels typically include mobile applications, web terminals, voice channels, image upload portals, and historical databases. Mobile application data is mainly obtained through embedded point logs, including user interface click events, sliding tracks, and input behaviors, which can reflect user operation habits and interaction intensity. Web terminal data relies more on browser interaction records and form submission information to supplement the service path of users in different terminal environments. Voice channel data acquisition relies on voice call recordings and speech recognition results, which can capture identity features and semantic content through voiceprint spectrum features and translated text. Image upload portals provide unstructured image information, which can be converted into structured features by combining optical character recognition methods. Service records in the historical database often contain timestamp, service type, and processing conclusion information, which can be standardized to form behavior time series data sequences.

[0036] The multi-dimensional dynamic portrait is generated based on the raw interaction data through the portrait construction model. The focus of this step is to extract features from different types of data and perform cross-modal fusion. Long short-term memory networks can model time series data such as interaction logs and historical service records, identify abnormal behavior patterns, and generate abnormal preference scores. Gradient boosting decision trees are suitable for processing structured numerical and categorical data, such as device type and channel usage frequency. The results obtained through training can represent the channel preference probability distribution. The bidirectional encoder representation module can combine the translated text and voiceprint features of speech recognition to extract emotion state labels in forward and backward semantic understanding. Mapping the above features from different sources to a unified representation space can form a multi-dimensional dynamic portrait that contains both preference tendency quantitative indicators and discrete features of emotional and behavioral habits.

[0037] This process emphasizes consistent processing of cross-channel features. For example, user behavior in mobile applications and voice channels may differ, requiring timestamp alignment, feature standardization, and normalization to avoid conflicts or shifts in the portrait. The generation of multi-dimensional dynamic portrait is not a one-time static output, but relies on continuous data stream updates, allowing the portrait to reflect changes in user status and preferences over time.

[0038] In actual implementation, the interactive channel data collection method can be adjusted according to different business needs. The mobile terminal application can upload user interaction events in real time through the log collection module embedded by the SDK, and the web page can collect form submission and page access path through the JavaScript listener. The voice channel data can be obtained through the recording and real-time speech recognition system integrated in the call center, and the voiceprint features are extracted by short-time Fourier transform or mel frequency cepstral coefficient. The image upload channel can deploy an edge OCR engine or call a remote recognition interface for real-time processing, and the recognition result is cleaned by a regular expression to ensure the consistency of the field format. The data in the historical database is usually large in size, and distributed collection and batch processing methods can be used to extract time series features in combination with sliding window technology.

[0039] In the portrait construction process, different modeling modules can be combined or replaced. For example, in time series data analysis, a gated recurrent unit can be used instead of a long short-term memory network to improve computational efficiency. In structured data analysis, in addition to gradient boosting decision trees, random forests or LightGBM models can also be used to improve training speed. In sentiment analysis, in addition to bidirectional encoders, sequence models based on Transformers can also be used to improve the ability to identify emotions in long sentences. To improve the real-time performance of the portrait update, a streaming computing framework can be introduced to ensure that the portrait is updated within milliseconds after the data arrives.

[0040] The embodiment can solve the problem of incomplete and lagging customer portraits caused by relying on only a single channel or static data in existing systems by collecting multi-source interaction data and generating multi-dimensional dynamic portraits. This approach enables the portrait to comprehensively cover user operation behavior, channel selection, and emotional state, and to remain dynamic as data updates. This can provide real, rich, and real-time user feature support for subsequent service strategy generation, thereby improving the consistency of service resource allocation and customer experience.

[0041] S20, loading a preset strategy set and generating a pre-grouping strategy set through a clustering module, matching the multi-dimensional dynamic portrait with the pre-grouping strategy set to obtain an initial service strategy;

[0042] In this embodiment, the operation of loading the preset strategy set is to call and parse the configured service strategy templates, which often include channel configuration, service resource ratio, and emotional coping methods. The establishment of the preset strategy set is derived from long-term accumulation of customer interaction data and service experience, such as claim response templates in the insurance business scenario and follow-up reminder templates in the health management scenario. The loading process of this set is usually completed through a strategy database calling interface to ensure that the strategy templates can quickly enter the calculation link.

[0043] The operation of generating the pre-grouping strategy set by the clustering module depends on the multi-dimensional feature extraction and vectorization representation of the strategy template. The features in the strategy template mainly include abnormal sensitivity parameters, channel adaptation types, emotional response modes, etc. The abnormal sensitivity parameters can be derived from the failure rate measurement of the strategy in the burst scenario in the historical test; the channel adaptation type is related to the service touch point corresponding to the strategy, such as mobile terminal priority or telephone priority; the emotional response mode is the response logic of the template when facing emotional customers. Convert these features into vectors and input them into the clustering algorithm (such as K-means or hierarchical clustering) to generate multiple strategy base groups, each base group containing several strategy templates with similar features, and generating a centroid coordinate as the representative feature of the class.

[0044] The process of matching the multi-dimensional dynamic portrait with the pre-grouping strategy set relies on similarity calculation. The multi-dimensional dynamic portrait already contains abnormal preference scores, channel preference probability distribution and emotional state labels, which need to be converted into feature vector form consistent with the centroid of the strategy base group. The common way is to convert the preference score into an abnormal feature value, the channel distribution into a probability vector, and the emotional label into a numerical feature. Then, the distance between the multi-dimensional dynamic portrait and each strategy base group is calculated by cosine similarity or Euclidean distance, and the closest strategy base group is selected as the target grouping. The strategy templates in the target grouping are further combined with the service scenario complexity and the adaptation score to select suitable strategies as the initial service strategy.

[0045] In actual implementation, the loading method of the pre-set strategy set can be adjusted according to different environments, which can be through a distributed strategy library to ensure fast calling, or through local caching to reduce delay. In the implementation of the clustering module, K-means clustering can be used to improve efficiency, or hierarchical clustering can be used to obtain a hierarchical strategy grouping structure. In similarity calculation, cosine similarity can be used to ensure the unified processing of different dimensional features, or Mahalanobis distance can be used to introduce a covariance matrix to enhance classification accuracy. In the strategy selection link, dynamic weights can be set, and for scenarios with high abnormal feature value weights, strategy templates with high fault tolerance are preferred.

[0046] By loading the pre-set strategy set and clustering grouping, the embodiment can avoid the blindness and inefficiency in the strategy template selection process. With clustering, the strategy base group matching the customer portrait can be quickly identified, thereby narrowing the candidate range and selecting a more suitable strategy in combination with the service complexity. This not only improves the efficiency of strategy matching, but also ensures the rationality of the initial service strategy in terms of situational adaptability and customer experience, providing higher quality input for subsequent optimization steps.

[0047] S30, combine the multi-dimensional dynamic portrait and the service stage identification information to generate a composite state vector, input the composite state vector into a pre-trained value network to obtain an action combination;

[0048] In the embodiment, the operation of combining the multi-dimensional dynamic portrait and the service stage identification information is to form a composite input vector capable of completely representing the customer state and the service process. The multi-dimensional dynamic portrait contains the abnormal preference score, the channel preference probability distribution and the emotion state label, which correspond to the behavioral risk tendency, the channel usage habit and the emotional expression characteristics of the customer respectively. The service stage identification information is derived from the business process engine and is usually represented as the node state in the service link, such as material collection, information review, payment confirmation in the insurance claim process, or health questionnaire filling, remote diagnosis and drug reminder in the medical health management. These identification information is parsed into numerical stage codes to depict the position of the service in the time process.

[0049] The generation process of the composite state vector includes feature extraction, splicing and standardization. First, the abnormal preference score in the multi-dimensional dynamic portrait is extracted as a numerical feature value; the channel preference probability distribution is extracted and represented as a vector form; the emotion state label is converted into a numerical code, such as realizing the dense vector representation of the emotion category through an embedding layer. Then, the service stage identification code is extracted to convert the process node state into a fixed-dimension stage code. After combining these features, the original composite vector is formed. In order to avoid the imbalance influence between different dimension features, the original composite vector is standardized to make it comparable in the numerical range, so as to ensure the stable training and reasoning of the subsequent value network.

[0050] When the composite state vector is input into the pre-trained value network, the neural network structure calculates the value expectation of each possible action through the forward propagation process. The value network is usually composed of multiple layers of neural networks, the input layer accepts the composite state vector, the intermediate layer contains nonlinear transformation and feature interaction mechanism, and the output layer gives the action value distribution. The value distribution of the action measures the expected return that may be brought by taking different actions in the current state, and the optimal action combination is selected from the actions with the maximum value expectation. The action combination usually includes channel switching, information pushing and resource allocation and other strategy instructions.

[0051] In different embodiments, the encoding of the service stage identification information can adopt a one-hot encoding manner or an embedding vector manner. The one-hot encoding can guarantee the uniqueness of the stage information, and the embedding vector can obtain the semantic distance between stages through training. For the standardization of the composite state vector, zero-mean normalization or maximum-minimum scaling can be used to adapt to different data distributions. On the value network structure, a multi-layer perceptron can be used, or a deep network combined with an attention mechanism can be used to highlight the role of key features in action value calculation. On the selection method of action combination, a greedy strategy can be used to directly select the action with the maximum value, or an epsilon-greedy or soft sampling strategy can be used to balance exploration and utilization.

[0052] The present embodiment can form a composite vector that can reflect the customer's personalized state and service process by combining multi-dimensional dynamic portraits and service stage identification information. After inputting the composite vector into the value network, an action combination highly related to customer behavior, emotion and channel habit can be generated, thereby providing a basis for optimizing the initial strategy. Compared with the traditional method relying on single-dimensional input, this composite input can more accurately depict the real needs of customers at a specific stage, thereby improving the pertinence and adaptability of subsequent service strategies.

[0053] S40, optimizing the initial service strategy by the action combination to generate a to-be-executed service strategy;

[0054] In the present embodiment, the initial service strategy is optimized by introducing the instruction information output by the value network based on the existing strategy, thereby generating a to-be-executed service strategy that can be directly applied. The initial service strategy includes service channel configuration, service content sequence and service resource ratio, and the action combination includes channel switching instruction, information push instruction and resource allocation instruction. The optimization process first analyzes the action combination and disassembles it into specific control instructions that can be directly applied to strategy update. The channel switching instruction corresponds to the service channel configuration, the information push instruction corresponds to the service content sequence, and the resource allocation instruction corresponds to the service resource ratio.

[0055] After the analysis is completed, the initial service strategy needs to be structured and processed to decouple the three elements for separate updating. For the channel switching instruction, the service channel configuration is updated to migrate the customer service from the current channel to a more suitable channel, such as from the mobile end to the artificial customer service or from the telephone to the online self-service platform. For the information push instruction, specific prompt content or notification information is inserted in the service content sequence, so that the service can respond to customer needs and emotional state in a timely manner. For the resource allocation instruction, the existing service resource ratio is adjusted, such as increasing the proportion of artificial customer service or reducing the consumption of computing resources for unnecessary tasks.

[0056] After the update of the three elements, the updated service channel configuration, information push content and resource allocation need to be re-integrated to form a whole strategy structure. In order to ensure the effectiveness and stability of the strategy execution, the execution time constraint condition is introduced in the integrated strategy, which limits the effective time period and trigger condition of the strategy, preventing outdated or invalid strategies from entering the execution phase. Finally, the formed to-be-executed service strategy is stored in the strategy execution queue to facilitate subsequent policy engine scheduling execution.

[0057] In different embodiments, the implementation of the channel switching instruction can adopt a static mapping table or a dynamic routing system. The static mapping table is suitable for scenarios with limited and infrequently changing channels, while the dynamic routing system can dynamically adjust the channel path based on real-time load and customer portrait. The implementation of the information push instruction can adopt templated push or adaptive push based on natural language generation, the former is convenient for rapid deployment, and the latter can achieve a higher degree of personalization. The resource allocation instruction can be implemented through fixed proportion adjustment or dynamically adjusted through real-time monitoring feedback mechanism, such as automatically allocating resources based on CPU utilization and customer waiting time. The execution time constraint condition can adopt time window constraint or event trigger-based constraint, such as policy taking effect within 24 hours after the customer submits a claim request.

[0058] This embodiment can dynamically adjust the channel configuration, information push and resource allocation in the customer service process by introducing action combination into the optimization process of the initial service strategy, thereby generating an executable service strategy. Compared with the traditional method relying on fixed rules, this optimization process can realize real-time adaptation, ensuring that the strategy can be continuously adjusted according to the customer portrait and interaction state. This not only improves the pertinence of strategy execution, but also improves the efficiency of service resource utilization, reduces unnecessary consumption, and enhances the coherence and timeliness of customer experience.

[0059] S50, constructing abnormal scene test data through a generative adversarial network, executing the to-be-executed service strategy to process the abnormal scene test data, and generating a verification result;

[0060] In this embodiment, the abnormal scenario test data is constructed by a generative adversarial network, which is composed of a generator and a discriminator. The generator is responsible for synthesizing data with abnormal characteristics, and the discriminator is used to verify whether the data has realistic feeling. In order to cover a wide range of abnormal situations, the generator receives two types of input. One is user portrait mutation data, such as sudden change of risk preference, dramatic fluctuation of emotional label, and shift of channel preference distribution. The other is environmental interference data, such as sudden increase of network delay, decline of device performance, and unavailability of channel. The generator fuses these two types of data and adds random noise to form an abnormal scenario test data set.

[0061] The service policy to be executed is loaded into a test sandbox environment, which isolates real users and systems to ensure that the test process does not interfere with the actual service. The abnormal scenario test data set is input into the sandbox environment, and the policy is executed in the sandbox. The response state and failure events of the policy are recorded during execution, and each failure event needs to have attributes such as timestamp, type label, and duration for subsequent analysis. The analysis module classifies and counts the collected failure events to form a failure type distribution, such as speech recognition failure, response timeout, and decision path error. Based on the total number of events and the failure distribution, combined with the continuous failure detection method, the fault tolerance rate index of the policy is calculated. Finally, the fault tolerance rate index and the failure type distribution are sorted into a verification result report to guide subsequent policy optimization.

[0062] In different embodiments, user portrait mutation data can be generated by a perturbation model based on historical data, or by randomly replacing key labels after clustering. Environmental interference data can be injected through a high-delay network tunnel, or simulated by limiting computing resources to simulate device performance degradation, or simulated by forcibly closing some communication interfaces to simulate channel unavailability. The injection method of random noise can use Gaussian noise, Poisson noise or uniform distribution noise to increase the diversity and complexity of the generated data.

[0063] The implementation of the test sandbox environment can use virtual containers or distributed simulation platforms. Virtual containers are suitable for single-machine rapid verification, while distributed simulation platforms are suitable for large-scale parallel testing. The failure event recording mechanism can be implemented based on a log collection framework or captured in real time through an event-driven listening mechanism. The calculation method of the fault tolerance rate index can be based on the proportion formula, that is, the ratio of the number of effective failure events to the total number of test cases; or a continuous failure coefficient can be added to measure whether there is a chain reaction in the system. The output format of the verification result report can use structured JSON format for easy automatic parsing, or use a visual dashboard to support manual analysis.

[0064] Example: In the medical health business field, a remote health management system may encounter high concurrency requests when patients have sudden health events. By generating adversarial networks to simulate a large number of patients simultaneously uploading heart rate data, emotional label fluctuations, and device delay problems, the service strategy to be executed can be executed in a sandbox to test its stability under high concurrency and emotional mutations. The verification result report can reveal that the response delay of the strategy is high in the heart rate surge scenario, thereby providing data support for optimizing the service process.

[0065] In the financial technology business field, taking insurance as a scenario, the claim service system may encounter abnormal situations such as users submitting claim applications, network fluctuations, and identity verification failures. By generating adversarial networks to build a dataset containing sudden claim peaks, delay fluctuations, and channel unavailability, the service strategy to be executed can be executed in a sandbox to find that the strategy fails to switch to a backup channel in time when the channel is unavailable. The verification result report reveals the risk points through the failure type distribution and fault tolerance rate indicators, helping to optimize the strategy and improve the stability and customer satisfaction of the claim process.

[0066] This embodiment can simulate abnormal scenarios through adversarial networks and execute the service strategy to be executed in a sandbox environment, which can comprehensively test the robustness of the strategy under extreme conditions without affecting the actual user experience. This process not only discovers the weaknesses of the strategy under different types of abnormalities, but also quantitatively evaluates the robustness of the strategy through fault tolerance rate indicators. Compared with relying on historical event playback or single stress testing, this mechanism can dynamically generate test data that covers more boundary conditions, thereby more effectively guiding strategy optimization and iteration.

[0067] S60, adjusting the service strategy to be executed based on the verification result to obtain a final service strategy;

[0068] In this embodiment, the service strategy to be executed is adjusted based on the verification result to obtain a final service strategy, mainly relying on the report generated by the verification link as input, which contains fault tolerance rate indicator values and failure type distribution. First, the verification result report needs to be parsed to extract the fault tolerance rate indicator values, which are used to judge whether the current strategy meets the expected stability requirements. At the same time, the main error categories in the report need to be identified, which usually include system resource errors, strategy logic errors, and environmental interference errors.

[0069] When the fault tolerance rate index value is lower than the threshold value, targeted adjustment operations need to be performed. The adjustment mode corresponding to system resource errors is to embed a resource fuse mechanism, for example, in the case of high concurrency, some requests are quickly rejected or transferred when the resource threshold is exceeded to avoid the overall system collapse. The adjustment mode corresponding to policy logic errors is to insert a decision verification node, for example, a logical verification point is added in the decision tree path to prevent the wrong path from being selected. The adjustment mode corresponding to environmental interference errors is to increase the standby communication channel, for example, switch to the standby channel when the main communication channel is unstable to ensure service continuity.

[0070] These defense strategies are added to the to-be-executed service strategy to form a defense-enhanced service strategy. The enhanced strategy needs to be subjected to local stress testing again to judge the optimization effect by monitoring the fault tolerance rate change value. If the fault tolerance rate change value reaches the set target, the defense-enhanced service strategy is taken as the final service strategy. If the target is not reached, the operation of matching and adding defense strategies needs to be performed again until the fault tolerance rate meets the requirements. Finally, the verified and adjusted strategy is output as the final service strategy and enters the subsequent execution and feedback link.

[0071] In different embodiments, the matching of defense strategies can be realized by a static mapping table or by an intelligent selection mechanism based on clustering. Static mapping is suitable for environments with clear rules, for example, fixed mapping of system errors to resource fuse mechanisms; while the clustering method dynamically selects appropriate defense strategies based on error log distribution.

[0072] The implementation of the resource fuse mechanism can use threshold monitoring or adaptive algorithms to dynamically adjust the fuse threshold according to the real-time load of the system. The implementation of the decision verification node can be realized by a logical expression judgment or a secondary prediction through an independent verification model to avoid the execution of the wrong path. The implementation of the standby communication channel can be multi-link redundancy or cross-protocol switching, for example, switching from a network channel to a short message channel.

[0073] Local stress testing can be based on random abnormal data or by constructing a special test set for the newly added defense module. In some embodiments, an incremental testing mechanism can also be introduced to verify only the adjusted module to reduce the testing time cost.

[0074] This embodiment can continuously iterate the to-be-executed service strategy before real execution through adjustment based on verification results, gradually repair the problems exposed by error categories, and form a final service strategy. This mechanism ensures that the strategy is not only effective in normal environments but also robust in complex and extreme environments. Compared with traditional methods that rely on fixed rules or single testing, this mechanism can realize dynamic closed-loop optimization, significantly improving service stability and fault tolerance capability.

[0075] S70, execute the final service strategy and record portrait change data and service indicators, and feed back the portrait change data and service indicators to the value network and portrait construction model.

[0076] In this embodiment, executing the final service strategy means that the verified and adjusted strategy is deployed to the actual interaction scenario. The execution process not only includes calling the service logic, but also covers operations such as channel switching, information pushing, and resource allocation. During the execution process, the system continuously collects the behavior feedback of the user in the interaction, such as interface operation track, interaction time length, and reaction speed, and combines the service call log to form portrait change data. The portrait change data reflects the differences in risk preference, channel preference, and emotional state of the user before and after the execution of the strategy.

[0077] The recording of the service indicators is performed in parallel with the portrait change data, and mainly includes response delay, processing time length, and customer satisfaction indicators. These service indicators are derived from the performance log of the service execution system and the user feedback collection module. By synchronously collecting and recording the portrait change data and the service indicators, it is ensured that the input data for subsequent feedback is complete and traceable.

[0078] The feedback link needs to convert the collected portrait change data into an input format suitable for the portrait construction model, and standardize the service indicators into data that can be used for value network training. The formatted and standardized data is input into the portrait construction model and the value network, respectively, the former is used to correct the multi-dimensional feature distribution of the user portrait, and the latter is used to update the value evaluation function of the strategy. Finally, this feedback mechanism ensures the continuous evolution ability of the strategy after execution, so that the system can iteratively optimize round by round.

[0079] In different embodiments, the monitoring granularity when executing the final service strategy can be different. In one way, the collection granularity is global, covering all fields of the user interaction log; in another way, only data directly related to risk preference and emotional state can be collected to reduce system overhead.

[0080] The generation method of the portrait change data can use an incremental update model, that is, only the change amount compared with the previous version of the portrait is calculated, or a full amount reconstruction method, that is, a complete portrait data is generated every time, and then compared with the historical data.

[0081] The recording method of the service indicators can rely on an independent monitoring agent to collect, such as a delay monitoring module deployed in the load balancing layer; or rely on internal system logs, and the service processing module directly outputs the processing time length and error rate.

[0082] The manner of feeding back to the portrait construction model can adopt batch updating, that is, inputting accumulated data every fixed time period; or real-time updating, that is, triggering parameter modification of the model immediately after each portrait change data is generated. The manner of feeding back to the value network can also be divided into two paths of full amount training and online reinforcement learning, the former is suitable for a scenario with large amount of accumulated data, and the latter is suitable for a scenario with high real-time requirement.

[0083] The embodiment can realize a closed loop from policy execution to policy optimization by recording the portrait change data and the service indicators after executing the final service policy and feeding back the data to the portrait construction model and the value network. This mechanism ensures that the policy is not fixed once, but can be continuously adapted to user behavior and environmental fluctuations in operation. The final effect is that the stability and precision of the policy are improved, while the need for manual intervention is reduced, and the overall service efficiency and customer satisfaction are improved.

[0084] The application relates to the technical field of artificial intelligence, can be applied to business scenarios such as financial technology and medical health, and discloses a service policy generation and optimization method, device, equipment and medium, which comprises the following steps: collecting original interaction data from multiple interaction channels, generating a multi-dimensional dynamic portrait based on the original interaction data through a portrait construction model; loading a preset policy set and generating a pre-grouping policy set through a clustering module, matching the multi-dimensional dynamic portrait with the pre-grouping policy set to obtain an initial service policy; combining the multi-dimensional dynamic portrait with service stage identification information to generate a composite state vector, inputting the composite state vector into a pre-trained value network to obtain an action combination; optimizing the initial service policy through the action combination to generate a to-be-executed service policy; constructing abnormal scene test data through a generative adversarial network, processing the abnormal scene test data through the to-be-executed service policy to generate a verification result; adjusting the to-be-executed service policy based on the verification result to obtain a final service policy; executing the final service policy and recording portrait change data and service indicators, and feeding back the portrait change data and the service indicators to the value network and the portrait construction model. The application generates a multi-dimensional dynamic portrait by fusing multi-source interaction data, dynamically optimizes an initial service policy in combination with an action combination output by a value network, constructs test data through a generative adversarial network in an abnormal scene to complete policy verification and adjustment, and finally realizes closed loop evolution of a service policy. The method can overcome the defects of the prior art that a service policy depends on static rules and lacks cross-channel adaptation and dynamic updating, thereby improving real-time adaptability and stability of customer service, significantly improving customer experience and improving service resource utilization efficiency.

[0085] In one embodiment, the step S10 comprises:

[0086] S101, application program burying point logs are collected, user interface operation events in the application program burying point logs are parsed, and an interface interaction sequence is generated;

[0087] S102, extract device type features including operating system type, screen size coefficient and device stability from the application's data logging;

[0088] S103, Collect voice call recordings, extract the voiceprint spectrum features of the voice call recordings, and transcribe the voice call recordings into text dialogue data through an automatic speech recognition engine;

[0089] S104, acquire an image and upload a document, and extract structured field information from the image-uploaded document through an optical character recognition interface;

[0090] S105, Collect historical service records, and format the historical service records to generate time-series behavioral data with timestamps;

[0091] S106, input the interface interaction sequence and behavior time series data into the long short-term memory network of the profile construction model, perform abnormal keyword time series analysis, and obtain abnormal preference scores;

[0092] S107, Input the structured field information and device type features into the gradient boosting decision tree of the profile construction model, determine the channel selection probability distribution, and generate the channel preference probability distribution;

[0093] S108, Input the text dialogue data and voiceprint spectrum features into the bidirectional encoder representation module of the portrait construction model, perform emotional semantic analysis, and obtain emotional state labels;

[0094] S109, Based on the abnormal preference score, channel preference probability distribution and emotional state label, construct a multi-dimensional dynamic profile.

[0095] In this embodiment, the collection of raw interactive data for multi-channel interaction scenarios starts with bidirectional access from both the client and server sides. Mobile applications and web pages output event streams such as clicks, exposures, dwell time, swipe depth, and control focus changes through embedded tracking components. The voice system outputs voice call recordings and call metadata, the image entry outputs image upload documents and file metadata, and the customer service and business systems output historical service records and processing results. All data is aligned with a unified time source and appended with user, device, and session identifiers. Before entering subsequent processing stages, data is de-identified, deduplicated, and placed for missing values. Time alignment uses a unified timestamp down to the millisecond level and establishes a cross-channel event sequence index to ensure consistency and replayability of data from different sources on the timeline.

[0096] The application program log point analysis is organized in the event level minimum unit, and the event includes page identification, component identification, interaction type, occurrence time, residence time, coordinate or path parameter and environment information, and an interface interaction sequence is generated through session boundary detection and time sorting. The session boundary can be determined by a long interval threshold or an explicit opening and closing signal, and after serialization, the events are discretized and vectorized, the event embedding is constructed to carry the page semantics and operation intention, and then a fixed or variable length sequence is formed according to the time sequence to provide a structured representation for the time sequence model input. Abnormal noise events are removed by threshold and density rules, and the main sequence reflecting the real behavior mode is reserved.

[0097] The device type feature comes from the request header and device information segment in the log point, the operating system type is encoded by discrete enumeration, the screen size coefficient is obtained by normalizing the resolution and pixel density, and the device stability is obtained by recording and counting the history of the device used by the customer. Stability calculation can be based on device replacement frequency, longest use time of a single device and interval between last replacement to generate a numerical score, and a dynamic sliding window smooth value is stored on the customer side to offset short-term fluctuations. The device type feature is used as an a priori environmental variable and a controllable condition variable in subsequent channel preference modeling.

[0098] The voice call recording processing link first performs endpoint detection, noise reduction and gain standardization, extracts the voiceprint spectrum feature in the stable segment, and recommends using the log-mel spectrum or mel frequency cepstrum coefficient and performing consistent configuration on frame length, frame shift and filter bank number to ensure the uniformity of feature size under different call lengths and sound field conditions. The parallel path calls the automatic speech recognition engine to generate text dialogue data, outputs structured text containing sentences, speaker separation, word-level timestamps and confidence, and if necessary, further improves readability and downstream matching accuracy through error correction and punctuation restoration. The acoustic and text two-way alignment anchor points are reserved for subsequent multi-modal fusion.

[0099] The structured field information of the image uploaded document is completed through image preprocessing and text detection and recognition two stages. The preprocessing stage includes tilt correction, perspective correction, denoising and contrast enhancement, the text detection uses layout analysis and region segmentation technology to output text blocks and key domain positions, the recognition stage sends the text blocks into the recognition engine for word block splicing and key-value pair extraction, and finally generates a structured representation including field name, field value and confidence, and improves the extraction reliability with double verification at the document level and field level. The references and consistency between fields are checked by a rule base or a lightweight learner, and contradictory items are put into the artificial review queue to ensure the quality of training data.

[0100] The historical service records are mapped to unified fields and timestamped to generate timestamped behavior time series data, which records request types, processing nodes, processing time, result status, exception labels, and channel records, and is organized as a time-ordered event chain by user and session dimensions. To adapt to time series modeling, long sequences are windowed and state snapshots are inserted at window boundaries to preserve context, missing fields are repaired by forward filling with the latest observation or model filling strategies, ensuring time series density and context integrity.

[0101] The interface interaction sequence and behavior time series data are input into a long short-term memory network to perform abnormal keyword time series analysis. The input includes event embedding and time interval encoding. Keyword trigger signals are obtained by searching for an abnormal dictionary or weakly supervised trigger in text and structured events. The time series model learns the dynamic relationship between keyword triggers and behavior chains at the sequence level and outputs an abnormal preference score between zero and one. In the training stage, historical cancellations, complaints, and timeout operations can be used as target events for weak labeling. Sequence-level or segment-level contrast loss is used to improve sensitivity to sudden behavior changes. In the inference stage, the score is smoothed and thresholded to reduce the impact of incidental noise.

[0102] Structured field information and device type features are input into a gradient boosting decision tree to determine the channel selection probability distribution. The input includes customer attributes, invoice field-derived quantities, device environment quantities, and context statistics. The output is a selection probability vector for each service channel. Training labels come from historical services where customers choose channels in different situations. Stratified sampling and class reweighting are used to alleviate the problem of insufficient samples in long-tail channels. The prediction output is temperature calibrated and probability normalized to form a channel preference probability distribution with confidence limits, which is used by downstream strategies to take a fallback path when uncertainty is high.

[0103] Text dialogue data and voiceprint spectral features are input into a bidirectional encoder representation module to perform sentiment semantic analysis. The text branch uses word or character granularity encoding and introduces speaker labels. The acoustic branch takes spectral frame sequences as input and adds speaker intensity and speech rate derived quantities. The two branches perform cross-modal attention fusion at aligned anchor points. The output is an emotional state label and corresponding confidence. The label set can include calm, positive, anxious, angry, and other multi-level states. Time series smoothing and cooling mechanisms are used to avoid frequent fluctuations in a short period of time. The fusion strategy supports early alignment and late feature concatenation to adapt to different dialogue lengths and noise levels.

[0104] The abnormal preference score, the channel preference probability distribution, and the emotion state label are combined in the same image space to form a multi-dimensional dynamic image. The combination process first completes the scale unification and time alignment, and then records the source, time window, and confidence boundary in the form of key-value pairs. After vectorization, it is written into the image storage. The image not only saves the instantaneous aspect but also maintains the trajectory over time. Through sliding window aggregation, exponential decay, and anomaly detection, stable preferences and short-term deviations are marked, providing searchable and interpretable input for subsequent strategy matching and value assessment.

[0105] The embodiment constructs a continuous pipeline from multi-channel collection to time series analysis, feature extraction, and multi-modal fusion. The abnormal preference score is obtained by long short-term memory network from interface interaction sequence and behavior time series data. The channel preference probability distribution is obtained by gradient boosting decision tree from structured field information and device type features. The emotion state label is obtained by bidirectional encoder representation module from text dialogue data and voiceprint spectrum features. The three are converged into a multi-dimensional dynamic image in a unified space-time coordinate. Since time alignment, scale unification, and uncertainty calibration are completed in the pre-processing stage, the image can reflect both long-term stable features and short-term fluctuation features, improving the sensitivity to risks and abnormalities, and the accuracy of channel allocation and emotion understanding. This provides a high signal-to-noise ratio input for subsequent strategy selection and strategy learning, reducing resource mismatch and experience fluctuations caused by misjudgment and missed judgment.

[0106] In one embodiment, the above step S20 includes:

[0107] S201, loading a preset strategy set containing a plurality of service strategy templates, extracting abnormal sensitivity parameters, channel adaptation types, and emotion response modes of the service strategy templates;

[0108] S202, performing feature vectorization processing on the abnormal sensitivity parameters, the channel adaptation types, and the emotion response modes by a clustering module to generate a strategy feature vector;

[0109] S203, performing grouping operation to group the service strategy templates into a plurality of strategy base groups based on the strategy feature vector, and determining the centroid coordinates of each strategy base group;

[0110] S204, establishing an association relationship between the strategy base groups and the service strategy templates, generating a strategy base group association mapping table, and storing the strategy base groups, corresponding centroid coordinates, and strategy base group association mapping table into a strategy grouping database;

[0111] S205, quantifying the abnormal preference score in the multi-dimensional dynamic image into an abnormal feature value, mapping the channel preference probability distribution in the multi-dimensional dynamic image into a channel feature vector, and converting the emotion state label in the multi-dimensional dynamic image into an emotion feature code;

[0112] S206, combine the abnormal feature value, the channel feature vector and the emotion feature code to form a multi-dimensional dynamic image feature vector, determine the cosine similarity between the multi-dimensional dynamic image feature vector and the centroid coordinates of each strategy base group, and select the strategy base group with the highest cosine similarity as the target strategy base group;

[0113] S207, select a strategy template as an initial service strategy according to the service scene complexity and the adaptive score of the service strategy template.

[0114] In the embodiment, the preset strategy set is composed of a group of executable service strategy templates, each template contains structured fields such as service channel, service content sequence, resource proportion and trigger condition, and is attached with three types of learnable attributes for grouping and matching: abnormal sensitivity parameter, channel adaptation type and emotion response mode. The abnormal sensitivity parameter is derived from the performance evaluation of the template in the abnormal situation under the historical scene, and the tolerance or response intensity of the abnormal behavior is recorded using a real number from zero to one; the channel adaptation type expresses the adaptation tendency of the template to the channels such as telephone, APP, webpage and offline in a multi-dimensional probability vector; the emotion response mode expresses the response strategy of the template to the states such as calm, positive, anxious and angry in a discrete code or vector embedding. The template is subjected to consistency check and field completion before being stored, so as to ensure the uniformity of field domain, numerical range and measurement unit, and facilitate subsequent vectorization and clustering.

[0115] The clustering module performs feature vectorization processing to map the abnormal sensitivity parameter, the channel adaptation type and the emotion response mode to the same measurement space. The mapping process includes scale unification, distribution alignment and weight configuration: scale unification eliminates dimensional differences through minimum maximum scaling or quantile scaling; distribution alignment relieves the long tail by using probability calibration or logarithmic transformation; weight configuration allows adjusting the contribution of the three types of attributes to the overall similarity in modeling configuration. After mapping, the strategy feature vector is obtained, the vector dimension is fixed and the field order is stable, supporting batch calculation of distance and similarity.

[0116] The grouping operation takes the strategy feature vector as input, divides the service strategy templates into multiple strategy base groups according to the distance or similarity between vectors, and calculates the centroid coordinates for each strategy base group. The distance metric can use Euclidean distance, Mahalanobis distance or cosine-based angle distance; the aggregation criterion can use a partition or density strategy to ensure that the same group is close and the groups are separated. The centroid coordinates are represented by the weighted average or geometric median of all vectors in the group, and the weight is derived from the historical effect score or the size of the covered population, which is used to represent the "typical strategy image" of the group. After grouping, a stable group identifier and version number are assigned to each strategy base group to facilitate online selection and offline evaluation traceability.

[0117] The association relationship between the strategy base group and the service strategy template is recorded in a mapping table, which includes group identification, template identification set, centroid coordinates, feature statistical summary (such as covariance, vector range) and validity flag. The mapping table is stored together with the strategy base group entity and the centroid coordinates in the strategy grouping database, and atomic writing is used to avoid cross-table inconsistency; the database layer realizes versioned storage and expired cleaning, supporting hot updating of group structure and centroid vector without affecting online retrieval. To improve retrieval efficiency, an additional inverted structure or a neighbor index based on vector index is established to realize fast positioning of target groups and candidate templates.

[0118] The multi-dimensional dynamic portrait comes from the upstream portrait construction process, including abnormal preference scores, channel preference probability distribution and emotion state labels. In order to align with the clustering space, the abnormal preference scores are first mapped to abnormal feature values, and the dominant effect of extreme scores on distance calculation is eliminated through monotonic transformation and temperature calibration; the channel preference probability distribution is mapped to the channel feature vector, and temperature smoothing and threshold truncation are performed as necessary to reduce weak signal noise; the emotion state label is converted into an emotion feature code, the text label uses one-hot or embedding representation, and the continuous emotion intensity uses normalized real value. The three types of vectors perform splicing and scale unification to form a multi-dimensional dynamic portrait feature vector, and are accompanied by a timestamp and a confidence boundary to ensure consistency with the strategy feature space.

[0119] The multi-dimensional dynamic portrait feature vector and the centroid coordinates of each strategy base group are evaluated for similarity, the cosine similarity is calculated, and the group with the highest score is selected as the target strategy base group. During the calculation process, the vector is unitized and the dot product is used to realize it, avoiding norm drift in high-dimensional space; when the difference between the highest score and the second highest score is below a threshold, an uncertainty branch is enabled to allow a small number of candidate groups to be retained for subsequent multi-arm selection or A / B evaluation. Similarity calculation retains intermediate decomposition results, which facilitates subsequent explanation of which dimensions contribute the most, thereby supporting interpretable selection.

[0120] The service scenario complexity comprehensively considers the service request type, system load index and customer value label and outputs a complexity score of zero to one through a rule or a learning model; the adaptation score measures the matching degree between a single service strategy template and the current multi-dimensional dynamic portrait and scenario complexity, and the calculation can adopt weighted scoring: the channel adaptation type of the template and the channel feature vector of the portrait are used for vector similarity, the emotional response mode of the template and the emotional feature code of the portrait are used for matching score, and the weights of each item are dynamically adjusted according to the complexity. The selection of the initial service strategy is performed in the target strategy base group, and the template with the highest adaptation score and meeting the operation constraints (such as resource proportion, time delay budget, compliance limit) is preferentially selected; when the scores of multiple candidates are similar, a lightweight exploration strategy can be introduced to improve the long-term benefit, and the selection record is written back to support online learning. The selected template does not change the original group structure, and is only instantiated on the execution side, ensuring the decoupling of stable grouping and online selection.

[0121] In the embodiment, a unified feature space is constructed by using the abnormal sensitivity parameter, the channel adaptation type and the emotional response mode, and clustering grouping is implemented, so that the service strategy template obtains a structured group organization and a representative centroid in the warehousing stage; by mapping the multi-dimensional dynamic portrait to the same vector space and using cosine similarity to locate the target strategy base group, the strategy retrieval converges from linear search of the whole library to neighbor search of the clustering center, so that the matching cost is reduced and the stability is improved; by combining the service scenario complexity and the adaptation score to select in the target group in a fine-grained manner, the initial service strategy can meet the operation constraints while fitting the risk, channel and emotion dimensions.

[0122] In one embodiment, the above step S30 comprises:

[0123] S301, obtaining service stage identification information, parsing the flow node state in the service stage identification information, and generating a service stage code;

[0124] S302, extracting an abnormal preference score in the multi-dimensional dynamic portrait as an abnormal feature value;

[0125] S303, extracting a channel preference probability distribution in the multi-dimensional dynamic portrait as a channel dimension vector;

[0126] S304, extracting an emotional state label in the multi-dimensional dynamic portrait as an emotional dimension code;

[0127] S305, splicing the abnormal feature value, the channel dimension vector, the emotional dimension code and the service stage code to generate an original composite vector;

[0128] S306, performing standardization processing on the original composite vector to generate a composite state vector;

[0129] S307, input the composite state vector into the pre-trained value network, perform forward propagation processing through the neural network layer of the value network, and obtain an action value distribution;

[0130] S308, select an action combination with the maximum value expectation from the action value distribution.

[0131] In the embodiment, the fusion of the multi-dimensional dynamic portrait and the service stage identification information is completed in units of single decision moments. First, service stage identification information is obtained, which is derived from a node state stream of an online service arrangement engine or a business process engine, including fields such as node name, stage sequence number, whether it is a branch point, whether it is a convergence point, and stage remaining time limit. The state stream is mapped to a service stage code by a state parser. The code contains both static position information (discrete embedding of the stage sequence number) and dynamic context information (normalized real value of the remaining time limit and binary bit of the branch flag). The coding dimension is fixed, the field order is fixed, a version number is used to mark the field set, and the alignment with the subsequent model input is ensured.

[0132] The abnormal preference score is derived from the portrait calculation module and is a real-valued quantification of the tendency of abnormal behavior. The score is calibrated by a confidence interval and scaled by a temperature to obtain an abnormal feature value. The calibration step introduces variance estimation within a historical window to suppress short-term fluctuations. Temperature scaling is used to flatten the dominant effect of extreme values on subsequent scoring. The abnormal feature value remains a one-dimensional real number, with an optional uncertainty estimate (such as standard error) for conservative decision-making in the value evaluation stage.

[0133] The channel preference probability distribution is a probability vector for channels such as phone, APP, web page, and offline. The vector is smoothed by a probability temperature and cropped by a threshold to obtain a channel dimension vector. Temperature smoothing reduces tail probability noise, and threshold cropping ensures that there is no numerical underflow in the total dimension. Subsequently, log-ratio transformation or centering processing is performed to make the distributions of different users comparable on the same scale. The output is still a fixed-length vector, with each dimension corresponding to a channel slot, without changing the channel enumeration set.

[0134] The emotion state label is a discrete or weakly continuous emotion label derived from the fusion of text dialogue data and voiceprint spectrum features. The label is mapped to an emotion dimension code, supporting two types of implementation: one-hot encoding for discrete labels and embedding encoding for introducing emotion similarity (e.g., the proximity relationship between anxiety and anger). If the upstream provides emotion intensity (e.g., in the 0-1 interval), the intensity coefficient is appended after embedding to form a weighted emotion vector to preserve the intensity information. The encoder maintains a stable mapping table from emotion labels to vectors, which is managed together with the model version.

[0135] The original composite vector is generated by splicing the abnormal feature value, channel dimension vector, emotion dimension code, and service stage code in a predetermined order. Before splicing, shape and missing checks are performed on each component. If the requirements are not met, a bottom-up filling mechanism is triggered: missing values are filled with a configured neutral value, and out-of-bound values are replaced with boundary values, and a quality flag bit is recorded for diagnosis. After the original composite vector is generated, it enters the standardization process, which includes two layers of processing: the first layer is a feature-level transformation that performs Z-score standardization or robust scaling on continuous components and keeps the original value of discrete embeddings; the second layer is batch-level alignment, which performs centering on each batch of vectors through sliding statistics or offline frozen statistics in online inference to ensure that the numerical distribution is consistent with the statistics during value network training. After completion, the composite state vector is output, with fixed dimensions, controlled range, and no missing points.

[0136] The composite state vector is input into the pre-trained value network, which consists of an input layer, several hidden layers, and an output layer. The input layer receives the complete vector, and the hidden layer uses a combination of full connection and gate units to process heterogeneous features; to avoid gradient deviation caused by scale difference, normalization and activation functions are enabled in the hidden layer. The network internally sets a context channel to emphasize the service stage code, and the channel weight is multiplied by the attention or gate coefficient and other components to achieve stage-sensitive value evaluation. Forward propagation produces an action value distribution, where each dimension of the vector corresponds to an executable atomic action or a compliant composite action label, and the numerical value is the expected value of the action in the current state. If the output layer contains illegal action bits (restricted by the scene or regulation), the value is reduced to the minimum value after inference through the mask mechanism to ensure that subsequent selection does not include non-compliant items.

[0137] The selection of action combination adopts the principle of value expectation maximization. First, filter out the constraint conflict items according to the masked action value distribution, and then calculate the weighted value in the remaining set. The weight can be determined by the stage coefficient derived from the service stage code, which is used to emphasize different action types in different stages (such as emphasizing verification-related actions in the identity verification stage). If there are multi-action coordination constraints, it is necessary to map from the value distribution to the combination space. The strategy executor searches for the set with the maximum weighted value and satisfying the mutual exclusion constraint in the combination space through heuristic search or beam search. After selection, the intermediate product of the current state-action pair (composite state vector, action value distribution, selected action set, and mask information) is retained to provide traceable evidence for subsequent offline review and online learning.

[0138] The embodiment unifies multi-dimensional dynamic images and service stage identification information, encodes and standardizes them into a composite state vector, inputs them into a value network training distribution alignment, reduces evaluation instability caused by numerical deviation, and cooperates abnormal characteristic values, channel dimension vectors and emotion dimension encoding in the same representation space. The value network can simultaneously perceive risk tendency, channel accessibility and emotion sensitivity in a single decision, thereby differentiating the action value. The selection is implemented based on the action value distribution combined with stage constraints and compliance masks to avoid inappropriate actions entering the execution link. Therefore, the action selection is closer to the current state and stage semantics, reducing invalid interactions and repeated communication. The output is resilient to abnormalities and emotional fluctuations, and the action combination still maintains predictable service effects in a high-noise environment. Subsequent strategy optimization and closed-loop training obtain more stable supervision signals.

[0139] In one embodiment, the step S40 described above comprises:

[0140] S401, parsing channel switching instructions, information pushing instructions and resource allocation instructions in the action combination;

[0141] S402, parsing the initial service strategy to separate service channel configuration, service content sequence and service resource ratio;

[0142] S403, updating the service channel configuration in response to the channel switching instructions;

[0143] S404, inserting information pushing content into the service content sequence in response to the information pushing instructions;

[0144] S405, adjusting the service resource ratio in response to the resource allocation instructions;

[0145] S406, integrating the updated service channel configuration, the service content sequence with inserted information pushing content and the adjusted service resource ratio to obtain an integrated strategy;

[0146] S407, adding an execution time constraint condition to the integrated strategy to generate a to-be-executed service strategy and storing the to-be-executed service strategy into a strategy execution queue.

[0147] In this embodiment, the action combination analysis takes a structured protocol as input, first extracts the instruction list, and then performs consistency verification and conflict resolution. The instruction list only contains three types of elements, namely channel switching instructions, information pushing instructions, and resource allocation instructions. The parsing link performs type identification, parameter unpacking, and legality check on each element. For example, the channel switching instruction needs to carry the target channel identifier and the effective range, the information pushing instruction needs to carry the content identifier and the insertion position, and the resource allocation instruction needs to carry the resource weight vector and the effective window. To avoid repeated execution, the parsing link calculates the idempotent key of the instruction and generates a deduplicated result set, and sorts based on the time tag to ensure a unique downstream processing order.

[0148] The initial service policy is decomposed into three independently maintained configuration bodies, namely service channel configuration, service content sequence, and service resource allocation. The service channel configuration records the current primary channel, backup channel, and cross-channel synchronization strategy, including session context synchronization switch and credential transfer method. The service content sequence is a finite sequence structure, with elements being operation units or information items, with sequence number, condition tag, and visibility flag. The service resource allocation defines the allocation proportion and upper threshold of each resource pool, such as the percentage allocation of intelligent agents and human agents and the concurrent upper limit. The decomposition link outputs the memory view of the three configuration bodies through a structured extractor, with version number and checksum, ensuring traceability of subsequent changes.

[0149] For channel switching instructions, the update logic follows a state machine-driven change process. First, evaluate the availability of the current channel and the reachability of the target channel, and determine the feasibility of switching through health probes and permission checks. Then set the new primary channel in the service channel configuration and generate a cross-channel synchronization mapping, including session identifier mapping and event mirroring rules, to ensure consistency of context between different channels. If the target channel does not meet the reachability condition, fall back to the primary channel unchanged and add a fallback strategy, mark the backup channel as candidate and write it to the evaluation list. All changes are completed within the transaction boundary to avoid configuration tearing caused by updating only part of the fields.

[0150] For information pushing instructions, the insertion behavior focuses on position decision, deduplication, and throttling. Position decision generates a candidate position set based on the insertion position, dependency tag, and the ordered relationship of the existing service content sequence. If the target position already exists, perform in-place replacement and merge the display constraints; if not, insert at the highest priority position in the candidate set. To avoid congestion caused by high-frequency pushing, query the number of times the same information has been sent within the time window before insertion, and if it exceeds the threshold, delay to the next window and add a deduplication mark. After insertion, rebuild the sequence number of the service content sequence to ensure that the subsequent executor processes in linear order.

[0151] For resource allocation instructions, the process is adjusted to the principle of safe convergence. First, verify that the resource weight vector is within the allowable interval and the sum is one. If it is out of bounds or the sum is not one, perform vector projection and normalization to restore the valid value range. Then, calculate the target allocation based on real-time capacity and queue length. Use linear interpolation to gradually approach the target value over several cycles to avoid resource shock caused by instantaneous jitter. If the target resource pool capacity is insufficient, automatically reduce its weight and correspondingly increase the weight of other pools, while marking the resource pool that needs to be expanded for the external orchestration system to trigger capacity adjustment.

[0152] After the completion of the three types of updates, the integration phase is entered to generate the integrated strategy. The integration phase includes consistency checking and dependency analysis. Consistency checking checks whether there are semantic conflicts in the service channel configuration and service content sequence, such as the main channel being voice and the first item in the sequence being an information item only visible on the interface. Such conflicts are resolved by automatically replacing them with equivalent content that can be presented by voice or enabling linkage reminders in the channel. Dependency analysis is used to handle conditional tags in the service content sequence, so that content related to the channel or resources is delayed or hidden when the conditions are not met. The integration result is written back to the memory view in a single transaction and generates a version increment, ensuring atomicity of changes.

[0153] The addition of time constraints is mainly in response to two types of constraints: response time and completion time. Response time is calculated based on historical service indicators and the urgency of the action combination. The urgency can be mapped from the priority label or service stage information in the action combination. Completion time is determined based on the estimated execution time of the service content sequence and resource availability. Both types of constraints are attached to the corresponding nodes or global scope of the integrated strategy in the form of numerical intervals, with violation handling rules such as automatically upgrading resources or switching channels. Time constraints and resource allocation maintain a linkage relationship. If the time limit is too tight and the resource weight is insufficient, temporary promotion of resource weight is triggered and then falls back after it is achieved.

[0154] The generation of the service strategy to be executed is completed after the integration and addition of time constraints described above. The generation process creates a unique identifier, records the source action combination summary, change summary, and version information, and solidifies it into an executable execution manifest. The manifest is submitted to the policy execution queue, which records priority, expiration time, retry policy, and idempotent key. Before entering the queue, perform duplicate detection. If there is an entry with the same idempotent key in the queue that is not expired, merge it into one execution and combine the time constraints to avoid repeated delivery. The entry uses a persistent message system or database table to implement, supporting sequential guarantees and failure retries to ensure that the strategy can be reliably landed in a distributed execution environment.

[0155] The embodiment combines the action group into channel switching instructions, information pushing instructions and resource allocation instructions, and respectively performs structured updates, so that the service channel configuration, the service content sequence and the service resource ratio configuration are changed consistently in a unified transaction, and the policy tearing caused by partial updates is avoided. The integrated policy forms an executable list after conflict checking and dependency analysis, and superimposes the time constraint conditions, so that the downstream execution has a clear time boundary and a default remedy path. Finally, the to-be-executed service policy is queued, and the idempotent and priority control are used to reduce repeated issuance and resource jitter. Thus, stable mapping from abstract actions to executable policies is realized, the policy implementation path is shortened, the cross-channel consistency and resource utilization are improved, and the time constraint drives the service response to be more consistent with the business urgency, thereby reducing the waiting time and invalid interaction probability.

[0156] In one embodiment, the step S50 comprises:

[0157] S501, creating user portrait mutation data by a user behavior generator of a generative adversarial network;

[0158] S502, injecting environment interference data by an environment generator of the generative adversarial network;

[0159] S503, fusing the user portrait mutation data and the environment interference data, and adding random noise to generate an abnormal scene test data set;

[0160] S504, loading the to-be-executed service policy into a test sandbox environment;

[0161] S505, inputting the abnormal scene test data set into the test sandbox environment;

[0162] S506, processing the abnormal scene test data set by executing the to-be-executed service policy in the test sandbox environment;

[0163] S507, recording failure response events in the policy execution process, and analyzing failure type distribution of the failure response events;

[0164] S508, determining a policy fault tolerance rate indicator, and generating a verification result report containing the fault tolerance rate indicator and the failure type distribution.

[0165] In this embodiment, the user behavior generator of the generative adversarial network takes historical interaction samples, portrait feature distribution, and stage context as input, and outputs user portrait mutation data. The generation process maps key dimensions such as risk preference, channel preference, and emotion encoding of the portrait vector to the latent space, samples the disturbance factor, and then restores it to a time-ordered interaction trajectory and portrait snapshot sequence through the decoder. The trajectory contains indicators such as click sequence, voice interaction proportion, and channel switching frequency. The snapshot sequence is arranged by timestamp, and each frame reflects the discrete transition of risk preference, channel preference reversal, and emotion from neutral to anxious. To ensure that the mutation does not deviate from the business reachable domain, the behavior generator performs reachability constraint verification on the sequence after each sampling, such as limiting the number of channel switches within a single window and limiting the duration of extreme emotions, and resamples or truncates the trajectory that does not meet the constraints.

[0166] The environment generator of the generative adversarial network constructs environmental disturbance data for key interference sources in the service link. The interference sources include network latency, packet loss, bandwidth jitter, speech recognition confusion, page rendering blocking, backend queuing delay, etc. The environment generator constructs the interference curve with a parameterized process driven by time series noise, which is aligned with the interaction trajectory time axis to form a one-to-one correspondence with the behavior event. For example, superimposing the recognition confusion matrix on the voice interaction segment to increase the probability of homophone replacement, superimposing transient latency spikes near the interface submission event to simulate weak networks, and superimposing identity verification retries in the channel switching window to simulate authentication jitter. The environment generator generates replayable configurations and random seeds for each interference segment to ensure that the same configuration remains reproducible in multiple verifications.

[0167] The user portrait mutation data and environmental disturbance data are fused along the time axis to form primary abnormal samples, and then random noise is added to obtain an abnormal scene test data set. The fusion step matches the behavior event timestamp with the interference segment time window, generates composite labels at the event level according to the coverage relationship, and records the type and intensity of interference received by each event. Random noise is controlled in two dimensions of amplitude and frequency spectrum. Amplitude noise is used to introduce fine-grained disturbances to avoid overfitting, and frequency spectrum noise is used to simulate periodic jitter and non-stationary fluctuations. Both are clipped within the safety boundary to maintain data validity. The data set structure includes an event table, a portrait snapshot table, and an interference configuration table, which are associated with session identifiers and timestamps for easy sandbox replay and index collection.

[0168] After the service policy to be executed is loaded into the test sandbox environment, cold start verification and dependency injection are entered. The cold start verification checks whether components such as channel drivers, content renderers, and resource schedulers required by the policy are available, and the dependency injection provides a simulated resource pool, a fake external interface, and a controllable authentication response for the policy executor. The test sandbox environment receives an abnormal scenario test data set, drives events to arrive in chronological order through a playback engine, and ensures that the same event is affected by the same type of interference in the same time window. The policy executor performs channel switching, information pushing, and resource allocation decision-making according to the policy list, and samples the execution track, output action, and system state into the event stream.

[0169] The failure response event recording module generates an execution result label and an exception label for each event during the playback process, aligns the result with the expectation using matching rules, and determines whether it fails. The failure response event is stored in a structured format, including a timestamp, a session identifier, a failure type, a trigger location, an interference context, a retry number, and a recovery time consumption. The failure type distribution is given by a two-level classifier, one layer of which distinguishes system errors, policy errors, and environment errors by rule mapping, and the second layer distinguishes subtypes by feature vectors, such as queue overflow and fuse trigger for system errors, content insertion failure and channel incompatibility for policy errors, and network timeout and speech recognition confusion for environment errors. The classifier is trained with labeled samples in the construction stage, and outputs the distribution according to the confidence in the execution stage and triggers the manual annotation queue for low-confidence samples to continuously calibrate.

[0170] The policy fault tolerance rate indicator is determined by the failure density and the chain degree. The failure density is calculated by the ratio of the number of effective failure events to the total number of test cases, and the chain degree is calculated by the maximum failure count and the total length in the continuous failure window. The fault tolerance rate is defined as one minus the product of the two, which is clipped to the closed interval of zero to one. The index calculation also supports scene and channel slicing to locate the weak links in specific channels or specific stages. The final output verification result report contains the fault tolerance rate indicator, the failure type distribution, and the playback path of the key failure samples, with reproducible experimental configurations and version information, supporting horizontal comparison and vertical tracking between different versions of the policy.

[0171] The embodiment aligns and fuses the interaction trajectory and interference scene through the user behavior generator and the environment generator. The abnormal scene test dataset can cover high-risk mutations and link interference under reproducible conditions, avoiding evaluation distortion caused by relying only on a small number of historical fault samples. The service strategy to be executed is placed in the test sandbox environment for playback and recording of failure response events. Then, two-level classification is performed to obtain the failure type distribution. The fault tolerance rate index combines the failure density and the continuous failure coefficient to give a quantitative measurement of system robustness. Thus, the weak points of the strategy in channel switching, information insertion and resource allocation can be identified before formal online, and the specific failure subclass and trigger condition can be located, providing a direct basis for subsequent adjustment. Meanwhile, the test is repeatable and the results are comparable, reducing the online regression cost and improving the reliability and efficiency of strategy iteration.

[0172] In one embodiment, the step S60 comprises:

[0173] S601, parsing the verification result report, extracting the fault tolerance rate index value and identifying the main error category;

[0174] S602, when the fault tolerance rate index value is lower than the preset threshold, according to the main error category, matching the corresponding defense strategy from the preset defense strategy library;

[0175] S603, for system error category, adding the matched resource fuse mechanism to the service strategy to be executed;

[0176] S604, for strategy error category, inserting the matched decision tree verification node into the service strategy to be executed;

[0177] S605, for environment error category, adding the matched backup communication channel to the service strategy to be executed;

[0178] S606, after adding the matched defense strategy to the service strategy to be executed, obtaining the defense enhanced service strategy;

[0179] S607, performing local stress test on the defense enhanced service strategy, and recording the fault tolerance rate change value;

[0180] S608, when the fault tolerance rate change value reaches the optimization target, taking the defense enhanced service strategy as the final service strategy;

[0181] S609, when the fault tolerance rate change value does not reach the optimization target, re-executing the defense strategy matching and adding operation.

[0182] In this embodiment, the verification result report analysis link is designed for both structured and semi-structured carriers. The structured carrier adopts a fixed field key set and a timestamp index to directly extract fault tolerance rate index values, failure type distribution, continuous failure window, and key failure sample identification. The semi-structured carrier completes extraction through field name regular mapping and key-value pair confidence scoring, and low-confidence entries enter the review queue. After reading the fault tolerance rate index value, boundary check and source check are first completed (compare the calculated version and test data set hash), and then enter the threshold comparison unit; the identification of main error categories is calculated based on the failure type distribution, and the multiplication value of the proportion, impact weight, and chain coefficient is added, the dominant category is determined according to the highest score, and when the scores are close, parallel category lists are constructed to support subsequent multi-path enhancement.

[0183] Threshold judgment supports both static threshold and dynamic threshold. Static threshold is maintained in the configuration center and is suitable for general regression testing; dynamic threshold is generated based on the sliding quantile of historical fault tolerance rate and scenario label (claim, consultation, complaint, etc.), and the calculation formula aggregates the nearly three-period mean, variance, and scenario weight to output the scenario threshold. The comparison logic adopts a closed-loop design of "strictly less than triggering", and records the triggering reason, threshold source and report version for tracing.

[0184] The defense strategy library is organized with metadata-oriented retrieval: each defense strategy contains error category label, applicable channel, trigger condition expression, execution action, rollback condition, observation indicator, dependency list and version constraint. The matching process first filters according to the main error category, and then scores according to channel consistency, trigger condition coverage and dependency satisfaction to generate a candidate set. If the candidate contains multiple different versions of the same action, select the highest compatible version according to the version constraint and environmental capability. The matching result is accompanied by explanation information (trigger expression and score details) for quick positioning of the reason when subsequent implantation fails.

[0185] The addition of the resource fuse mechanism takes high-risk nodes as the implantation point. High-risk nodes are jointly determined by failure heat and resource occupation peak, and meet the condition of "failure proportion exceeding threshold and resource indicator out of bounds". The fuse configuration includes metric source (CPU, memory, thread, queue length or external dependency time consumption), threshold form (absolute threshold, percentile threshold or adaptive threshold), semi-open detection strategy and cooling time. The trigger path is: the metric sample enters the sliding window, the window statistical result is compared with the threshold, and when the threshold is exceeded, the degradation path is switched (such as degrading to read-only query or static guidance), and after the cooling time expires, it enters the semi-open state for small flow detection, and if the detection is successful, it is closed and recovered. The implantation action inserts a guardian node in the strategy execution graph and reconnects the in and out edges, so that subsequent actions follow the degradation path during the fuse period.

[0186] The insertion of decision tree check nodes targets the strategy error category. The nodes perform consistency and safety checks before decision branching points, inputting portrait key fields (abnormal preference score, channel preference probability distribution, emotional state label), stage context, and action parameters to be executed. The check tree is constructed with a mixture of rule leaves and statistical leaves: rule leaves are used for hard constraints (e.g., "high anxiety state prohibits aggressive marketing"), and statistical leaves are used for probabilistic constraints (e.g., "channel preference distribution deviates too much from the current channel, requiring secondary confirmation"). When a rejection leaf is hit, the original action is blocked and an alternative action or manual review is triggered. When a warning leaf is hit, a secondary confirmation or throttling parameter is added to the original action. The inserted nodes maintain execution logs and hit distribution for subsequent retraining.

[0187] The addition of backup communication channels targets the environmental error category. Channel descriptions include channel type (SMS, email, APP push, voice callout), encoding and bandwidth model, arrival rate and time delay distribution estimation. The health checker continuously performs lightweight probes on alternative channels to establish a health state machine. The router makes choices based on health status, user channel accessibility, and content sensitivity level: sensitive content prefers channels with high security level, and time-sensitive content uses low-latency backup channels when the main channel is unhealthy. To avoid repeated notifications, the deduplication unit constructs an idempotent key with session ID, content fingerprint, and time window to ensure that users only perceive valid notifications once in multi-channel sending scenarios.

[0188] After the matching defense strategy is completed and implanted, the service strategy is generated with defense enhancement. The generation process includes dependency verification (components and permissions), conflict detection (mutual exclusion between new rules and existing rules), performance budget evaluation (estimation of time delay and resource overhead introduced by new nodes), and versioned archive. When the performance budget exceeds, it automatically falls back to a lightweight enhancement configuration (such as shortening the fuse observation window, relaxing the check threshold, or selecting a lighter backup channel), and records the compromise list for repair in subsequent iterations.

[0189] Local stress testing is performed around newly added or changed nodes. The test case generator synthesizes targeted samples according to error categories to ensure that the added logic is triggered and boundary conditions are covered; the playback engine replays samples in a test sandbox, collecting enhancement-related indicators, including failed response event count, queue length peak, average response time delay, and user-perceptible behavior (such as channel switching frequency). The fault tolerance rate change value is represented as the difference between the new and old fault tolerance rates, with confidence interval and sample coverage rate statistics; when the difference is negative or the confidence interval crosses zero, it is determined to be substandard, and the next round of matching and addition is entered. Each iteration generates a change summary (change point, introduced / removed strategy, indicator change), which is used as a subsequent data asset deposit.

[0190] The decision-making process compares the change in the fault tolerance rate with the optimization target to determine convergence. When the target is achieved, the enhanced service strategy is released as the final service strategy, and the version, dependencies, and scope of effectiveness are registered in the strategy repository. If the target is not achieved, the process automatically returns to the matching and adding process. If there is no significant improvement for several consecutive rounds, protection is triggered (e.g., limiting the magnitude of changes in a single round or forcing manual review) to avoid oscillations.

[0191] Example Explanation: In an insurance business scenario, the system first collects raw customer interaction data from multiple interaction channels, including mobile application event logs, call recordings between customers and human agents or intelligent voice systems, customer-uploaded claims images or forms, and historical service records. This raw data is then processed by multiple modules of the profile building model to transform it into a multi-dimensional dynamic profile. For example, event logs are parsed into customer clicks and path behaviors on the interface, further extracting device operating system type and screen features to reflect the customer's digital capabilities; voice recordings are analyzed using voiceprint analysis and speech recognition to obtain dialogue text and acoustic features, combined with a bidirectional encoder structure to form emotional semantic tags, capturing anxiety or dissatisfaction during the claims process; uploaded documents are processed using optical character recognition to extract structured fields, reflecting the customer's policy information and claims details; historical service records are formatted to form timestamped behavioral time-series data, reflecting the customer's long-term preferences and interaction patterns. Finally, the model integrates abnormal preference scores, channel preference probability distributions, and emotional state tags to generate a dynamically evolving customer profile.

[0192] After obtaining a customer profile, the system loads a pre-defined set of strategies. This set contains various insurance service strategy templates, each recording sensitivity parameters, channel adaptation methods, and emotional response patterns for different scenarios. These strategies are vectorized and grouped using a clustering module, forming pre-grouped strategy sets, and a mapping relationship is established in the strategy grouping database. Subsequently, the similarity between the feature vectors of the customer profile and the centroids of each strategy base group is calculated. The strategy base group that best matches the customer profile is selected, and the most suitable service strategy is determined as the initial service strategy based on service scenario complexity and adaptation score. For example, when an elderly customer submits a claim, the system may select a strategy base group with high fault tolerance and low complexity to reduce operational difficulty and cognitive burden.

[0193] Subsequently, the system combines the multi-dimensional dynamic profile with service stage identification information to form a composite state vector, which is then input into a pre-trained value network. The value network calculates a series of action combinations through forward propagation of the composite vector, each corresponding to operational suggestions such as channel switching, information push notifications, or resource allocation. For example, when the value network identifies that a customer is experiencing negative emotional fluctuations and a long waiting time during the claims process, it might output the action combination of "switching to a telephone agent + pushing a claims progress reminder + increasing human resource allocation."

[0194] Based on the action combination, the initial service strategy is optimized to generate a to-be-executed service strategy. In this process, the channel switching instruction in the action combination is mapped to update the service channel configuration, for example, switching from online self-service claims to manual telephone channel; the information push instruction is converted into inserting new message content, such as claim status reminder or supplementary material prompt; the resource allocation instruction adjusts the proportion of manpower and computing power in the service queue to prioritize processing of urgent or emotionally volatile customers. After these operations, the strategy is integrated into an executable form and time constraint conditions are added to ensure real-time response in the process of claims settlement or customer communication. Finally, the to-be-executed service strategy is stored in the strategy execution queue.

[0195] To ensure the reliability of the strategy in complex and abnormal environments, the system constructs abnormal scenario test data through a generative adversarial network. The user behavior generator constructs customer portrait mutations, such as frequently submitting claims materials in a short period of time or suddenly transferring to a new channel; the environment generator injects external interference, such as simulating system delay or channel congestion. These data are fused with random noise to form a complete abnormal test set, and drive the to-be-executed service strategy to run in a sandbox environment. Failure response events are recorded during execution, failure types are classified and counted, and the fault tolerance rate of the strategy is calculated, and finally a verification result report is output.

[0196] According to the verification result, the system adjusts the to-be-executed service strategy. If the fault tolerance rate is lower than the preset threshold, the defense mechanism is triggered. For system errors, increase the resource fuse mechanism to prevent service collapse; for strategy errors, insert decision tree verification nodes in the decision-making process to avoid inappropriate action execution; for environmental errors, add backup communication channels to ensure that messages can be delivered to customers. After enhancing the defense, the system generates a service strategy after defense enhancement, and re-verifies the fault tolerance rate change in local stress testing. If the fault tolerance rate reaches the optimization target, the strategy is confirmed as the final service strategy, otherwise the matching and enhancement process is repeated until the target is met.

[0197] The final service strategy is executed in actual operation, and customer interaction behavior, channel selection changes and emotional state fluctuations are continuously collected during the process to generate portrait change data. At the same time, service indicators such as processing time, customer satisfaction index and response delay are also recorded. These data are formatted and standardized and fed back to the portrait construction model and value network. The portrait model absorbs changes in customer behavior and preferences using incremental training, and the value network updates parameter weights based on service indicators, so as to have better adaptation ability in the next round of strategy generation. The whole process forms a closed loop, so that the insurance service strategy continuously evolves to better cope with the dynamics and complexity of customer demand.

[0198] In insurance business scenarios, the implementation of this process transforms the service process, which originally relied on human experience and static rules, into a dynamic, real-time, and personalized intelligent service. For example, when a high-net-worth client experiences anxiety during the claims process, the system can automatically detect this and immediately switch to a VIP human agent, along with claims progress updates and priority allocation of computing power, significantly improving customer satisfaction. For elderly clients with limited digital literacy, the system uses clustering to identify their profile characteristics, prioritizing low-complexity, high-fault-tolerance strategy templates and embedding backup channels, such as SMS or human telephone, to ensure unimpeded information transmission. The entire process not only reduces service costs in the claims process but also enables continuous optimization of service strategies through data feedback mechanisms, driving insurance services towards proactive adaptation and intelligent evolution.

[0199] In the healthcare field, the system first collects raw user interaction data from multiple channels, including event logs from health management applications, voice interactions during remote health consultations, user-uploaded health reports or images, and long-term accumulated personal health records. Application event logs reflect user actions within health management applications, such as dietary record input, exercise task completion status, or medication reminder responses. Voice interaction data is transcribed into text through voiceprint spectrum analysis and speech recognition, and combined with a semantic analysis model to extract user emotional fluctuations and language preferences during consultations. Uploaded health reports use character recognition technology to extract structured fields such as blood pressure, blood sugar, and heart rate values. Personal health records are formatted and converted into time-series data with timestamps, presenting long-term changes in health status. The system inputs this data into a health profile model to generate a multi-dimensional dynamic health profile that includes abnormal preference scores, channel preference distribution, and emotional tags.

[0200] After obtaining a health profile, the system loads a pre-defined set of health intervention strategies. This set includes service strategy templates for different types of users, covering exercise programs, dietary recommendations, and psychological support. The clustering module vectorizes these strategy templates based on sensitivity parameters, channel adaptation methods, and emotional response patterns, forming multiple strategy base groups, and establishes mapping relationships to store them in the strategy database. The system calculates the similarity between the health profile feature vector and the centroid coordinates of each base group, selects the most matching strategy base group, and chooses a strategy template based on the complexity of the health scenario to obtain the initial service strategy. For example, for users with significant mood swings, a strategy template with a high weighted emotional response pattern might be selected as the initial strategy.

[0201] Next, the system combines the multi-dimensional dynamic health portrait with health management stage identification information to generate a composite state vector and inputs it into the pre-trained value network. The value network performs forward propagation through neural network layers and outputs a set of action combinations, including diet intervention frequency adjustment, exercise plan switching, psychological counseling content pushing, etc. For example, when the system identifies that the user has recently been insufficient in exercise and has negative emotions, the action combination may include "switching exercise mode to low-intensity activity + pushing positive health content + increasing psychological support resource allocation".

[0202] Based on the action combination, the initial service strategy is optimized to form a health service strategy to be executed. The system analyzes the action combination, switches the exercise plan to a configuration with higher adaptability, inserts additional diet reminder content, adjusts the allocation ratio of psychological health resources, and integrates the optimized content into a complete service strategy, and adds time constraints for execution to ensure that the intervention measures take effect within a reasonable time range. Finally, the to-be-executed strategy is stored and enters the execution queue.

[0203] Before the strategy is executed, the system constructs abnormal scenario test data through the generative adversarial network, simulates user health portrait mutation (such as a large fluctuation in blood sugar level in a short period of time), environmental interference (such as failure to complete the exercise plan for a long time), and integrates noise to generate abnormal test sets. The to-be-executed strategy is loaded into the test sandbox for execution, and failure response events such as failure to trigger an early warning or invalid resource allocation are recorded during execution. The failure distribution is counted and the fault tolerance rate is calculated to generate a verification report.

[0204] According to the verification result, if the fault tolerance rate is lower than the threshold, the system will automatically adjust the to-be-executed strategy. For system-type errors, a resource fuse mechanism is added to avoid excessive invocation of health reminders; for strategy-type errors, decision tree verification nodes are inserted to ensure the rationality of action combinations; for environmental-type errors, backup channels are added, such as switching from mobile reminders to phone or email pushes. The defense-enhanced health service strategy is generated after enhancement and is executed again for local stress testing until the fault tolerance rate reaches the optimization target.

[0205] Finally, the enhanced service strategy is run in the real user health management process, executing exercise guidance, diet reminders, and emotional intervention, and collecting new behavior data in the process. The system generates portrait change data in real time, including changes in user exercise habits, improvements in diet adherence, and emotional label changes, while recording service indicators such as task completion rate, health improvement index, and user satisfaction. These data are formatted and input into the portrait model for incremental training, and the value network is updated to have higher adaptive ability in future strategy generation. In this way, the service system in the medical health scenario can continuously evolve and be more robust and personalized in dealing with long-term health management and sudden situations of users.

[0206] The embodiment combines fault tolerance rate index value driven diagnosis and structured implantation of three types of countermeasures, forms a closed loop for weakness positioning and repair, and no longer depends on large-scale rule rewriting but point enhancement for repair action, shortens the path from discovery to improvement; the resource fuse mechanism cuts off the resource avalanche path, the decision tree check node suppresses inappropriate action, the standby communication channel improves the information arrival rate, the three cooperate in the same strategy graph, so that the failure density and the continuous failure coefficient are synchronously reduced, and the fault tolerance rate change value can be quantitatively improved; the local stress test focuses on incremental changes, reduces verification cost and verification time delay, and at the same time introduces risks through version management and conflict detection control, and finally improves the robustness and maintainability of the online strategy without increasing unnecessary complexity.

[0207] In an embodiment, a service strategy generation and optimization apparatus is provided, which corresponds to the service strategy generation and optimization method in the above embodiment. Referring to Figure 3 , Figure 3 A functional module schematic diagram of a preferred embodiment of the service strategy generation and optimization apparatus of the present application. The image construction module 10, the strategy clustering module 20, the value network module 30, the strategy optimization module 40, the anomaly verification module 50, the defense enhancement module 60 and the feedback update module 70. The detailed description of each functional module is as follows:

[0208] The image construction module 10 is used to collect raw interaction data from multiple interaction channels, and generate a multi-dimensional dynamic image based on the raw interaction data through an image construction model;

[0209] The strategy clustering module 20 is used to load a preset strategy set and generate a pre-packaged strategy set through a clustering module, match the multi-dimensional dynamic image with the pre-packaged strategy set, and obtain an initial service strategy;

[0210] The value network module 30 is used to combine the multi-dimensional dynamic image and service stage identification information to generate a composite state vector, input the composite state vector into a pre-trained value network, and obtain an action combination;

[0211] The strategy optimization module 40 is used to optimize the initial service strategy through the action combination, and generate a to-be-executed service strategy;

[0212] The anomaly verification module 50 is used to construct anomaly scenario test data through a generative adversarial network, execute the to-be-executed service strategy to process the anomaly scenario test data, and generate a verification result;

[0213] The defense enhancement module 60 is used to adjust the to-be-executed service strategy based on the verification result, and obtain a final service strategy;

[0214] The feedback update module 70 is used to execute the final service strategy and record the profile change data and service indicators, and to feed back the profile change data and service indicators to the value network and profile construction model.

[0215] In one embodiment, the image construction module 10 is specifically used for:

[0216] Collect application event logs, parse user interface operation events in the application event logs, and generate a sequence of interface interactions.

[0217] Extract device type features, including operating system type, screen size factor, and device stability, from the application's event logs;

[0218] Collect voice call recordings, extract the voiceprint spectrum features of the voice call recordings, and use an automatic speech recognition engine to transcribe the voice call recordings into text dialogue data;

[0219] The system collects images and uploads them to documents, then extracts structured field information from the uploaded images using an optical character recognition interface.

[0220] Collect historical service records, and format the historical service records to generate time-series behavioral data with timestamps;

[0221] The interface interaction sequence and behavioral time-series data are input into the long short-term memory network of the profile construction model, and abnormal keyword time-series analysis is performed to obtain abnormal preference scores;

[0222] The structured field information and device type features are input into the profile to construct the gradient boosting decision tree of the model, determine the channel selection probability distribution, and generate the channel preference probability distribution;

[0223] The text dialogue data and voiceprint spectral features are input into the bidirectional encoder representation module of the portrait construction model to perform emotional semantic analysis and obtain emotional state labels.

[0224] Based on the abnormal preference score, channel preference probability distribution, and emotional state label, a multi-dimensional dynamic profile is constructed.

[0225] In one embodiment, the strategy clustering module 20 is specifically used for:

[0226] Load a preset policy set containing multiple service policy templates, and extract the abnormal sensitivity parameters, channel adaptation types, and sentiment response patterns of the service policy templates;

[0227] The clustering module performs feature vectorization processing on the anomaly sensitivity parameter, the channel adaptation type, and the emotion response pattern to generate a strategy feature vector.

[0228] performing a grouping operation to group the service policy templates into a plurality of policy base groups based on the policy feature vectors, and determining a centroid coordinate of each policy base group;

[0229] establishing an association relationship between the policy base groups and the service policy templates, generating a policy base group association mapping table, and storing the policy base groups, corresponding centroid coordinates, and the policy base group association mapping table into a policy grouping database;

[0230] quantifying an abnormal preference score in the multi-dimensional dynamic portrait into an abnormal feature value, mapping a channel preference probability distribution in the multi-dimensional dynamic portrait into a channel feature vector, and converting an emotional state label in the multi-dimensional dynamic portrait into an emotional feature code;

[0231] combining the abnormal feature value, the channel feature vector, and the emotional feature code to form a multi-dimensional dynamic portrait feature vector, determining a cosine similarity between the multi-dimensional dynamic portrait feature vector and the centroid coordinates of each policy base group, and selecting a policy base group with the highest cosine similarity as a target policy base group;

[0232] selecting a policy template as an initial service policy according to the service scene complexity and the adaptation score of the service policy template.

[0233] In an embodiment, the value network module 30 is specifically configured to:

[0234] obtain service stage identification information, parse a process node state in the service stage identification information, and generate a service stage code;

[0235] extract an abnormal preference score in the multi-dimensional dynamic portrait as an abnormal feature value;

[0236] extract a channel preference probability distribution in the multi-dimensional dynamic portrait as a channel dimension vector;

[0237] extract an emotional state label in the multi-dimensional dynamic portrait as an emotional dimension code;

[0238] splice the abnormal feature value, the channel dimension vector, the emotional dimension code, and the service stage code to generate an original composite vector;

[0239] perform a standardization process on the original composite vector to generate a composite state vector;

[0240] input the composite state vector into a pre-trained value network, perform a forward propagation process through a neural network layer of the value network, and obtain an action value distribution;

[0241] select an action combination with the maximum value expectation from the action value distribution.

[0242] In an embodiment, the policy optimization module 40 is specifically configured to:

[0243] parsing the channel switching instruction, the information pushing instruction and the resource allocation instruction in the action combination;

[0244] parsing the initial service policy to separate the service channel configuration, the service content sequence and the service resource ratio;

[0245] updating the service channel configuration in response to the channel switching instruction;

[0246] inserting the information pushing content into the service content sequence in response to the information pushing instruction;

[0247] adjusting the service resource ratio in response to the resource allocation instruction;

[0248] integrating the updated service channel configuration, the service content sequence with the inserted information pushing content and the adjusted service resource ratio to obtain an integrated policy;

[0249] adding an execution time constraint condition to the integrated policy to generate a to-be-executed service policy and storing the to-be-executed service policy in a policy execution queue.

[0250] In an embodiment, the anomaly verification module 50 is specifically configured to:

[0251] creating user portrait mutation data through a user behavior generator of a generative adversarial network;

[0252] injecting environmental interference data through an environment generator of the generative adversarial network;

[0253] fusing the user portrait mutation data and the environmental interference data and adding random noise to generate an anomaly scenario test data set;

[0254] loading the to-be-executed service policy into a test sandbox environment;

[0255] inputting the anomaly scenario test data set into the test sandbox environment;

[0256] processing the anomaly scenario test data set in the test sandbox environment by executing the to-be-executed service policy;

[0257] recording failure response events in the policy execution process and analyzing the failure type distribution of the failure response events;

[0258] determining a policy fault tolerance rate indicator and generating a verification result report containing the fault tolerance rate indicator and the failure type distribution.

[0259] In an embodiment, the defense enhancement module 60 is specifically configured to:

[0260] parsing the verification result report, extracting the fault tolerance rate index value and identifying the main error category;

[0261] when the fault tolerance rate index value is lower than a preset threshold, according to the main error category, matching a corresponding defense strategy from a preset defense strategy library;

[0262] for the system error category, adding a matched resource fusing mechanism to the to-be-executed service strategy;

[0263] for the strategy error category, inserting a matched decision tree verification node into the to-be-executed service strategy;

[0264] for the environment error category, adding a matched backup communication channel to the to-be-executed service strategy;

[0265] after adding the matched defense strategy to the to-be-executed service strategy, obtaining a defense-enhanced service strategy;

[0266] performing a local stress test on the defense-enhanced service strategy, and recording a fault tolerance rate change value;

[0267] when the fault tolerance rate change value reaches an optimization target, taking the defense-enhanced service strategy as a final service strategy;

[0268] when the fault tolerance rate change value does not reach the optimization target, re-executing the defense strategy matching and adding operations.

[0269] In one embodiment, a computer device is provided, which can be a server, and an internal structure diagram thereof can be as shown in Figure 4 The computer device includes a processor, a memory, a network interface and a database connected through a system bus. The processor of the computer device is used to provide determination and control capabilities. The memory of the computer device includes a non-volatile and / or volatile storage medium, an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external client through a network connection. The computer program is executed by the processor to implement the functions or steps of the server side of the service strategy generation and optimization method.

[0270] In one embodiment, a computer device is provided, which can be a client, and an internal structure diagram thereof can be as shown in Figure 5As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide determination and control capability. The memory of the computer device includes a non-volatile storage medium, an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The network interface of the computer device is used to communicate with the external server through the network connection. The computer program is executed by the processor to realize the function or step of the client side of the service strategy generation and optimization method.

[0271] In one embodiment, a computer device is provided, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor executing the computer program to implement the following steps:

[0272] Collecting raw interaction data from multiple interaction channels, generating a multi-dimensional dynamic portrait based on the raw interaction data through a portrait construction model;

[0273] Loading a preset strategy set and generating a pre-grouping strategy set through a clustering module, matching the multi-dimensional dynamic portrait with the pre-grouping strategy set to obtain an initial service strategy;

[0274] Combining the multi-dimensional dynamic portrait with service stage identification information to generate a composite state vector, inputting the composite state vector into a pre-trained value network to obtain an action combination;

[0275] Optimizing the initial service strategy through the action combination to generate a to-be-executed service strategy;

[0276] Constructing abnormal scenario test data through a generative adversarial network, executing the to-be-executed service strategy to process the abnormal scenario test data to generate a verification result;

[0277] Adjusting the to-be-executed service strategy based on the verification result to obtain a final service strategy;

[0278] Executing the final service strategy and recording portrait change data and service indicators, feeding back the portrait change data and service indicators to the value network and the portrait construction model.

[0279] In one embodiment, a computer readable storage medium is provided, which stores a computer program, the computer program being executed by a processor to implement the following steps:

[0280] Collecting raw interaction data from multiple interaction channels, generating a multi-dimensional dynamic portrait based on the raw interaction data through a portrait construction model;

[0281] Load a preset policy set and generate a pre-grouping policy set through a clustering module, match the multi-dimensional dynamic portrait with the pre-grouping policy set, and obtain an initial service policy;

[0282] Combine the multi-dimensional dynamic portrait with service stage identification information to generate a composite state vector, input the composite state vector into a pre-trained value network, and obtain an action combination;

[0283] Optimize the initial service policy through the action combination to generate a to-be-executed service policy;

[0284] Construct abnormal scene test data through a generative adversarial network, execute the to-be-executed service policy to process the abnormal scene test data, and generate a verification result;

[0285] Adjust the to-be-executed service policy based on the verification result to obtain a final service policy;

[0286] Execute the final service policy and record portrait change data and service indicators, and feed back the portrait change data and service indicators to the value network and the portrait construction model.

[0287] It should be noted that the functions or steps described above in relation to the computer readable storage medium or the computer device can correspond to the relevant descriptions of the server side and the client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0288] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiments of the methods can be included. Any reference to memory, storage, database or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0289] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above-mentioned functional units and modules is exemplified, and in actual application, the above-mentioned functions can be completed by different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above.

[0290] It should be noted that if non-company software tools or components appear in the embodiments of the present application, they are only used for example introduction and do not represent actual use. The above-described embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A method for generating and optimizing service strategies, characterized in that, Includes the following steps: Collect raw interaction data from multiple interaction channels, and generate multi-dimensional dynamic profiles based on the raw interaction data by building a profile model. Load a preset strategy set and generate a pre-grouped strategy set through a clustering module. Match the multi-dimensional dynamic profile with the pre-grouped strategy set to obtain the initial service strategy. The multi-dimensional dynamic profile and service stage identification information are combined to generate a composite state vector. The composite state vector is then input into a pre-trained value network to obtain an action combination. The initial service strategy is optimized by the combination of actions to generate a service strategy to be executed. Anomaly scenario test data is constructed by generating an adversarial network, the service to be executed is executed to process the anomaly scenario test data, and verification results are generated. The service strategy to be executed is adjusted based on the verification results to obtain the final service strategy; The final service strategy is executed and profile change data and service metrics are recorded. The profile change data and service metrics are then fed back to the value network and profile construction model.

2. The method for generating and optimizing service strategies as described in claim 1, characterized in that, Collect raw interaction data from multiple interaction channels, and generate a multi-dimensional dynamic profile based on the raw interaction data through a profile building model, including: Collect application event logs, parse user interface operation events in the application event logs, and generate a sequence of interface interactions. Extract device type features, including operating system type, screen size factor, and device stability, from the application's event logs; Collect voice call recordings, extract the voiceprint spectrum features of the voice call recordings, and use an automatic speech recognition engine to transcribe the voice call recordings into text dialogue data; The system collects images and uploads them to documents, then extracts structured field information from the uploaded images using an optical character recognition interface. Collect historical service records, and format the historical service records to generate time-series behavioral data with timestamps; The interface interaction sequence and behavioral time-series data are input into the long short-term memory network of the profile construction model, and abnormal keyword time-series analysis is performed to obtain abnormal preference scores; The structured field information and device type features are input into the profile to construct the gradient boosting decision tree of the model, determine the channel selection probability distribution, and generate the channel preference probability distribution; The text dialogue data and voiceprint spectral features are input into the bidirectional encoder representation module of the portrait construction model to perform emotional semantic analysis and obtain emotional state labels. Based on the abnormal preference score, channel preference probability distribution, and emotional state label, a multi-dimensional dynamic profile is constructed.

3. The method for generating and optimizing service strategies as described in claim 1, characterized in that, A preset policy set is loaded, and a pre-grouped policy set is generated through a clustering module. The multi-dimensional dynamic profile is matched with the pre-grouped policy set to obtain an initial service policy, including: Load a preset policy set containing multiple service policy templates, and extract the abnormal sensitivity parameters, channel adaptation types, and sentiment response patterns of the service policy templates; The clustering module performs feature vectorization processing on the anomaly sensitivity parameter, the channel adaptation type, and the emotion response pattern to generate a strategy feature vector. Perform a grouping operation to group the service policy template into multiple policy base groups based on the policy feature vector, and determine the centroid coordinates of each policy base group; Establish the association between the policy base group and the service policy template, generate a policy base group association mapping table, and store the policy base group, the corresponding centroid coordinates, and the policy base group association mapping table in the policy grouping database; The abnormal preference scores in the multidimensional dynamic profile are quantified into abnormal feature values, and the channel preference probability distribution in the multidimensional dynamic profile is mapped into a channel feature vector. The emotional state labels in the multidimensional dynamic profile are converted into emotional feature codes. The abnormal feature values, the channel feature vector, and the emotion feature encoding are combined to form a multidimensional dynamic profile feature vector. The cosine similarity between the multidimensional dynamic profile feature vector and the centroid coordinates of each policy base group is determined, and the policy base group with the highest cosine similarity is selected as the target policy base group. Select a strategy template as the initial service strategy based on the complexity of the service scenario and the adaptation score of the service strategy template.

4. The method for generating and optimizing service strategies as described in claim 1, characterized in that, Combining the multi-dimensional dynamic profile with the service stage identifier information, a composite state vector is generated. This composite state vector is then input into a pre-trained value network to obtain an action combination, including: Obtain service stage identification information, parse the process node status in the service stage identification information, and generate service stage code; Extract the abnormal preference score from the multidimensional dynamic profile as an abnormal feature value; Extract the channel preference probability distribution from the multi-dimensional dynamic profile as the channel dimension vector; The emotional state labels in the multidimensional dynamic portrait are extracted as the emotional dimension encoding; The abnormal feature value, the channel dimension vector, the sentiment dimension code, and the service stage code are concatenated to generate the original composite vector; The original composite vector is normalized to generate a composite state vector; The composite state vector is input into a pre-trained value network, and forward propagation is performed through the neural network layers of the value network to obtain the action value distribution. Select the action combination with the highest expected value from the action value distribution.

5. The method for generating and optimizing service strategies as described in claim 1, characterized in that, The initial service strategy is optimized by the combination of actions to generate a service strategy to be executed, including: Analyze the channel switching instruction, information push instruction, and resource allocation instruction in the action combination; The initial service strategy is analyzed to separate the service channel configuration, service content sequence, and service resource allocation. In response to the channel switching command, update the service channel configuration; In response to the information push instruction, insert information push content into the service content sequence; In response to the resource allocation instruction, adjust the service resource allocation ratio; By integrating the updated service channel configuration, the service content sequence with inserted push notifications, and the adjusted service resource allocation, a consolidated strategy is obtained. Add execution time constraints to the integrated strategy, generate a service strategy to be executed, and store the service strategy to be executed in the strategy execution queue.

6. The method for generating and optimizing service strategies as described in claim 1, characterized in that, Anomaly scenario test data is constructed by generating an adversarial network, and the service strategy to be executed is executed to process the anomaly scenario test data to generate verification results, including: User profile mutation data is created by a user behavior generator using a generative adversarial network; Injecting environmental interference data through an environment generator in a generative adversarial network; The user profile mutation data and the environmental interference data are combined, and random noise is added to generate an abnormal scenario test dataset; Load the service policy to be executed into the test sandbox environment; Input the abnormal scenario test dataset into the test sandbox environment; In the test sandbox environment, the service strategy to be executed is used to process the abnormal scenario test dataset; Record failure response events during strategy execution and analyze the distribution of failure types in the failure response events; Determine the strategy's fault tolerance metric and generate a verification result report that includes the fault tolerance metric and the distribution of failure types.

7. The method for generating and optimizing service strategies as described in claim 6, characterized in that, Based on the verification results, the service strategy to be executed is adjusted to obtain the final service strategy, including: Analyze the verification result report, extract the fault tolerance index value, and identify the main error categories; When the fault tolerance index value is lower than the preset threshold, the corresponding defense strategy is matched from the preset defense strategy library according to the main error category; For system error categories, add a matching resource circuit breaker mechanism to the policy of the service to be executed; For policy error categories, insert a matching decision tree verification node into the policy to be executed; For environmental error categories, add a matching backup communication channel to the service policy to be executed; After adding the matching defense strategy to the service strategy to be executed, the enhanced defense service strategy is obtained. Perform local stress tests on the enhanced defense service strategy and record the changes in fault tolerance rate; When the change in the fault tolerance rate reaches the optimization target, the enhanced defense service strategy will be used as the final service strategy. If the change in the fault tolerance rate does not reach the optimization target, the defense strategy matching and addition operations are re-executed.

8. A service strategy generation and optimization apparatus, characterized in that, The service strategy generation and optimization apparatus includes: The profile building module is used to collect raw interaction data from multiple interaction channels and generate multi-dimensional dynamic profiles based on the raw interaction data through the profile building model. The strategy clustering module is used to load a preset strategy set and generate a pre-grouped strategy set through the clustering module, and match the multi-dimensional dynamic profile with the pre-grouped strategy set to obtain an initial service strategy; The value network module is used to combine the multi-dimensional dynamic profile and service stage identification information to generate a composite state vector, and input the composite state vector into a pre-trained value network to obtain an action combination. The strategy optimization module is used to optimize the initial service strategy through the action combination and generate a service strategy to be executed. The anomaly verification module is used to construct anomaly scenario test data through a generative adversarial network, execute the service to be executed policy to process the anomaly scenario test data, and generate verification results. A defense enhancement module is used to adjust the service strategy to be executed based on the verification results to obtain the final service strategy. The feedback update module is used to execute the final service strategy and record profile change data and service metrics, and to feed back the profile change data and service metrics to the value network and profile construction model.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and a service policy generation and optimization program stored in the memory and executable on the processor. When the service policy generation and optimization program is executed by the processor, it implements the steps of the service policy generation and optimization method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a service policy generation and optimization program, which, when executed by a processor, implements the steps of the service policy generation and optimization method as described in any one of claims 1-7.

Images