Data disaster tolerance verification method and system based on AI technology

Abstract

The invention discloses a multi-node rapid disaster recovery switching method and system based on AI, and the method comprises the steps: collecting disaster recovery configuration information and backup data features, and building a data dependence relation and influence range association map; a directed acyclic graph of the fault combination is constructed, and a q-learning algorithm is adopted to optimize verification time sequence arrangement; performing risk scoring according to the verification result, and generating a verification report; training a local disaster recovery efficiency evaluation model based on verification results of the multiple data centers; carrying out weighted aggregation on model parameters by using a FedAvg algorithm through a federated learning mechanism, and generating a global evaluation model; and an evaluation result is fed back to the disaster recovery system for continuous optimization. According to the method, through collaborative modeling of all the data centers, dispersed professional knowledge is integrated on the premise of protecting data privacy, the verification coverage rate is increased, the service influence is minimized, the verification accuracy is enhanced, and the technical problems that a traditional disaster recovery verification method is low in efficiency, incomplete in coverage, incapable of sharing experience and the like are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of data disaster recovery technology, specifically to a data disaster recovery verification method and system based on AI technology, which is used to improve the intelligence and efficiency of data disaster recovery verification. Background Technology

[0002] Data disaster recovery verification is a crucial technical means to ensure the continuity of critical business operations. With the deepening of digital transformation, traditional disaster recovery verification methods are no longer sufficient to meet the complexity and real-time requirements of modern IT architectures. Currently, this field mainly involves core technical aspects such as backup data integrity verification, fault scenario simulation, and recovery capability assessment.

[0003] Existing disaster recovery verification technologies mainly include two types of solutions: script-based automated verification and rule-based fault injection. Script-based verification solutions periodically check backup data using predefined test cases, which can detect basic data corruption issues. Rule-driven fault injection solutions test the system's recovery capabilities by simulating preset fault types, thus improving the verification coverage to some extent.

[0004] The most mature disaster recovery verification technology currently employs a time-window-based batch verification mechanism. This solution sets a fixed verification cycle and performs large-scale data consistency checks and fault recovery drills during off-peak business periods. Its core principle is to divide the verification task into multiple independent verification units, execute them one by one in a serial or parallel manner, and finally summarize and generate a verification report.

[0005] However, existing technologies have two key problems: First, the verification efficiency is low. Traditional methods require a lot of manual intervention and have a long verification cycle, making it difficult to adapt to the 24 / 7 business continuity requirements. Second, the level of intelligence is insufficient. There is a lack of adaptive generation capabilities for complex fault scenarios and in-depth interpretation capabilities for verification results, resulting in poor accuracy of RTO / RPO prediction.

[0006] Therefore, it is necessary to develop a data disaster recovery verification method and system based on AI technology to improve the intelligence and efficiency of disaster recovery verification. Summary of the Invention

[0007] The purpose of this invention is to provide a data disaster recovery verification method and system based on AI technology, which aims to solve the technical problems of low verification efficiency and insufficient intelligence in existing disaster recovery verification technologies, and to achieve efficient and accurate data disaster recovery verification, thereby ensuring the continuity of enterprise critical business.

[0008] To achieve the above objectives, this invention provides a data disaster recovery verification method based on AI technology, comprising: acquiring backup data and historical fault data from multiple data centers; employing a variational autoencoder to extract features and perform semantic analysis on the backup data, constructing a multidimensional feature representation including metadata profiles and business semantic tags; based on the multidimensional feature representation and the historical fault data, using an Actor-Critic dual-network structure reinforcement learning strategy network to learn fault modes and dynamically generate scenario-based fault injection instructions that conform to the actual business distribution; according to the scenario-based fault injection instructions, constructing a directed acyclic graph of fault events and optimizing the timing arrangement through time-optimal directed q-analysis technology, executing scenario-based fault injection and acquiring verification data; using a pre-trained Transformer model to perform feature encoding on the verification data, and employing transfer learning technology and a causal inference model for intelligent analysis, outputting an interpretable verification report containing data integrity, consistency, and recoverability assessment results; based on the performance indicators and verification results in the interpretable verification report, using a federated learning framework to aggregate and update model parameters of each data center node, achieving cross-data center policy synchronization and performance optimization.

[0009] Furthermore, a variational autoencoder is used to perform feature extraction and semantic analysis on the backup data, constructing a multi-dimensional feature representation that includes metadata profiles and business semantic tags, including:

[0010] Based on the backup data, a variational autoencoder is used to compress the high-dimensional original backup data into a low-dimensional latent space to obtain the latent feature representation of the data.

[0011] The latent feature representation is then used to reconstruct the data using a decoder and calculate the reconstruction error, thereby learning the intrinsic structure and semantic features of the data.

[0012] Based on the aforementioned intrinsic structure and semantic features, a feature vector indexing system is established using an extensible Trie construction algorithm. The branching factor and tree depth are dynamically adjusted according to the data scale and query pattern to generate an efficient feature retrieval structure.

[0013] For the relationships between the backup data, a data relationship graph is constructed using graph embedding technology and vector mapping is performed using the Node2Vec algorithm to form business semantic tags containing topology information;

[0014] By integrating the latent feature representation, the feature retrieval structure, and the business semantic label, a unified feature representation framework is constructed to obtain the multidimensional feature representation.

[0015] Furthermore, a multi-dimensional feature representation is constructed, including metadata profiles and business semantic tags, comprising:

[0016] Based on the basic attributes and statistical characteristics of the backup data, a metadata profile containing data attributes, statistical characteristics, and quality indicators is generated through data quality analysis and importance assessment.

[0017] The business meaning of the backup data is understood and classified using natural language processing technology and domain knowledge base, and business semantic tags reflecting the business meaning and importance of the data are constructed.

[0018] Based on the dependencies and references between the backup data, graph theory algorithms are used to analyze data association patterns and calculate the scope of influence, and an association graph describing the data dependencies and scope of influence is established.

[0019] The metadata profile, the business semantic tags, and the association graph are uniformly encoded and standardized using a feature fusion algorithm to form the multidimensional feature representation.

[0020] Furthermore, a reinforcement learning strategy network with an Actor-Critic dual-network structure is used for fault mode learning to dynamically generate scenario-based fault injection instructions that conform to the distribution of real business operations, including:

[0021] Based on the fault types, frequency of occurrence, and scope of impact in the historical fault data, a fault knowledge base containing the correlation between fault occurrence patterns and business characteristics is constructed through data mining and pattern recognition to obtain prior knowledge of fault patterns.

[0022] Using prior knowledge of the aforementioned fault modes and the current system state, a strategy decision is made through an Actor network, including fault type selection, impact scope determination, and injection timing control, to generate an initial fault injection strategy.

[0023] For the initial fault injection strategy, the Critic network analysis is used to verify the effectiveness, business impact and recovery performance after the strategy is executed, and the strategy value assessment results are output.

[0024] Based on the strategy value evaluation results, the parameters of the Actor network and Critic network are updated through the strategy gradient algorithm to optimize the fault injection strategy generation capability.

[0025] Based on the optimized network parameters and current verification requirements, the Monte Carlo tree search algorithm is used to explore the optimal fault combination scheme and generate the scenario-based fault injection instruction.

[0026] Furthermore, a directed acyclic graph of fault events is constructed using time-optimal directed q-analysis techniques, and the timing arrangement is optimized. Contextualized fault injection is then performed, and verification data is obtained, including:

[0027] Based on the fault events in the scenario-based fault injection instructions, the causal relationships and temporal constraints between the events are analyzed, and a directed acyclic graph with fault events as nodes and dependencies as edges is constructed.

[0028] For the directed acyclic graph, the q-learning algorithm is used to learn the verification effect of fault combinations under different time windows, and the optimal timing arrangement that maximizes verification coverage and minimizes business impact is calculated.

[0029] According to the optimal timing arrangement, the fault injection operation is executed sequentially in chronological order, and the system response is monitored in real time to collect verification data including system state changes, recovery process and performance indicators.

[0030] Furthermore, a pre-trained Transformer model is used for feature encoding, and transfer learning techniques and causal inference models are employed for intelligent analysis. The resulting interpretable verification report, including assessments of data integrity, consistency, and recoverability, is output.

[0031] Based on the verification requirements, data features, and fault scenario information in the verification data, the self-attention mechanism of the pre-trained Transformer model is used to encode multimodal information and generate a unified vector representation.

[0032] Using the unified vector representation, feature extraction is performed on a base model pre-trained on a large-scale disaster recovery scenario dataset, and fine-tuning techniques are used to adjust parameters for the current specific verification scenario to obtain an adaptively enhanced feature representation.

[0033] Based on the adaptive enhancement feature representation, a structured causal model is used to analyze the causal relationship between data integrity, consistency and recoverability indicators, and to identify the root causes and critical paths affecting disaster recovery effectiveness.

[0034] Based on the critical path and the current verification scenario, the graph reordering approximate nearest neighbor search technique is used to match similar historical verification cases to obtain verification decision support information.

[0035] By integrating the root cause, the critical path, and the verification decision support information, an interpretable verification report containing verification conclusions, key indicators, reasoning processes, and key evidence is generated based on an attention mechanism, thus obtaining the interpretable verification report.

[0036] Furthermore, the federated learning framework is used to aggregate and update model parameters across data center nodes, enabling cross-data center policy synchronization and performance optimization, including:

[0037] At each data center node, a verification result collector and performance monitoring component are deployed to collect performance metrics such as verification accuracy, execution efficiency, resource consumption, and fault detection rate in real time, thereby obtaining local training data.

[0038] Based on the local training data, each node independently performs incremental training of the validation model and calculates the gradient update of the model parameters to generate local gradient information.

[0039] The local gradient information is securely processed using homomorphic encryption and differential privacy techniques, and then uploaded to the central coordinator via a secure aggregation protocol.

[0040] In the central coordinator, the FedAvg algorithm is used to perform weighted aggregation based on the data quality, number of training samples and historical contribution of each node to generate updated global model parameters.

[0041] The updated global model parameters are securely distributed to each data center node and merged with the local model to achieve cross-data center policy synchronization and performance optimization.

[0042] Furthermore, it also includes integrating scalable Trie construction algorithms to improve data retrieval efficiency:

[0043] Based on the access frequency, size, and type characteristics of backup data, a multi-level index structure and an adaptive sharding strategy are designed using an extensible Trie construction algorithm to build a dynamically optimized data index system.

[0044] Based on the data scale and memory resource status, the branching factor and depth of the Trie tree are dynamically adjusted, and Bloom filters and caching mechanisms are integrated to achieve millisecond-level data location and retrieval.

[0045] Furthermore, it also includes using graph reordering approximate nearest neighbor search techniques for fault mode identification:

[0046] Historical failure cases and verification results are represented as high-dimensional feature vectors, and a similarity calculation framework based on graph structure is constructed to establish a failure mode knowledge base.

[0047] Based on the fault mode knowledge base and current fault characteristics, the accuracy and efficiency of similarity calculation are optimized by graph reordering algorithm, which can quickly identify fault scenarios with similar patterns and provide decision support.

[0048] This invention also provides a data disaster recovery verification system based on AI technology, including a processor and a memory, wherein the processor is used to execute the above-described method.

[0049] The present invention also provides a data disaster recovery verification system based on AI technology, including a processor and a memory, wherein the processor is used to execute the above-described method.

[0050] The beneficial effects of this invention include:

[0051] 1. By using AI technology to extract features and perform semantic analysis on backup data, a multi-dimensional feature representation is constructed, enabling a deep understanding and accurate identification of the data, providing a comprehensive and accurate data foundation for subsequent intelligent verification and decision-making.

[0052] 2. Based on a reinforcement learning policy network, fault scenarios that conform to the distribution of real business are dynamically generated, realizing adaptive scenario-based fault injection and improving the authenticity and coverage of verification.

[0053] 3. By employing transfer learning and causal reasoning models, the validation data is intelligently analyzed, and an interpretable validation report is output, which enhances the credibility and operability of the validation results.

[0054] 4. By using the federated learning framework, cross-datacenter policy synchronization and performance optimization are achieved, improving overall performance while ensuring data privacy.

[0055] 5. By integrating time-optimal directed q-analysis, scalable Trie construction algorithm, and graph reordering approximate nearest neighbor search, the overall verification efficiency and accuracy of the system are significantly improved. Compared with traditional methods, the verification time can be shortened by 60-80%, the verification efficiency can be improved by 3-5 times, and the prediction error of RTO and RPO can be controlled within 5%. Attached Figure Description

[0056] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the embodiments will be briefly described below. These drawings are incorporated in and constitute a part of this specification. They illustrate embodiments conforming to this disclosure and, together with the specification, serve to explain the technical solutions of this disclosure. It should be understood that the following drawings only show some embodiments of this disclosure and should not be considered as limiting the scope. Those skilled in the art can obtain other related drawings based on these drawings without creative effort.

[0057] Figure 1 This is a flowchart of a data disaster recovery verification method based on AI technology according to the present invention;

[0058] Figure 2 This is a detailed flowchart of S1 of the present invention;

[0059] Figure 3 This is a flowchart illustrating how the federated learning framework of this invention achieves cross-datacenter policy synchronization and performance optimization. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. The components of the embodiments of this disclosure described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this disclosure provided in the accompanying drawings is not intended to limit the scope of the claimed disclosure, but merely represents selected embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without inventive effort are within the scope of protection of this disclosure.

[0061] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0062] In this document, the term "and / or" merely describes a relationship, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0063] like Figure 1 As shown, this invention provides a data disaster recovery verification method based on AI technology, including the following steps:

[0064] Step S1: Obtain backup data and historical fault data from multiple data centers, and use a variational autoencoder to perform feature extraction and semantic analysis on the backup data to construct a multi-dimensional feature representation that includes metadata profiles and business semantic tags.

[0065] In this step, the AI ​​data profiling module first performs in-depth analysis and feature mining on massive backup data. The system uses a variational autoencoder (VAE) as its core algorithm. The encoder compresses the high-dimensional original backup data into a low-dimensional latent space, and then the decoder reconstructs the data. In this process, the system automatically learns the inherent structure and semantic features of the data. The probabilistic modeling characteristics of the variational autoencoder enable it to capture the uncertainty of data distribution, providing an important basis for subsequent anomaly detection and risk assessment.

[0066] To achieve efficient management and rapid retrieval of extracted features, the system introduces a scalable Trie construction algorithm to establish an indexing system for feature vectors. The core innovation of this algorithm lies in its dynamic optimization strategy, which can adjust the branching factor and tree depth of the Trie tree in real time based on data scale and query patterns. When the data volume increases, the algorithm automatically increases the branching factor to reduce the tree depth, thereby maintaining query performance; when memory resources are scarce, it appropriately increases the tree depth to reduce memory consumption. This adaptive mechanism ensures that the system maintains optimal query efficiency under different scales and resource conditions.

[0067] In the semantic tagging stage, the system uses graph embedding technology to model the complex relationships between backup data. By constructing a data relationship graph, each data entity is treated as a node in the graph, and dependencies, references, and temporal sequences between data are treated as edges. Then, graph embedding algorithms such as Node2Vec or GraphSAGE are used to map these relationships to a low-dimensional continuous vector space. This mapping not only preserves the topological structure information between data but also discovers implicit business semantic relationships, providing rich feature representations for data classification, clustering, and anomaly detection.

[0068] Ultimately, the system integrates metadata profiles, business semantic tags, and relationship graphs to form a multi-dimensional feature representation system for backup data. The metadata profile includes the data's basic attributes, statistical characteristics, and quality indicators; the business semantic tags reflect the data's business meaning and importance; and the relationship graph describes the dependencies and scope of influence between data points. This multi-dimensional feature representation provides a comprehensive and accurate data foundation for subsequent intelligent verification and decision-making.

[0069] Step S2: Based on the multidimensional feature representation and the historical fault data, a reinforcement learning strategy network with an Actor-Critic dual-network structure is used to learn fault modes and dynamically generate scenario-based fault injection instructions that conform to the real business distribution.

[0070] The scenario-based fault injection engine employs deep reinforcement learning to learn and simulate fault patterns in real-world environments. The system analyzes historical fault data, business operation logs, and system monitoring metrics to construct a fault knowledge base encompassing dimensions such as fault type, frequency of occurrence, scope of impact, and recovery time. Based on this historical data, the reinforcement learning network can learn the inherent patterns of fault occurrence and the correlation patterns between business characteristics.

[0071] In terms of network architecture design, the system adopts an Actor-Critic dual-network structure to separate policy learning and value assessment. The Actor network is responsible for generating fault injection strategies based on the current system state and business characteristics, including decisions such as fault type selection, impact scope determination, and injection timing control. The Critic network evaluates the value of the strategies generated by the Actor, judging the effectiveness of the strategies by analyzing the verification results, business impact, and recovery performance after strategy execution. The two networks continuously optimize through interactive learning, ensuring that the generated fault scenarios are both realistic and have verification value.

[0072] To accurately model the temporal dependencies between fault events, the system introduces a time-optimal directed q-analysis technique. This technique first constructs a directed acyclic graph (DAG) to represent the causal relationships and temporal constraints between different fault events. Nodes in the graph represent specific fault events, and edges represent the dependencies and temporal order between events. Then, a q-learning algorithm is used to optimize the time series of fault injection. By learning the verification effects of fault combinations under different time windows, the optimal timing arrangement that maximizes verification coverage and minimizes business impact is found.

[0073] In terms of strategy space exploration, the system employs the Monte Carlo tree search algorithm to find the optimal fault combination scheme. This algorithm constructs a search tree to represent different fault combination paths, where each node represents a fault state and edges represent the addition or change of fault events. The algorithm iteratively optimizes through four steps: selection, expansion, simulation, and backpropagation. While balancing exploration and utilization, it gradually converges to a fault combination scheme that can fully verify the system's disaster recovery capability.

[0074] Ultimately, the system can generate scenario-based fault injection commands covering various types of faults, including hardware failures, network interruptions, data corruption, human error, and natural disasters. These commands are not only comprehensive in terms of fault types, but also highly consistent with real business environments in terms of occurrence probability, impact, and timing, providing realistic and reliable test scenarios for disaster recovery verification.

[0075] Step S3: Based on the scenario-based fault injection instruction, construct a directed acyclic graph of fault events and optimize the timing arrangement using time-optimal directed q-analysis technology, execute scenario-based fault injection, and obtain verification data.

[0076] Based on the fault events in the scenario-based fault injection instructions, the system first analyzes the causal relationships and timing constraints between events, and constructs a directed acyclic graph with fault events as nodes and dependencies as edges. This graph clearly describes the sequential dependencies between fault events, ensuring that the fault injection process conforms to the logical order in the actual business scenario.

[0077] For the constructed directed acyclic graph, the system employs a q-learning algorithm to learn the verification effectiveness of fault combinations under different time windows. The algorithm calculates the optimal timing arrangement that maximizes verification coverage and minimizes business impact by evaluating verification coverage, business impact, and resource consumption under different timing schemes. This optimization not only considers the comprehensiveness of verification but also takes into account the control of impact on normal business operations, achieving a balance between verification effectiveness and business continuity.

[0078] According to the optimal timing schedule, the system executes fault injection operations sequentially. During execution, the system monitors the system response in real time, collecting multi-dimensional verification data including system state changes, recovery processes, and performance metrics. This data comprehensively records the system's behavior and recovery capabilities in the face of different types of faults, providing rich evidence for subsequent intelligent analysis.

[0079] By employing time-optimal directed q-analysis, the system can minimize the interference of the verification process on normal business operations while ensuring verification coverage. Compared to traditional linear or random fault injection methods, this method based on directed graphs and reinforcement learning can more accurately simulate real-world fault scenarios, improving the accuracy and efficiency of verification.

[0080] Step S4: The verification data is feature-encoded using a pre-trained Transformer model, and intelligent analysis is performed using transfer learning techniques and causal inference models to output an interpretable verification report containing assessment results of data integrity, consistency, and recoverability.

[0081] The intelligent verification decision center first employs a pre-trained Transformer model to perform feature encoding and semantic understanding of complex verification tasks. Leveraging its powerful self-attention mechanism, the Transformer model captures long-distance dependencies between different components in the verification task and encodes multimodal information such as verification requirements, data features, and fault scenarios into a unified vector representation. This encoding not only preserves the semantic content of the original information but also enhances the correlation between different pieces of information, providing high-quality input features for subsequent intelligent decision-making.

[0082] Regarding model adaptability, the system fully leverages the knowledge gained from pre-training on large-scale datasets through transfer learning techniques. Specifically, the system first trains a base model on a large dataset encompassing various disaster recovery scenarios, learning general validation patterns and decision-making rules. Then, for the specific validation scenario, fine-tuning techniques are used to locally adjust the model parameters, enabling the model to quickly adapt to new data distributions and business characteristics. This transfer learning approach not only significantly reduces model training time but also improves validation accuracy in small-sample scenarios.

[0083] To gain a deeper understanding of the intrinsic relationships among validation metrics, the system employs a causal reasoning framework to establish a causal relationship model for these metrics. By analyzing the causal chains among key metrics such as data integrity, consistency, and recoverability, the system can identify the root causes and critical paths affecting disaster recovery effectiveness. This structured causal model not only quantifies the contribution of different factors to the validation results but also predicts the changing trends of validation metrics under specific intervention measures, providing a scientific basis for optimizing disaster recovery strategies.

[0084] In terms of verification decision support, the system employs graph reordering approximate nearest neighbor search technology to intelligently match the current verification scenario with historical verification cases. This technology first represents verification cases as high-dimensional feature vectors, then constructs an index system based on a graph structure, and optimizes the accuracy and efficiency of similarity calculation through a reordering algorithm. When faced with a new verification task, the system can quickly retrieve the most similar historical cases and provide reference for current decisions based on the verification results and processing experience of these cases.

[0085] Ultimately, the system generates a highly interpretable verification report based on the attention mechanism. The report not only includes verification conclusions and key metrics, but also details the reasoning process and crucial evidence used to arrive at those conclusions. By visually displaying the attention weight distribution, users can clearly understand which factors significantly impact the verification results and how these impacts are transmitted and accumulated. This interpretable design greatly improves the credibility and operability of the verification results.

[0086] Step S5: Based on the performance metrics and verification results in the interpretability verification report, the model parameters of each data center node are aggregated and updated through the federated learning framework to achieve cross-data center policy synchronization and performance optimization.

[0087] The self-healing optimization closed-loop module constructs a distributed continuous learning system that enables knowledge sharing and collaborative optimization across data centers while ensuring data privacy and security. The system first deploys verification result collectors and performance monitoring components on each data center node to collect key performance indicators in real time, including verification accuracy, execution efficiency, resource consumption, and fault detection rate, as well as valuable experience such as new fault modes and abnormal data characteristics discovered during the verification process.

[0088] In the implementation of the federated learning framework, each data center node independently trains its model based on locally collected data, without transmitting raw data to the central server. Each node uses local data to incrementally train the validation model, calculates gradient updates for model parameters, and then uploads this gradient information to the central coordinator via a secure aggregation protocol. The secure aggregation protocol employs homomorphic encryption and differential privacy technologies to ensure that sensitive information from each node is not leaked during parameter transmission.

[0089] After receiving gradient information uploaded by each node, the central coordinator intelligently aggregates these gradients using the FedAvg (Federated Average) algorithm. This algorithm is not a simple averaging operation, but rather a weighted average based on factors such as data quality, number of training samples, and historical contribution of each node, ensuring that the experience of high-quality nodes receives greater weight. Simultaneously, the coordinator performs consistency checks and anomaly detection on the aggregated parameters, filtering out potentially malicious updates or anomalous parameters.

[0090] During the parameter distribution phase, the central coordinator securely distributes the updated global model parameters to each data center node. Upon receiving the new parameters, each node integrates them with its local model and fine-tunes them according to local business characteristics. This combination of global parameters and local adaptation ensures both the model's universality and the personalized needs of each data center.

[0091] Through this closed-loop mechanism of continuous learning and optimization, the system can continuously accumulate and share best practices for disaster recovery verification, thereby continuously improving the intelligence level of the entire distributed verification network. When a new failure mode or verification method is discovered in a data center, this knowledge can be quickly propagated to other nodes, achieving collaborative evolution and performance optimization of the entire network.

[0092] The following is a specific example illustrating the implementation process of the AI-based data disaster recovery verification method of the present invention:

[0093] A financial institution has three data centers located in different regions and needs to comprehensively verify the data disaster recovery capabilities of its core transaction system. The system first executes step S1 to obtain backup data and historical fault data from these three data centers. The backup data includes critical business data such as the core transaction database, customer information data, and transaction records, totaling approximately 15TB. The historical fault data contains records of various faults that occurred over the past two years, including detailed descriptions and handling processes for network outages, storage failures, and data corruption.

[0094] The system employs a variational autoencoder to extract features and perform semantic analysis on these backup data. First, information such as the database table structure, field types, and data distribution is input into the encoder, compressing it into a 128-dimensional latent space to obtain a low-dimensional representation of the data. For example, for customer transaction logs, the system captures characteristics such as high data update frequency, numerous related tables, and high business value. Then, the data is reconstructed using a decoder, the reconstruction error is calculated, and the inherent structure of the data is learned. The system discovers a strong correlation between the transaction logs and the account balance tables, and that data integrity has a significant impact on business continuity.

[0095] Next, based on these learned features, the system employs a scalable Trie construction algorithm to build a feature index, setting the branching factor to 16 and the tree depth to 4 levels to achieve fast feature retrieval. The system also uses the Node2Vec algorithm to analyze the relationships between data, constructing a data relationship graph containing 2000 nodes and 15000 edges, and mapping it to a 64-dimensional vector space to form business semantic tags containing topological structure information. Finally, the system integrates these feature representations to construct a multi-dimensional feature representation including metadata profiles and business semantic tags, laying the foundation for subsequent analysis.

[0096] In step S2, based on these multidimensional feature representations and historical fault data, the system employs a reinforcement learning strategy network with an Actor-Critic dual-network structure to learn fault modes. The system first analyzes historical fault data, finding that storage faults account for 35%, network faults for 28%, data corruption for 20%, and other faults for 17%, and identifies that faults mostly occur during maintenance windows following peak business periods. These statistical features are used to construct a fault knowledge base as prior knowledge for the strategy network.

[0097] The Actor network generates an initial fault injection strategy based on the current system state and prior knowledge. For example, the system decides to simulate a storage controller failure in the Guangzhou data center, affecting 30% of the transaction database's storage volumes, with the execution scheduled during the system maintenance window at 2 AM the following day. The Critic network evaluates the value of this strategy, finding that it covers 73% of critical data paths, has a manageable impact on business operations, and has moderate recovery difficulty, giving it a high value score of 0.82. Based on this evaluation, the system updates the network parameters using a policy gradient algorithm, and after approximately 500 iterations of training, the network converges to a stable state.

[0098] Ultimately, the system employed a Monte Carlo tree search algorithm to explore an optimal combination of five key fault scenarios: 1) storage controller failure in the Guangzhou data center; 2) network partitioning in the Shanghai data center; 3) data file corruption in the Beijing data center; 4) cross-center data synchronization interruption; and 5) errors in the backup system index table. The system generated detailed scenario-based fault injection instructions, including specific parameters, execution conditions, and expected impacts for each fault.

[0099] In step S3, the system constructs a directed acyclic graph of fault events using time-optimal directed q-analysis based on the generated scenario-based fault injection instructions. The graph contains 5 main fault nodes and 12 dependency edges; for example, data synchronization interruption must be executed after network partitioning, and backup index error must be executed after data file corruption. The system uses a q-learning algorithm to optimize the execution timing, learning the optimal execution order within a 4-hour verification window: first, the storage controller fault (t=0 min), then the network partition (t=40 min), followed by data file corruption (t=90 min), then the data synchronization interruption (t=120 min), and finally the backup index error (t=180 min).

[0100] Following this sequence, the system executed fault injection operations sequentially. During fault execution, the system monitored over 200 metrics in real time, including system response time, resource utilization, error logs, and data consistency check results. For example, during a storage controller fault, the system recorded an increase in storage I / O latency from 2ms to 120ms, a rise in database query response time from 50ms to 2.5s, and an automatic failover time of 28 seconds, during which approximately 1% of transaction requests failed. All this monitoring data was collected into a structured validation dataset for subsequent analysis.

[0101] In step S4, the system uses a pre-trained Transformer model to encode features from the collected validation data. The model has 12 attention heads and 768 hidden layer dimensions, enabling it to handle multimodal data. The system inputs monitoring metrics, log content, and fault event sequences into the model to generate a unified vector representation. Then, the system uses a pre-trained base model on 100 disaster recovery scenarios and 10,000 hours of validation data for feature extraction. Furthermore, it fine-tunes the model using transfer learning techniques to better adapt it to the characteristics of the current financial trading system, making the model more suitable for the features of financial data.

[0102] The system employs a structured causal model to analyze the causal relationships among data integrity, consistency, and recoverability indicators. The analysis revealed that data write delays caused by storage controller failures are the primary cause of data consistency issues (contributing 62%), while backup index errors are a key factor affecting recovery time (contributing 78%). The system identified the critical causal path of "storage failure → data write delay → transaction rollback → data inconsistency," providing a clear direction for optimizing disaster recovery strategies.

[0103] The system also used graph reordering approximate nearest neighbor search technology to match three highly similar failure scenarios from the historical case library, including a real case that occurred last year at another financial institution. These cases provided valuable handling experience, such as the suggestion that "in the event of a storage controller failure, priority should be given to ensuring the complete writing of the transaction consistency log, and a short delay in other non-critical data is acceptable."

[0104] Finally, the system integrated the analysis results and generated a detailed interpretability verification report. The report showed a data integrity score of 92 out of 100, a consistency score of 85, and a recoverability score of 89. The report identified two main risk points in the system: first, the cross-center data synchronization mechanism lacked sufficient retry mechanisms during network fluctuations, resulting in approximately 3% of data synchronization delays exceeding the RPO target; second, the backup system's index update strategy was not flexible enough, potentially leading to index inconsistencies under high-concurrency scenarios. The report also provided specific optimization recommendations, including improving the synchronization protocol and optimizing the index update strategy.

[0105] In step S5, based on the results of the verification report, the system aggregates and updates the model parameters of each data center node using a federated learning framework. Each data center node independently deploys a verification result collector and a performance monitoring component to collect local verification metrics. For example, the Beijing node recorded a detection accuracy of 96%, a recovery time of 35 minutes, and a peak resource consumption of 80% in a data file corruption scenario; the Shanghai node demonstrated an 87% fault detection rate and a recovery time of 42 minutes in a network partition scenario.

[0106] Each node independently trains and validates its model based on local data. The Beijing node, after 200 rounds of training, showed a significant improvement in its ability to identify network partition scenarios, as indicated by the model parameter gradient. The Shanghai node, after 180 rounds of training, improved its accuracy in selecting data recovery strategies by 12%. The Guangzhou node, after 220 rounds of training, demonstrated outstanding performance in storage fault prediction. This local gradient information was uploaded to the central coordinator after being homomorphically encrypted.

[0107] The central coordinator uses the FedAvg algorithm for weighted aggregation, assigning weights based on the data quality and training sample size of each node. The weights for the Beijing, Shanghai, and Guangzhou nodes are 0.4, 0.35, and 0.25, respectively. The aggregated global model parameters improved the validation scenario coverage by 15% and the fault prediction accuracy by 9%, outperforming the model of any single node.

[0108] The updated global model parameters are securely distributed to each node and fused with the local model. For example, the Guangzhou node uses a global parameter weight of 0.7 and a local parameter weight of 0.3 for fusion, and then fine-tunes it using 5,000 local sample data points to better adapt the model to the characteristics of the local environment. In this way, the three data centers form a closed-loop system of collaborative learning and continuous optimization, continuously improving disaster recovery capabilities, and allowing each node to benefit from the experience of other nodes while maintaining data privacy.

[0109] This complete verification cycle was completed within 12 hours, saving approximately 70% of the time compared to traditional methods. The system accurately predicted the Recovery Time Objective (RTO) to be 45 minutes (the actual recovery time was 43 minutes), with a prediction error of only 4.7%, far below the industry average. More importantly, the risk points identified by the system were validated in real-world business operations a month later—a network fluctuation did indeed cause a data synchronization problem, but because the system had optimized the synchronization protocol in advance, it successfully avoided potential data inconsistencies. This successful disaster recovery verification not only demonstrated the system's effectiveness but also provided strong assurance for the business continuity of financial institutions.

[0110] In another embodiment, such as Figure 2 As shown, a variational autoencoder is used to extract features and perform semantic analysis on the backup data, constructing a multi-dimensional feature representation that includes metadata profiles and business semantic tags, including:

[0111] Step S1.1: Based on the backup data, use a variational autoencoder to compress the high-dimensional original backup data into a low-dimensional latent space to obtain the latent feature representation of the data.

[0112] The system first preprocesses the backup data, including data cleaning, standardization, and normalization, to ensure the quality and consistency of the input data. Then, the preprocessed high-dimensional raw backup data is input into the encoder network of the variational autoencoder. The encoder network typically consists of multiple layers of neural networks, which map the high-dimensional data to a low-dimensional latent space through layer-by-layer dimensionality reduction. In this process, the encoder not only learns the compressed representation of the data but also captures the probability distribution characteristics of the data, giving the latent feature representation good generalization and expressive power.

[0113] Latent feature representations are typically low-dimensional vectors, with dimensions much smaller than the original data, but containing key information and the inherent structure of the original data. This dimensionality reduction process not only improves the efficiency of subsequent processing but also filters out noise and redundant information in the data, extracting truly valuable features.

[0114] Step S1.2: Reconstruct the data using a decoder and calculate the reconstruction error for the latent feature representation, thereby learning the intrinsic structure and semantic features of the data.

[0115] The latent feature representations obtained in step S1.1 are input into the decoder network of the variational autoencoder. The decoder network attempts to remap the low-dimensional features back to the original data space, generating reconstructed data. The system calculates the reconstruction error by comparing the original data and the reconstructed data. This error reflects the model's ability to capture the original data structure.

[0116] During training, the system optimizes model parameters by minimizing the weighted sum of reconstruction error and KL divergence (a measure of the difference between the latent feature distribution and the standard normal distribution). This training method allows the model to learn the intrinsic structure and semantic features of the data while ensuring reconstruction accuracy.

[0117] By analyzing the distribution and characteristics of reconstruction errors, the system can identify abnormal patterns and special structures in the data, providing an important basis for subsequent anomaly detection and risk assessment. At the same time, the features in the latent space also have clear semantic meanings, which can be used for semantic understanding and classification of the data.

[0118] Step S1.3: Based on the intrinsic structure and semantic features, a feature vector index system is established using an extensible Trie construction algorithm. The branching factor and tree depth are dynamically adjusted according to the data scale and query mode to generate an efficient feature retrieval structure.

[0119] Based on the intrinsic structure and semantic features learned in step S1.2, the system uses a scalable Trie construction algorithm to establish an indexing system for feature vectors. A Trie tree is a tree-like data structure suitable for prefix queries and fast retrieval. By using each dimension of the feature vector as a different level of the tree, it achieves efficient indexing of multi-dimensional features.

[0120] The key innovation of the algorithm lies in its dynamic optimization strategy: the system adaptively adjusts the branching factor and tree depth of the Trie tree based on the actual data scale and query pattern. When processing large-scale data, the algorithm increases the branching factor to reduce the tree depth, thereby reducing the query path length and improving retrieval speed; when memory resources are limited, it appropriately increases the tree depth to reduce the number of branches and reduce memory usage.

[0121] In addition, the algorithm integrates Bloom filters and caching mechanisms to further optimize retrieval performance. Bloom filters are used to quickly determine whether a feature is likely to exist in the index, avoiding unnecessary tree traversals; the caching mechanism records the results of hot queries, reducing redundant calculations.

[0122] With this scalable index structure, the system can complete the retrieval of specific features in milliseconds, significantly improving the efficiency of data processing and verification.

[0123] Step S1.4: For the relationships between the backup data, a data relationship graph is constructed using graph embedding technology and vector mapping is performed using the Node2Vec algorithm to form business semantic tags containing topology information.

[0124] The system analyzes the relationships between backup data, including references, dependencies, and temporal sequences, and constructs a complex network graph with data entities as nodes and relationships as edges. This data relationship graph comprehensively reflects the topology and interaction patterns of the data ecosystem.

[0125] For the constructed data relationship graph, the system uses the Node2Vec graph embedding algorithm for vector mapping. Node2Vec is a graph embedding method based on random walks. It samples the context of nodes by performing biased random walks on the graph, and then learns the vector representation of the nodes using a method similar to Word2Vec.

[0126] The key parameters of the algorithm are p and q, which control the walk strategy: p controls the probability of a random walk returning to the previous node, and q controls the tendency of a random walk to explore unseen nodes. By adjusting these two parameters, a balance can be achieved between a depth-first strategy (exploring local structure) and a breadth-first strategy (exploring global structure), thereby capturing network structure information at different scales.

[0127] The vector representation obtained by mapping not only preserves the topological structure information between data, but also reflects the business semantic relationships between the data. In the vector space, semantically similar data will cluster together, forming natural semantic clusters, providing an intuitive semantic understanding for subsequent data analysis and decision-making.

[0128] Step S1.5: Integrate the latent feature representation, the feature retrieval structure, and the business semantic label to construct a unified feature representation framework and obtain the multidimensional feature representation.

[0129] The system integrates the latent feature representations, feature retrieval structures, and business semantic labels obtained in the preceding steps to construct a unified feature representation framework. During the integration process, the system employs a feature fusion algorithm to weight and standardize features from different sources, ensuring that various features can be compared and analyzed within a unified metric space.

[0130] Latent feature representation provides the basic features and inherent structure of the data; feature retrieval structure supports efficient similarity queries and pattern matching; and business semantic tags provide semantic understanding and business relevance of the data. These three elements complement each other, together forming a comprehensive and accurate multidimensional feature representation.

[0131] The resulting multidimensional feature representation is a highly structured data model capable of describing the characteristics and relationships of backup data from multiple dimensions. This multidimensional feature representation provides a solid data foundation for subsequent intelligent verification and decision-making, enabling the system to accurately understand the business value, risk level, and associated impact of the data, thereby making more precise disaster recovery verification decisions.

[0132] In another embodiment, a multidimensional feature representation including metadata profiles and business semantic tags is constructed, including:

[0133] Step A1.1: Based on the basic attributes and statistical characteristics of the backup data, generate a metadata profile containing data attributes, statistical characteristics, and quality indicators through data quality analysis and importance assessment.

[0134] The system first performs a comprehensive attribute extraction on the backup data, including basic attributes such as data type, size, creation time, and modification frequency. Then, it conducts multi-dimensional statistical analysis, calculating statistical characteristics such as data distribution, central tendency, and dispersion.

[0135] Based on the extracted basic attributes and statistical features, the system performs data quality analysis, evaluating quality dimensions such as data completeness, accuracy, consistency, and timeliness. By setting quality thresholds and rule sets, it identifies and marks quality problems and abnormal patterns in the data.

[0136] Simultaneously, the system assesses the importance of data, determining its criticality and scope of impact on business operations based on business rules and historical usage patterns. The assessment considers multiple factors, including the data's business value, usage frequency, and dependencies.

[0137] Ultimately, the system integrates basic attributes, statistical features, quality indicators, and importance scores to generate a structured metadata profile that comprehensively describes the characteristics and status of the data.

[0138] Step A1.2: For the business meaning of the backup data, use natural language processing technology and domain knowledge base to perform semantic understanding and classification annotation, and construct business semantic tags that reflect the business meaning and importance of the data.

[0139] The system performs natural language processing on the text information in the backup data, including text segmentation, keyword extraction, topic modeling and other techniques, to extract semantic information from unstructured or semi-structured data.

[0140] Based on a domain knowledge base and a business terminology table, the system performs semantic understanding and mapping, matching extracted keywords and topics with standard business concepts to identify the business meaning and domain of the data.

[0141] The system then performs multi-level classification and labeling, marking the data into different business categories, functional modules, and application scenarios based on the business classification system, forming a hierarchical classification structure.

[0142] At the same time, the system assesses the business importance of the data, and based on business rules and criticality analysis, determines the degree of impact and criticality of the data on specific business processes, and sets corresponding importance levels.

[0143] Ultimately, the system integrates semantic understanding results, classification labels, and importance ratings to construct a business semantic labeling system that reflects the meaning and value of data services.

[0144] Step A1.3: Based on the dependencies and references between the backup data, use graph theory algorithms to analyze the data association patterns and calculate the scope of influence, and establish an association graph describing the data dependencies and scope of influence.

[0145] The system analyzes the dependencies and references between backup data, identifies upstream and downstream relationships, calling relationships and inheritance relationships between data entities, and constructs an initial data association network.

[0146] Based on the constructed network of connections, the system applies graph theory algorithms for in-depth analysis, including connectivity analysis, centrality calculation, and community detection, to identify the core patterns and key nodes of data association.

[0147] The system then calculates the scope of impact, and through propagation models and cascading analysis, assesses the potential scope of impact and propagation path of a single data change or failure on the entire data ecosystem.

[0148] At the same time, the system performs critical path analysis to identify critical paths of data dependencies in business processes, assess the importance and vulnerability of each node on the path, and provide a basis for risk assessment.

[0149] Ultimately, the system integrates the correlation analysis results to establish a multi-level correlation graph, which intuitively displays the dependency structure, scope of influence, and critical path between data, providing a topological perspective for disaster recovery decision-making.

[0150] Furthermore, in step A1.3, the system uses graph theory algorithms to analyze data association patterns and calculate the scope of influence based on the dependencies and references between backup data, establishing an association graph describing data dependencies and the scope of influence. This step is crucial for understanding the internal structure of the data ecosystem and the risk propagation path; its implementation process is detailed below.

[0151] The system first identifies and extracts various relationships between backup data through multiple methods. For structured databases, the system analyzes foreign key relationships, view dependencies, and trigger call chains between tables; for file systems, it analyzes file reference and inclusion relationships; and for application systems, it analyzes data flow and API call relationships defined in configuration files. For example, in a core financial trading system, the system identifies a typical data dependency chain: "Customer Master Data → Account Information → Transaction History → Balance Summary → Report Data." These initially identified relationships form the skeleton of the initial relational network.

[0152] To further enrich the network of relationships, the system also employs data flow analysis technology to discover hidden indirect dependencies by tracing the flow of data within the system. For example, by analyzing log data, the system found that although there is no direct reference relationship between the risk control rule table and the transaction authorization table, changes to risk control rules can indirectly affect the transaction processing flow by influencing authorization decisions, constituting an important implicit dependency. These implicit dependencies are added to the initial network, forming a more complete data relationship network.

[0153] After initial network construction, the system applies graph theory algorithms for in-depth analysis. First, the system calculates network connectivity metrics, including global and local connectivity. Global connectivity reflects the tightness of the entire data ecosystem; high connectivity indicates strong interdependence among data points, potentially leading to rapid fault propagation. Local connectivity reflects the degree of data aggregation within a specific region, helping to identify relatively independent data modules. In a real-world case, system analysis revealed that the customer data module had an internal connectivity of 0.78, while its connectivity with external systems was only 0.23, indicating that this module was relatively independent and suitable as an independent recovery unit.

[0154] Next, the system calculates the centrality metrics for each node, including degree centrality, betweenness centrality, and eigenvector centrality. Degree centrality represents the number of directly connected neighbors of a node, reflecting the direct influence of the data; betweenness centrality represents the frequency with which a node lies on the shortest path between other node pairs, reflecting the "bridging" role of data in information flow; eigenvector centrality considers the importance of neighboring nodes, reflecting the overall influence of the data in the entire network. Through these metrics, the system can identify key data nodes. For example, when analyzing the data relationships of an e-commerce platform, the system found that the eigenvector centrality of the order master table was as high as 0.92, far higher than other tables, indicating that it is a core node in the entire data system, and its integrity is crucial to the entire business process.

[0155] The system also employs community discovery algorithms, such as the Louvain method or the InfoMap algorithm, to divide the data association network into multiple relatively independent modules or "communities." Data nodes within each community are closely interconnected, while connections to other communities are relatively limited. This community structure helps in understanding the natural grouping and functional modules of the data. In the analysis of a certain financial system, the system automatically identified six main data communities, corresponding to functional modules such as customer management, account management, transaction processing, risk control, report analysis, and system management. Data changes within each community primarily affect other data within that community, with relatively limited cross-community impact.

[0156] To assess the potential impact of faults or changes, the system implements impact range calculation based on a network propagation model. The system employs a cascading failure model to simulate how faults propagate within the network. Specifically, when a data node fails, the system calculates the probability that all nodes directly connected to that node will be affected, and then recursively calculates the impact of these affected nodes on their neighbors until the propagation effect decreases below a preset threshold. In this way, the system can generate a "heatmap" showing the probability and extent to which different nodes are affected. For example, for the main configuration table of the core transaction database, the system calculates that its failure will directly affect 23 related tables and indirectly affect up to 142 downstream data objects, covering the entire transaction processing flow and 80% of the report generation process.

[0157] The system also performs critical path analysis to identify bottlenecks and single points of failure in the data flow process. Using the Critical Path Method (CPM) or Project Evaluation and Review Technique (PERT), the system analyzes the time and resource dependencies of each stage in the data processing flow. By calculating the criticality and resilience of the path, the system can identify critical data links without redundancy or alternative paths. In the analysis of a supply chain system, the system found that the data path from raw material inventory to production planning and then to purchase orders is a critical path. Failure at any node on this path will directly lead to the interruption of the entire production planning system; therefore, a dedicated backup and recovery strategy needs to be designed for each data node on this path.

[0158] The system's impact scope analysis also considers the timeliness and business importance of the data. The system integrates business criticality scoring into the propagation model, giving higher weight to data nodes that are more critical to the business in the impact scope calculation. Simultaneously, the system considers the timeliness requirements of the data, distinguishing between real-time, near-real-time, and historical data, and employing different impact assessment strategies for different data types. For example, for real-time transaction data, the system sets a lower tolerance threshold and a higher propagation rate; while for historical archived data, more lenient assessment parameters are used.

[0159] After completing all analyses, the system integrates the results into a multi-layered relationship graph. The core layer of the graph displays direct dependencies and main flow paths between data points; the extension layer shows indirect dependencies and potential impact paths; the aggregation layer displays data communities and functional module divisions; and the risk layer displays the distribution of key nodes, bottleneck paths, and vulnerabilities. This multi-layered visualization makes complex data relationships intuitive and understandable, facilitating disaster recovery decisions and risk assessments.

[0160] Taking an e-commerce platform as an example, the system's constructed relationship graph clearly demonstrates the complete dependency chain from user data to order data, then to inventory data and logistics data. The graph identifies the configuration data of the order processing engine as a critical node, whose impact covers three core business processes: order creation, payment processing, and inventory updates. The system calculated that a failure at this node would affect 30% of online transactions within 5 minutes and 95% of order processing capacity within 30 minutes. Based on this analysis, the operations team designed a real-time mirror backup and automatic failover mechanism for this configuration data, reducing potential business downtime from hours to minutes.

[0161] Through this in-depth graph theory analysis and correlation graph construction, the system can not only understand the static dependencies between data, but also predict the dynamic failure propagation paths and impact ranges, providing a scientific basis for disaster recovery strategy design and risk management. This disaster recovery approach based on data correlation transcends the traditional single-point backup and recovery model, achieving a global understanding and precise protection of the entire data ecosystem.

[0162] Step S1.4: The metadata profile, the business semantic tags, and the association graph are uniformly encoded and standardized using a feature fusion algorithm to form the multidimensional feature representation.

[0163] The system employs a feature fusion algorithm to uniformly process the metadata profiles, business semantic tags, and association graphs generated in the preceding steps. First, features from different sources are standardized to ensure that all types of features are compared at the same scale and unit.

[0164] Then, the system employs a weighted combination strategy, assigning different weights based on the importance and reliability of the features to achieve optimized fusion of multi-source features. Feature weights can be set through expert experience or automatically optimized using machine learning methods.

[0165] The system also uses dimensionality reduction techniques, such as principal component analysis (PCA) or t-SNE, to process high-dimensional features, thereby reducing the dimensionality of the feature space while retaining key information and improving the efficiency of subsequent processing.

[0166] At the same time, the system performs feature selection, filtering out the most representative and discriminative feature subsets based on the information gain or relevance of the features, thereby reducing redundancy and noise.

[0167] Ultimately, the system integrates the processed features into a unified multidimensional feature representation. This representation retains the key information of the original features while possessing structured, standardized, and computable characteristics, providing an ideal data foundation for subsequent intelligent analysis and decision-making.

[0168] In another embodiment, a reinforcement learning strategy network with an Actor-Critic dual-network structure is used for fault mode learning to dynamically generate scenario-based fault injection instructions that conform to the actual business distribution, including:

[0169] Step S2.1: Based on the fault types, occurrence frequency and impact range in the historical fault data, a fault knowledge base containing the correlation between fault occurrence patterns and business characteristics is constructed through data mining and pattern recognition to obtain prior knowledge of fault patterns.

[0170] The system performs comprehensive analysis and mining of historical fault data to identify the characteristic patterns and occurrence regularities of various faults. First, the system classifies and statistically analyzes the fault data, including statistical analysis of dimensions such as fault type distribution, occurrence frequency, duration, and impact range.

[0171] Then, the system employs time series analysis to study the temporal patterns of fault occurrence, including periodic patterns, seasonal variations, and long-term trends. Through techniques such as autoregressive models and Fourier transforms, the system is able to capture the temporal correlation characteristics of fault occurrence.

[0172] The system also analyzes the correlation between faults and business characteristics, identifying which business characteristics are highly correlated with specific types of faults through correlation analysis and association rule mining. These association rules can be expressed in the form of "if a business characteristic meets condition X, then the probability of fault type Y occurring increases".

[0173] Simultaneously, the system performs fault link and propagation mode analysis to study how faults spread within the system and affect other components. By constructing fault propagation graphs and causal chains, the system can understand the cascading effects and impact paths of faults.

[0174] Ultimately, the system integrates the analysis results into a structured fault knowledge base, containing multi-dimensional information such as fault type characteristics, occurrence patterns, business relationships, and propagation modes. This knowledge base serves as prior knowledge for reinforcement learning, guiding the policy network to generate more realistic and effective fault scenarios.

[0175] Step S2.2: Utilize the prior knowledge of the fault modes and the current system state to make strategy decisions through the Actor network, including fault type selection, scope of influence determination and injection timing control, and generate an initial fault injection strategy.

[0176] The system takes prior knowledge of failure modes and the current system state as input and makes policy decisions through an Actor network. An Actor network is a deep neural network, typically composed of multiple fully connected networks or recurrent neural networks, which can map input states to specific decision actions.

[0177] In the fault type selection phase, the Actor network generates probability distributions for different fault types based on the current system state and historical fault distributions, and selects specific fault types based on sampling from these distributions. The network considers factors such as the actual occurrence frequency of fault types, their correlation with the current business state, and verification coverage requirements.

[0178] In the impact scope determination phase, the Actor network assesses the verification value and business risks of different impact scopes to determine the system components and data range that should be affected by the failure. The network weighs verification coverage against business impact, selecting the impact scope that maximizes verification effectiveness while controlling business risks.

[0179] In the injection timing control phase, the Actor network determines the optimal time to perform fault injection based on system load, business importance, and time window constraints. The network considers factors such as avoiding peak business periods, the completion status of critical transactions, and verification time windows to ensure that fault injection is executed at the appropriate time.

[0180] Through these three key decisions, the Actor network generates a complete initial fault injection strategy, including detailed information such as the specific fault type, scope of impact, and execution timing.

[0181] Step S2.3: For the initial fault injection strategy, use Critic network analysis to verify the effect, business impact and recovery performance after the strategy is executed, and output the strategy value assessment result.

[0182] The system evaluates the effectiveness and impact of the initial fault injection strategy through simulation or small-scale trials. The Critic network is the core component of value assessment, typically composed of multi-layer neural networks, capable of mapping policy state-action pairs to value scores.

[0183] In the verification effectiveness evaluation phase, the Critic network analysis strategy contributes to verification coverage, fault detection rate, and anomaly discovery rate. The network evaluates whether the strategy can effectively verify the various key components and functions of the system, and whether it can discover potential disaster recovery issues.

[0184] In the business impact assessment phase, Critic network analysis examines the degree of disruption the strategy causes to normal business operations, including performance impact, service interruption time, and changes in user experience. The network team then evaluates whether the strategy is implemented within acceptable business impact limits.

[0185] In the recovery performance evaluation phase, the Critic network analysis system assesses its recovery capabilities and efficiency in the face of injection failures, including recovery time, data integrity, and service continuity. The network evaluates whether the strategy can effectively test the system's recovery mechanism.

[0186] The Critic network takes these factors into account and outputs a comprehensive strategy value score, which reflects the strategy's overall performance across three dimensions: validation effectiveness, business impact, and recovery performance.

[0187] Step S2.4: Based on the strategy value evaluation results, update the parameters of the Actor network and Critic network through the strategy gradient algorithm to optimize the fault injection strategy generation capability.

[0188] The system updates and optimizes the parameters of the Actor and Critic networks based on the policy value evaluation results output by the Critic network using the policy gradient algorithm. Policy gradient is a commonly used optimization method in reinforcement learning, which adjusts network parameters according to the policy execution results, gradually optimizing the policy towards higher value.

[0189] During the Actor network update phase, the system uses the policy gradient formula to calculate the parameter update direction. Specifically, based on the Critic network's value assessment, the system calculates the gradient direction of each decision action and then adjusts the parameters along directions that improve policy value. This enables the Actor network to generate more valuable fault injection policies.

[0190] In the Critic network update phase, the system uses temporal difference learning or Monte Carlo methods to optimize its value assessment capabilities. The system compares the gap between the Critic's value predictions and the actual observed returns, improving the prediction accuracy of the Critic network by minimizing this gap.

[0191] Meanwhile, the system employs experience replay technology, storing historical policy execution records in a replay buffer and randomly sampling from it for batch learning. This method can break the correlation between samples, improving the stability and efficiency of learning.

[0192] Through continuous learning and optimization, the generation and evaluation capabilities of both networks will be continuously improved, enabling the entire policy network to generate increasingly high-quality fault injection policies that meet actual needs.

[0193] Step S2.5: Based on the optimized network parameters and current verification requirements, the Monte Carlo tree search algorithm is used to explore the optimal fault combination scheme and generate the scenario-based fault injection instruction.

[0194] Based on the optimized network parameters and the specific verification requirements, the system employs the Monte Carlo Tree Search (MCTS) algorithm to explore the optimal fault combination scheme. MCTS is a powerful decision tree search algorithm capable of efficiently finding near-optimal solutions in a large decision space.

[0195] During the search process, the system constructs a decision tree, where each node represents a fault state and edges represent decisions to add or change the fault. The depth of the tree represents the length of the fault sequence, and leaf nodes represent complete fault combination schemes.

[0196] MCTS iteratively optimizes through four key steps: Selection, Expansion, Simulation, and Backpropagation. In the Selection phase, the algorithm uses the UCB (Upper Confidence Bound) formula to balance exploration and exploitation, selecting the most promising nodes. In the Expansion phase, the algorithm adds new child nodes to the selected nodes, expanding the search space. In the Simulation phase, the algorithm uses fast heuristics or Actor networks to simulate the execution of decision sequences until the termination state, evaluating the value of each solution. In the Backpropagation phase, the algorithm propagates the evaluation results back along the path, updating node statistics.

[0197] Through numerous iterations, MCTS can gradually converge to high-quality fault combination solutions. The system selects the most valuable path in the search tree to generate the final scenario-based fault injection instructions. These instructions not only contain detailed information about individual faults but also the combination relationships and execution order between faults, forming a complete fault scenario.

[0198] In another embodiment, a directed acyclic graph of fault events is constructed using time-optimal directed q-analysis techniques, and the timing arrangement is optimized. Contextualized fault injection is then performed, and verification data is obtained, including:

[0199] Step S3.1: Based on the fault events in the scenario-based fault injection instruction, analyze the causal relationships and temporal constraints between the events, and construct a directed acyclic graph with fault events as nodes and dependencies as edges.

[0200] The system performs in-depth analysis of each fault event in the scenario-based fault injection instructions, identifying causal relationships, dependencies, and timing constraints between events. The analysis considers various relationship types, including necessary preconditions, triggering relationships, mutual exclusion relationships, and parallel possibilities.

[0201] Based on the analysis results, the system constructs a directed acyclic graph (DAG) model of the fault events. In this graph, each node represents a specific fault event, and edges represent dependencies or temporal constraints between events. The weight of the edges can represent the strength of the dependency or the time interval requirement.

[0202] The system ensures that the constructed graph is acyclic, which guarantees the executability of fault injection. If circular dependencies are detected, the system analyzes the conflict relationships and makes adjustments to ensure the rationality of the final graph structure.

[0203] After the graphical model is constructed, the system performs topology sorting analysis to identify the execution order constraints of events and the critical path. Critical path analysis helps the system understand which event sequences have the greatest impact on the overall execution time.

[0204] This graph-based modeling method enables the system to intuitively and accurately represent the organizational structure and dependencies of events in complex fault scenarios, laying the foundation for subsequent timing optimization.

[0205] Step S3.2: For the directed acyclic graph, the q-learning algorithm is used to learn the verification effect of fault combinations under different time windows, and the optimal timing arrangement that maximizes verification coverage and minimizes business impact is calculated.

[0206] The system treats the directed acyclic graph of fault events as a Markov decision process (MDP) and employs a q-learning algorithm to learn the optimal event scheduling strategy. In this MDP, the state represents the set of executed fault events, the action represents selecting the next event to execute, and the reward function combines metrics of validation coverage and business impact.

[0207] Q-learning is a model-free reinforcement learning algorithm that learns the optimal policy by iteratively updating a q-value table. The q-value table Q(s,a) represents the long-term expected reward of performing action a in state s. The algorithm iteratively updates the q-values ​​using the Bellman equation: Q(s,a) ← Q(s,a) + α[r + γ·max(Q(s',a')) - Q(s,a)], where α is the learning rate, γ is the discount factor, r is the immediate reward, and s' is the new state after performing action a.

[0208] The system is designed with a comprehensive reward function to balance verification effectiveness and business impact. The reward function consists of two parts: positive rewards (increased verification coverage, increased fault detection rate) and negative penalties (degraded business performance, prolonged service downtime). The system optimizes strategies for different verification scenarios by adjusting the weights of these two parts.

[0209] To address the complex state space, the system employs function approximation methods (such as deep q-networks) to replace traditional tabular q-learning, improving learning efficiency and generalization ability. The system uses techniques such as experience replay and target networks to stabilize the training process.

[0210] Through extensive iterative learning, the q-learning algorithm gradually converges to the optimal timing arrangement that balances verification coverage and business impact, providing a scientific execution plan for fault injection.

[0211] Step S3.2 is a key technical step in this invention, which utilizes the q-learning algorithm in reinforcement learning to intelligently optimize the timing of fault injection. The implementation process, core principles, and specific applications of this step will be explained in detail below.

[0212] The system first transforms the directed acyclic graph of fault events constructed in step S3.1 into a Markov decision process (MDP), which is a prerequisite for applying the q-learning algorithm. In this MDP, the state space is defined by the set of executed fault events, with each state representing a combination of faults that has been injected; the action space is the next fault event that can be executed, constrained by the dependencies in the directed graph; the state transition function describes the change in the system state after executing a fault event; and the reward function comprehensively considers the verification coverage gain and business impact, expressing the immediate benefit of executing a certain action in a numerical way.

[0213] The core of the q-learning algorithm is to learn an action-value function Q(s,a), which represents the long-term cumulative reward obtained by following the optimal policy after performing action a in state s. The system approximates this value function through iterative updates: Q(s,a)←Q(s,a)+α[r+γ·max(Q(s',a'))-Q(s,a)]. Here, α is the learning rate, controlling the update step size; γ is the discount factor, balancing immediate reward and future reward; r is the immediate reward obtained by performing action a; s' is the new state after performing action a; and max(Q(s',a')) is the maximum value of all possible actions in the new state. Through repeated iterations, the Q-function gradually converges to the optimal action value.

[0214] In practical implementation, the system designs a composite reward function to precisely quantify the value of each fault injection decision. This reward function includes multiple dimensions: verification coverage gain reflects the new information brought to the overall verification by the newly injected fault, such as covering previously untested components or functions; business impact measures the degree of interference of the fault on normal business operations, including negative impacts such as performance degradation and service interruption time; resource utilization efficiency considers the consumption of computing, storage, and network resources during the verification process; and time window utilization evaluates whether the distribution of verification tasks within the allocated time window is reasonable. These dimensions are combined by weight to form the final reward value, and the weights can be dynamically adjusted according to specific verification objectives and business priorities.

[0215] Considering the complexity of fault verification scenarios, the system employs a function approximation method to handle large-scale state spaces. Specifically, the system uses a Deep Q-Network (DQN) instead of traditional tabular q-learning, using a neural network to approximate the Q-function. DQN contains multiple hidden layers, enabling it to automatically extract state features and generate action value estimates. To stabilize the training process, the system employs an experience replay technique, storing historical state-action-reward-next state samples in a replay buffer. During training, samples are randomly selected in batches for updates, breaking the correlation between samples. Simultaneously, the system also uses a target network technique, maintaining a separate network to generate target values ​​and periodically synchronizing parameters from the main network to reduce instability during training.

[0216] During the learning process, the system needs to balance exploration and exploitation. To this end, the system employs an ε-greedy strategy: selecting the action with the highest estimated value (exploitation) with a probability of 1-ε, and randomly selecting a legal action (exploration) with a probability of ε. The ε value gradually decreases as the learning process progresses, initially prioritizing exploration and later prioritizing exploitation. In certain key validation scenarios, the system also incorporates a Boltzmann exploration strategy, sampling according to the probability distribution of action values, ensuring a higher probability of selecting high-value actions while still maintaining a degree of exploration.

[0217] The learning process also requires consideration of the specific constraints of different time windows. For example, high-impact fault injection can be performed during off-peak business periods, while excessive interference should be avoided during peak business periods. The system addresses this issue through a time-related reward adjustment function: during predefined critical business periods, the penalty weight for business impact is increased; during maintenance windows or low-load periods, the penalty weight is reduced accordingly. This dynamic adjustment ensures that verification tasks can intelligently adapt to business time patterns.

[0218] To illustrate with a practical example, a system needs to schedule seven interdependent fault injection events within an 8-hour verification window. Traditional methods might simply schedule these events at equal intervals or randomly, often with unsatisfactory results. However, after approximately 5000 rounds of simulation training, the q-learning algorithm learned an optimized timing arrangement: high-resource-consuming fault injections (such as full data consistency checks) are scheduled during periods of low resource utilization; faults that may trigger automatic recovery (such as network partitions) are scheduled within windows with sufficient observation time; and highly correlated faults (such as primary / standby failover and data synchronization interruptions) are scheduled within appropriate time intervals, allowing for observation of their interaction effects without confusing phenomena with their causes due to their proximity. This arrangement improves verification coverage by approximately 40% while reducing the impact on normal business operations by approximately 60%, a significant improvement over manually planned solutions.

[0219] Another advantage of the q-learning algorithm is its ability to adapt to uncertainty. In practical verification, the effect and duration of fault injection are often uncertain. Traditional deterministic planning struggles to handle such variations, while q-learning-based methods are naturally suited to stochastic environments, dynamically adjusting subsequent decisions based on observed system responses. For example, if a fault injection causes a system recovery time that exceeds expectations, the algorithm automatically adjusts the timing of subsequent fault injections, ensuring the consistency and effectiveness of the verification process.

[0220] As validation experience accumulates, the system's q-learning model continuously improves. After each validation, the system adds the observed state transitions and reward data to the training dataset, updating the model through a combination of supervised learning and reinforcement learning. This continuous learning mechanism enables the system to adapt to constantly changing IT infrastructure and business needs, thereby improving the effectiveness and efficiency of the validation strategy.

[0221] It's worth noting that the system also implements scenario-based q-learning extensions. For different types of verification objectives (such as performance disaster recovery verification, data consistency verification, and end-to-end fault recovery verification), the system trains specialized q-functions to capture the optimal decision-making patterns under different scenarios. Before verification execution, the system selects the appropriate model based on the current verification objective, or combines the outputs of multiple models to further improve the accuracy of the decision.

[0222] By optimizing the timing arrangement through the q-learning algorithm, the system can maximize the acquisition of effective verification information within a limited verification time window, while minimizing interference with normal business operations. This AI-based verification scheduling method not only improves the comprehensiveness and accuracy of verification but also significantly reduces resource consumption and business risks during the verification process, bringing a qualitative leap to disaster recovery verification.

[0223] Step S3.3: According to the optimal timing arrangement, perform the fault injection operation in chronological order, monitor the system response in real time, and collect verification data including system state changes, recovery process and performance indicators.

[0224] The system constructs a detailed fault injection execution plan based on optimal timing. The plan includes detailed information such as the specific execution time, duration, impact scope, and execution method for each fault event. The system ensures the feasibility and security of the plan, including setting up emergency rollback mechanisms and safety boundary conditions.

[0225] According to the execution plan, the system triggers fault injection operations sequentially through automated tools or API interfaces. The injection methods include various types such as resource limitations (CPU / memory usage, network bandwidth limitations), service interruptions (process termination, service restart), data problems (data corruption, delay, loss), and environmental anomalies (network partitions, clock skew).

[0226] During the fault injection process, the system deploys a comprehensive monitoring system to capture system responses and changes in real time. Monitoring metrics include infrastructure layer metrics (CPU, memory, network, disk, etc.), middleware layer metrics (database connections, message queue length, etc.), application layer metrics (response time, error rate, throughput, etc.), and business layer metrics (transaction success rate, user experience, etc.).

[0227] Meanwhile, the system records the disaster recovery process in detail, including key time points such as fault detection time, alarm trigger time, recovery operation start time, and service recovery time, as well as various operations and status changes during the recovery process.

[0228] The system integrates all collected monitoring data, status changes, and recovery process information into a structured verification dataset, providing a comprehensive and detailed data foundation for subsequent intelligent analysis.

[0229] In another embodiment, a pre-trained Transformer model is used for feature encoding, and transfer learning techniques and causal inference models are employed for intelligent analysis. The resulting interpretable verification report, including assessments of data integrity, consistency, and recoverability, is output.

[0230] Step S4.1: Based on the verification requirements, data features, and fault scenario information in the verification data, use the self-attention mechanism of the pre-trained Transformer model to encode multimodal information and generate a unified vector representation.

[0231] The system first preprocesses and standardizes the validation data, converting data from different sources and formats into a standard format that the model can process. The processed data includes multimodal data such as structured monitoring metrics, semi-structured log data, and unstructured error messages.

[0232] The system then encodes these multimodal data using a pre-trained Transformer model. The Transformer model is known for its powerful self-attention mechanism, which effectively captures long-range dependencies in the data. The self-attention mechanism generates a representation containing global contextual information by calculating the correlation between each element in the sequence and all other elements.

[0233] For different types of data, the system employs specific encoding strategies: structured data is converted into vectors through a feature embedding layer; text data is represented through word embeddings and positional encoding; and temporal data captures temporal characteristics by adding temporal encoding. These initial encodings are then fed into the Transformer's multi-head self-attention layer and feedforward network layer to generate deep semantic representations.

[0234] Multimodal fusion is a key challenge. The system integrates information from different modalities into a unified vector representation through cross-modal attention mechanisms or specific fusion layers. This fusion takes into account the correlation and complementarity between different modalities, ensuring that important information is not lost during the fusion process.

[0235] The resulting unified vector representation is a high-dimensional vector that encodes comprehensive information about verification requirements, data features, and fault scenarios, providing a rich semantic foundation for subsequent intelligent analysis.

[0236] Step S4.2: Using the unified vector representation, feature extraction is performed on the base model pre-trained on a large-scale disaster recovery scenario dataset, and fine-tuning techniques are used to adjust the parameters for the current specific verification scenario to obtain an adaptively enhanced feature representation.

[0237] The system utilizes a pre-trained base model on a large-scale disaster recovery scenario dataset to further extract features from the unified vector representation generated in step S4.1. The pre-trained model has been trained on a large dataset containing various disaster recovery scenarios, fault types, and system responses, and has learned general knowledge and patterns in the disaster recovery field.

[0238] By inputting unified vector representations into the pre-trained model, the system can extract more abstract and semantically rich features. These features reflect the correlation between the current verification scenario and known disaster recovery modes, helping the system identify typical failure modes and recovery behaviors.

[0239] For the specific validation scenario, the system employs fine-tuning techniques to locally adjust the parameters of the pre-trained model. During fine-tuning, the system uses a small amount of highly relevant validation data collected in the current environment to train the model in a targeted manner. Different strategies can be used for fine-tuning, such as fixing the feature extractor and only fine-tuning the top layer, fine-tuning all parameters with a low learning rate, or progressively decreasing the learning rate layer by layer.

[0240] The purpose of fine-tuning is to better adapt the model to the specific validation environment and business characteristics, thereby improving the accuracy and relevance of the analysis. Through fine-tuning, the model can focus on failure modes and system behaviors unique to the current environment, overcoming the distributional differences between the pre-training data and the current environment.

[0241] After fine-tuning, the system obtained adaptively enhanced feature representations that incorporated both the general knowledge learned during the pre-training phase and the specific characteristics of the current validation environment, providing ideal input for subsequent causal analysis.

[0242] Step S4.3: Based on the adaptive enhancement feature representation, a structured causal model is used to analyze the causal relationship between data integrity, consistency and recoverability indicators, and to identify the root causes and critical paths affecting disaster recovery effectiveness.

[0243] Based on adaptively enhanced feature representation, the system constructs a structured causal model (SCM) for disaster recovery verification indicators. This model uses a directed acyclic graph to represent the causal relationships between variables, with nodes representing key indicators (such as data integrity, consistency, and recoverability) and edges representing the direction and intensity of causal influence.

[0244] The model construction combines domain knowledge and data-driven methods: first, an initial causal structure is set based on disaster recovery theory and expert experience; then, causal discovery algorithms (such as PC algorithm and FCI algorithm) are used to learn and verify causal relationships from the data; finally, the strength of causal relationships is quantified through structural equation modeling (SEM).

[0245] Based on the constructed causal model, the system conducts comprehensive causal analysis: it assesses the impact of changes in specific factors on the outcome through intervention analysis (do-calculus); it explores hypothetical scenarios of "what if x did not happen" through counterfactual analysis; and it traces the propagation chain of causal effects through path analysis.

[0246] The system focuses on identifying the root causes and critical paths that affect disaster recovery effectiveness. Root causes are those factors located at the beginning of the causal chain and having a significant impact on the final outcome; critical paths are the causal propagation paths from root causes to the final outcome with the greatest impact. The system quantitatively assesses the importance of different factors by calculating indicators such as path effect strength and node centrality.

[0247] The causal analysis results not only explain the reasons for the formation of the current verification results, but also predict the possible changes in the verification indicators under different intervention measures, providing a scientific basis for the optimization of disaster recovery strategies.

[0248] Step S4.4: Based on the critical path and the current verification scenario, use graph reordering approximate nearest neighbor search technology to match similar historical verification cases and obtain verification decision support information.

[0249] The system intelligently matches the feature vector of the current verification scenario with the historical verification case library to find historical cases with similar characteristics. The historical case library stores a large number of past verification records, each containing information such as verification scenario characteristics, execution results, solutions, and lessons learned.

[0250] The matching process employs a graph reordering approximate nearest neighbor search technique, which is an efficient similarity search method, particularly suitable for high-dimensional feature spaces. The implementation consists of two stages: first, a candidate case set is quickly retrieved using a basic index structure (such as LSH or a tree index); then, the candidate set is precisely ordered using a graph reordering algorithm to improve the accuracy of similarity calculation.

[0251] The graph re-ranking algorithm constructs a similarity graph where nodes are candidate cases and edges connect highly similar case pairs. By performing random walks or diffusion processes on the graph, the algorithm can discover deep-seated similarity relationships, overcoming the limitations of traditional Euclidean distance or cosine similarity. The algorithm considers the correlation and weight differences between features, enabling it to identify semantically similar cases that may have different feature representations.

[0252] The system extracts validation decision support information from matched historical cases, including validation results, solutions adopted, success stories, and precautions. This information is sorted and filtered according to relevance and applicability to ensure that the provided reference information has practical guiding value.

[0253] Drawing on historical experience can not only help in understanding current validation results, but also provide validated solutions to the problems discovered, greatly improving the efficiency and accuracy of decision-making.

[0254] Step S4.5: Integrate the root cause, the critical path, and the verification decision support information, and generate an interpretable verification report containing verification conclusions, key indicators, reasoning processes, and key evidence based on an attention mechanism, thus obtaining the interpretable verification report.

[0255] The system integrates the root cause analysis, critical path identification, and historical case references obtained in the preceding steps to construct a knowledge graph for the verification report. This graph organizes various types of information according to logical relationships and importance, forming the report's skeletal structure.

[0256] Based on the constructed knowledge graph, the system employs an attention mechanism to generate interpretable verification reports. This attention mechanism assigns different weights based on the importance and relevance of the content, ensuring that the report highlights key information while maintaining content integrity. Specifically, the system uses a hierarchical attention network to calculate attention scores at the word, sentence, and paragraph levels, thereby determining the display order and level of detail for the content.

[0257] The report is comprehensive and structured, including the following core sections: the verification conclusions section summarizes the overall disaster recovery capability assessment results and key findings; the key indicators section presents the quantitative results of key indicators such as data integrity, consistency, and recoverability; the reasoning process section details the logical chain and causal relationship from evidence to conclusions; and the key evidence section provides specific data and observations supporting the conclusions.

[0258] The system also generates visualizations, including indicator dashboards, cause-and-effect diagrams, and time-series change graphs, to intuitively display verification results and the analysis process. The visualizations are adaptively designed, dynamically adjusting the content and level of detail based on user focus.

[0259] The generated verification report not only tells users "what the results are", but also explains "why the results are like this" and "how to improve", which improves the understandability and operability of the verification results and provides clear guidance for disaster recovery strategy optimization.

[0260] In another embodiment, such as Figure 3 As shown, a federated learning framework is used to aggregate and update model parameters across data center nodes, enabling cross-data center policy synchronization and performance optimization, including:

[0261] Step S5.1: Deploy a verification result collector and performance monitoring component on each data center node to collect performance indicators such as verification accuracy, execution efficiency, resource consumption, and fault detection rate in real time, and obtain local training data.

[0262] The system deploys a verification result collector and performance monitoring components on each data center node in a distributed architecture. These components collect various performance metrics and verification results from the nodes in real time through API integration, log analysis, and monitoring probes.

[0263] The core metrics collected include four dimensions: verification accuracy metrics (such as data consistency verification accuracy, recovery point verification accuracy, etc.), reflecting the correctness of the verification results; execution efficiency metrics (such as verification task completion time, resource utilization, etc.), reflecting the efficiency of the verification process; resource consumption metrics (such as CPU utilization, memory usage, I / O load, etc.), reflecting the impact of verification on system resources; and fault detection rate metrics (such as fault identification rate, false alarm rate, false negative rate, etc.), reflecting the fault detection capability of verification.

[0264] The system also collects fine-grained data during the verification process, including the detection accuracy for each fault type, the recovery success rate for different data types, and the execution time distribution for various verification tasks. This fine-grained data helps to more accurately identify the model's strengths and weaknesses.

[0265] All collected data underwent preprocessing and standardization, transforming it into a format suitable for machine learning. Preprocessing included steps such as outlier detection, missing value handling, and feature normalization to ensure data quality and consistency.

[0266] The processed data becomes the local training dataset, used for subsequent local model training and optimization. The system ensures the timeliness and representativeness of the data by periodically updating the training data through a sliding window mechanism, ensuring that the model can adapt to dynamic changes in the environment.

[0267] Step S5.2: Based on the local training data, each node independently performs incremental training of the validation model and calculates the gradient update of the model parameters to generate local gradient information.

[0268] Each data center node independently performs incremental training of the validation model based on locally collected training data. Incremental training is an efficient model update method that updates the existing model using only newly collected data, without retraining the entire model. This method is particularly suitable for distributed environments, significantly reducing computational overhead and training time.

[0269] During training, the system uses Mini-batch Stochastic Gradient Descent (SGD) or its variants (such as Adam and RMSProp) as the optimization algorithm. Each node starts with the current global model parameters and then performs several rounds of training iterations using local data. The training objective is to minimize the prediction error (such as mean squared error and cross-entropy), and the gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm.

[0270] To prevent overfitting and improve generalization ability, the system employs various regularization techniques during training, such as L1 / L2 regularization, Dropout, and batch normalization. The system also uses an early stopping strategy, stopping training when performance on the validation set no longer improves, thus avoiding overfitting to local data.

[0271] After training, the system calculates the gradient update information of the model parameters. Specifically, the system calculates the difference between the trained model parameters and the parameters before training (global model), obtaining a parameter update vector. This vector reflects the direction and magnitude of the local data's optimization of the model.

[0272] Finally, the system generates a local gradient information package, which includes parameter gradients, training metadata (such as the number of samples and training epochs), and performance metrics (such as local validation accuracy). This information will be used in the subsequent secure aggregation process.

[0273] Step S5.3: The local gradient information is processed securely using homomorphic encryption and differential privacy technology, and then uploaded to the central coordinator via a secure aggregation protocol.

[0274] The system performs secure processing on local gradient information to ensure that sensitive data from each node is not leaked during parameter aggregation. This security processing primarily employs two core technologies: homomorphic encryption and differential privacy.

[0275] Homomorphic encryption allows computation to be performed in an encrypted state without first decrypting the data. The system uses an additive homomorphic encryption scheme (such as the Paillier cryptosystem) to encrypt gradient parameters, ensuring that the original gradient values ​​are not leaked even during transmission and aggregation. During encryption, each node encrypts its own gradient information using a public key, while only the central coordinator possesses the private key required for decryption.

[0276] Differential privacy is a technique that protects privacy by adding carefully calibrated noise to the data. The system adds an appropriate amount of random noise (such as Laplace noise or Gaussian noise) to the gradient vector based on gradient sensitivity and a pre-defined privacy budget (ε). After adding noise, even if an attacker obtains the complete gradient information, they cannot accurately infer information about any specific sample.

[0277] Secure aggregation protocols are the core mechanism for ensuring secure parameter transmission and aggregation. The system employs Secure Multi-Party Computation (SMPC) protocols, such as SecAgg, which allow multiple participants to collaboratively compute functions (such as summation) without revealing their individual inputs. The protocol is robust and can handle anomalies such as participant disconnections.

[0278] During communication, the system employs secure transmission protocols such as TLS / SSL to ensure that gradient information is not eavesdropped on or tampered with during network transmission. Simultaneously, the system implements strict identity authentication and authorization mechanisms to ensure that only legitimate nodes can participate in the parameter aggregation process.

[0279] By combining these security measures, the system can achieve knowledge sharing and collaborative optimization across data centers while ensuring privacy and data security.

[0280] Step S5.4: In the central coordinator, the FedAvg algorithm is used to perform weighted aggregation based on the data quality, number of training samples and historical contribution of each node to generate updated global model parameters.

[0281] The central coordinator receives the safely processed gradient information from each node and performs aggregation. The core algorithm for aggregation is the Federated Avg algorithm, which is the most basic and effective aggregation method in federated learning.

[0282] The aggregation process first requires decrypting the encrypted gradient information uploaded by each node. The central coordinator uses its private key to decrypt the encrypted data and obtain the original gradient update values. For gradients processed using differential privacy, the system accepts the noise contained within them; this noise usually cancels out after aggregation across a large number of nodes, having a limited impact on the final result.

[0283] The basic idea of ​​the FedAvg algorithm is to perform a weighted average of the parameter updates of each node. Unlike a simple average, the system assigns different weights based on multiple factors: data quality indicators reflect the representativeness and reliability of the node's data; the number of training samples represents the scale of data contributed by each node; and historical contribution evaluates the effectiveness and value of the node's historical updates. The system uses a weighted formula... Calculate the global parameters, where α i w is the weight of node i. i This involves updating the parameters of node i.

[0284] The system also implements an anomaly detection mechanism to identify and filter potential malicious updates or anomalous parameters. Detection methods include statistical outlier analysis, directional consistency tests, and impact simulation assessments. For detected anomalous updates, the system can reduce their weight or exclude them completely to ensure the reliability of the aggregation results.

[0285] After aggregation, the system performs a global model consistency check to ensure that the model parameters are within a reasonable range and have good generalization performance. If a global model performance degradation is detected, the system can initiate a rollback mechanism to restore the model parameters to the previous stable version.

[0286] Finally, the central coordinator generates updated global model parameters, ready to be distributed to each data center node. These global parameters incorporate the knowledge contributions of all nodes, representing the system's current optimal verification strategy and performance level.

[0287] Step S5.4 is the core technical step in this invention for realizing federated learning model aggregation. It integrates the local model parameters distributed across various data centers into a higher-performing global model using the FedAvg (Federated Averaging) algorithm. The implementation principle of the FedAvg algorithm and its specific application process in this system will be explained in detail below.

[0288] The basic idea of ​​the FedAvg algorithm is to aggregate the parameters of multiple local models through weighted averaging to form a global model that incorporates the knowledge of each node. In this system, the central coordinator first receives encrypted model parameters from each data center node. These parameters are typically represented as multidimensional tensors, containing weights and biases from the neural network. For example, a typical disaster recovery verification model may contain approximately 5 million parameters, distributed across the input layer, multiple hidden layers, and the output layer. Parameter updates are transmitted in the form of gradients or complete model parameters, and their security is protected using homomorphic encryption to ensure that data is not leaked during transmission.

[0289] Before aggregation officially begins, the central coordinator first performs a quality assessment on each node to determine its weight in the aggregation process. This assessment is based on three key dimensions: first, data quality, measured by evaluating each node's data distribution, noise level, and feature completeness; second, the number of training samples, reflecting the amount of information contributed by each node; and third, historical contribution, assessed by recording the performance of each node's model in past aggregations to evaluate its reliability and consistency. The system uses a composite scoring function to integrate these three dimensions into a comprehensive weighting coefficient.

[0290] To illustrate with a practical example, suppose the system has three data center nodes A, B, and C, each training its own model parameter sets θA, θB, and θC based on local data. Evaluation shows that node A has a data quality score of 0.85, a sample size of 10,000, and a historical contribution of 0.92; node B has a data quality score of 0.78, a sample size of 8,000, and a historical contribution of 0.85; and node C has a data quality score of 0.90, a sample size of 6,000, and a historical contribution of 0.88. The system converts these metrics into final weights using a weighting function: wA = 0.42, wB = 0.31, wC = 0.27. These weights will be used in subsequent parameter aggregation.

[0291] The core step of the FedAvg algorithm is parameter weighted aggregation. For each parameter position i in the model, the global model parameter calculation formula is: Where θ k,i w is the parameter value of node k at position i. k This is the weight of node k. In practice, this calculation is usually performed efficiently using matrix or tensor operations. For example, for the three node cases mentioned earlier, the global value of the i-th parameter is calculated as:

[0292]

[0293] However, simple weighted averaging can face some practical challenges. The system enhances the robustness and effectiveness of the FedAvg algorithm through several technical improvements. First is the anomaly detection mechanism, where the system uses statistical methods to detect anomalous parameter updates. If a node's parameter update differs significantly from the majority of nodes, it may indicate a problem with the node's data or that it has been attacked. In such cases, the system dynamically reduces its weight or completely excludes its contribution. For example, if some parameter updates of node C deviate from the mean by more than three standard deviations, the system may reduce its weight at those parameter positions by 50%.

[0294] The second improvement is a hierarchical aggregation strategy. For large-scale distributed systems, which may have dozens or hundreds of data centers, directly aggregating parameters from all nodes would incur communication and computational burdens. The system adopts a hierarchical structure, first performing local aggregation between geographically proximate or business-related nodes, and then globally aggregating these intermediate results. This approach reduces communication overhead while preserving the representativeness of the model. For example, in a financial institution with global data centers, the system might first perform regional aggregation within the Asia-Pacific, Europe, and Americas regions, and then aggregate the results from these three regions.

[0295] The third improvement is adaptive learning rate adjustment. Different parameters may update at different nodes, and the system dynamically calculates a suitable learning rate for each parameter position, making the aggregation process smoother. Specifically, a smaller learning rate is used at positions with larger parameter update magnitudes and greater differences between nodes; conversely, a larger learning rate is used to accelerate convergence. This method effectively prevents excessive oscillations in model parameters and accelerates the model's stable convergence.

[0296] The fourth improvement is knowledge distillation enhancement. In addition to parameter-level aggregation, the system also collects the prediction results of each node model on a set of common data, integrating these "soft labels" into the global model through knowledge distillation. This method can convey the model's decision boundary information, not just parameter values, achieving better performance on certain tasks. For example, the system might use a set of synthetic disaster recovery scenario data, collect the risk assessment results of each node model for these scenarios, and then incorporate this assessment knowledge into the global model through distillation.

[0297] In practice, the aggregation process of the FedAvg algorithm is iterative. In each round of communication, the central coordinator sends the current global model parameters to each node. Each node uses its local data to train and update the model, and then sends the updated parameters or parameter gradients back to the central coordinator. The coordinator applies the FedAvg algorithm to aggregate these updates, generate new global model parameters, and begin the next round of iteration. This process typically continues for multiple rounds until the global model performance reaches a preset target or the parameter changes fall below a threshold.

[0298] In terms of performance optimization, the system also implements several techniques to improve the efficiency of the FedAvg algorithm. First, communication compression technology reduces the amount of data transmitted through methods such as quantization, sparsity, or differential coding. For example, the system may only transmit updates of the first 20% of the parameter changes, which typically captures about 80% of the effective information. Second, an asynchronous aggregation mechanism does not require all nodes to submit updates simultaneously; instead, aggregation occurs only after a sufficient number of node updates have been received, improving the system's response speed and fault tolerance. Third, adaptive allocation of computing resources dynamically adjusts the batch size and number of iterations for each node participating in training based on its hardware capabilities and load conditions.

[0299] A significant advantage of the FedAvg algorithm in this system is its protection of data privacy. Data centers only need to share model parameters, not the raw data, greatly reducing the risk of data leakage. This is particularly important for industries such as finance and healthcare that handle sensitive business data. Furthermore, through techniques such as homomorphic encryption, even if parameters are intercepted during transmission, attackers cannot extract valuable information.

[0300] Specifically, in disaster recovery verification scenarios, the FedAvg algorithm has brought significant value. Different data centers typically face different types of disaster risks and failure modes. Through federated learning, the system can learn a wider range of fault response knowledge. For example, data centers located in earthquake zones have accumulated rich experience in handling physical facility failures, while data centers facing frequent cyberattacks have unique insights into security protection. The FedAvg algorithm successfully integrates these dispersed expertise, creating a "super expert" model capable of handling a broader range of disaster scenarios.

[0301] In one real-world example, after the system was deployed in 15 data centers globally, through 10 rounds of FedAvg aggregation, the global model improved fault prediction accuracy by 17.5%, reduced recovery time estimation error by 23.2%, and increased validation scenario coverage by 31.8%. More importantly, when a data center encountered an unprecedented complex fault scenario, thanks to the experience contributions from other data centers, the global model was able to provide effective coping strategies, while a locally trained model might completely fail.

[0302] This distributed collaborative learning mechanism, implemented using the FedAvg algorithm, not only improves the comprehensiveness and accuracy of disaster recovery verification but also establishes a continuously optimizing knowledge-sharing ecosystem. As data centers accumulate new verification experience, the global model continues to evolve, providing increasingly robust assurance for the business continuity of the entire organization.

[0303] Step S5.5: Securely distribute the updated global model parameters to each data center node and merge them with the local model to achieve cross-data center policy synchronization and performance optimization.

[0304] The central coordinator securely distributes the updated global model parameters to each data center node. The distribution process employs secure transmission protocols and encryption mechanisms to ensure the security of the parameters during transmission. The system also implements a digital signature mechanism to ensure that the parameters received by each node genuinely originate from the central coordinator, preventing man-in-the-middle attacks.

[0305] After receiving the global parameters, each node first performs integrity verification and version checks to ensure that the parameters have not been tampered with and are the latest version. Then, the node merges the global parameters with the local model to achieve knowledge sharing and policy synchronization.

[0306] Parameter fusion employs multiple strategies, selecting the most suitable method based on node characteristics and business needs: the direct replacement strategy completely replaces local model parameters with global parameters, suitable for scenarios with small differences between nodes; the interpolation fusion strategy uses the formula α*w_global+(1-α)*w_local to weight and combine global and local parameters, balancing generality and specificity; and the selective update strategy only updates parameters of specific layers or modules, preserving local knowledge and adaptability.

[0307] After fusion, the nodes undergo local adaptive optimization, fine-tuning and optimizing the model based on local business characteristics and data distribution. The fine-tuning process uses small-scale locally collected data and employs a small learning rate for a limited number of training rounds to ensure that the model can better adapt to the local environment while maintaining global knowledge.

[0308] The system also establishes a continuous monitoring and feedback mechanism to evaluate performance changes of the updated model. If a performance degradation is detected, the system will automatically adjust the fusion strategy or trigger a rollback mechanism to ensure that the model update does not negatively impact business operations.

[0309] Through this iterative parameter distribution, fusion, and optimization mechanism, the system achieves knowledge sharing and policy synchronization across data centers, enabling the entire distributed network to continuously learn and evolve, thereby continuously improving the intelligence level and performance efficiency of disaster recovery verification.

[0310] In another embodiment, integrating a scalable Trie construction algorithm improves data retrieval efficiency:

[0311] Step S6: Based on the access frequency, size, and type characteristics of the backup data, a multi-level index structure and adaptive sharding strategy are designed using an extensible Trie construction algorithm to build a dynamically optimized data index system;

[0312] The system first performs multi-dimensional feature analysis on the backup data, including access frequency statistics, data size distribution, and type feature identification. By analyzing the access patterns and hotspot distribution of different data, the system can identify frequently accessed data sets and the temporal variation patterns of access patterns.

[0313] Based on the analysis results, the system adopts a scalable Trie construction algorithm to design a multi-level index structure. A Trie tree (prefix tree) is a tree-like data structure, particularly suitable for prefix queries and string matching. The system extends and optimizes the traditional Trie structure to adapt to the indexing needs of large-scale backup data.

[0314] A multi-level index structure organizes data into a hierarchical index system: the first level performs coarse-grained segmentation based on data type and business classification; the second level performs medium-grained indexing based on data characteristics and attributes; and the third level establishes fine-grained indexes based on specific content and identifiers. This hierarchical design enables queries to quickly narrow down the scope and improves retrieval efficiency.

[0315] The system implements an adaptive sharding strategy, dynamically adjusting the data sharding scheme based on data size and distribution characteristics. For large datasets, the system employs horizontal sharding (partitioning by record) or vertical sharding (partitioning by field) techniques to distribute data across multiple nodes, achieving parallel processing and load balancing. The sharding strategy considers the principle of data locality, placing highly correlated data in the same shard as much as possible to reduce cross-shard queries.

[0316] The index structure and sharding strategy are not statically fixed, but dynamically optimized as data scale and access patterns change. The system monitors query performance and resource utilization, periodically evaluates the current index efficiency, and automatically adjusts index parameters and sharding layout based on the evaluation results to ensure optimal retrieval performance at all times.

[0317] Step S7: Based on the data scale and memory resource status, dynamically adjust the branching factor and depth of the Trie tree, and integrate Bloom filters and caching mechanisms to achieve millisecond-level data location and retrieval.

[0318] The system continuously monitors changes in data scale and available memory resources, and dynamically adjusts the key parameters of the Trie tree based on the monitoring results. The branching factor determines the number of child nodes of each node, affecting the width and depth of the tree; the tree depth directly relates to the query path length and query efficiency.

[0319] As the amount of data increases, the system tends to increase the branching factor, making the tree "wider" and reducing its depth, thereby reducing the number of node levels that need to be traversed during the query process. Specifically, the system calculates the optimal branching factor value by analyzing the query path length and node access frequency to minimize the average query path.

[0320] When memory resources are scarce, the system will appropriately reduce the branching factor, increase the tree depth, and reduce overall memory usage. The system uses a memory usage prediction model to calculate memory requirements under different parameter configurations and selects the configuration with the lowest memory usage while meeting performance requirements.

[0321] The system integrates a Bloom filter to improve query efficiency, especially for "not found" checks. A Bloom filter is a space-efficient probabilistic data structure used to quickly determine whether an element might exist in a set. When a query request is received, the system first performs a quick check using a Bloom filter. If the filter returns "not found," the result can be returned directly, avoiding unnecessary tree traversal.

[0322] The system also implements a multi-layered caching mechanism, including hot data caching, query result caching, and path caching. The caching strategy combines LRU (Least Recently Used) and LFU (Least Frequently Used) algorithms, and takes into account the timeliness and update frequency of the data, automatically adjusting the cache content and size.

[0323] Through the synergistic work of these optimization techniques, the system can maintain millisecond-level data location and retrieval performance even when handling petabyte-scale data, greatly improving data processing efficiency during disaster recovery verification.

[0324] In another embodiment, graph reordering approximate nearest neighbor search technique is used for fault mode identification:

[0325] Step S8: Represent historical failure cases and verification results as high-dimensional feature vectors and construct a similarity calculation framework based on graph structure to establish a failure mode knowledge base;

[0326] The system performs structured processing and feature extraction on the collected historical failure cases and verification results. Each failure case is represented as a high-dimensional feature vector containing multiple dimensions, including key attributes such as failure type, symptom characteristics, impact range, duration, and system response characteristics.

[0327] Feature representation employs a combination of techniques: categorical features are represented by one-hot encoding or embedded vectors; numerical features are normalized; text descriptions are converted into semantic vectors using NLP techniques (such as TF-IDF, Word2Vec, or BERT); and temporal data are encoded into fixed-length vectors using temporal convolutional networks or recurrent neural networks.

[0328] The system constructs a similarity calculation framework based on a graph structure, treating each failure case as a node in the graph, with edges between nodes representing the similarity relationships between cases. The weights of the edges reflect the similarity between two cases, typically calculated based on distance metrics of feature vectors (such as Euclidean distance, cosine similarity, or Mahalanobis distance).

[0329] The system employs multiple strategies to optimize the graph structure: the k-nearest neighbor graph retains only the edges between each node and its k most similar nodes, reducing the complexity of the graph; the similarity threshold filtering retains only edges with similarity exceeding a certain threshold, ensuring that the connected cases are indeed similar; and the hierarchical index divides cases into different subgraphs according to major categories, improving search efficiency.

[0330] The established fault mode knowledge base not only includes individual fault cases but also typical fault modes identified through clustering and pattern mining. The system uses graph clustering algorithms (such as the Louvain method and Infomap algorithm) to identify natural clusters in fault cases, with each cluster representing a typical fault mode or scenario. These typical patterns are stored in the knowledge base as "anchor points" to accelerate subsequent retrieval of similar cases.

[0331] Step S9: Based on the fault mode knowledge base and current fault features, optimize the accuracy and efficiency of similarity calculation through graph reordering algorithm, quickly identify fault scenarios with similar patterns and provide decision support.

[0332] When the system encounters new fault characteristics, it first converts them into feature vector representations consistent with the knowledge base, ensuring consistency in format and dimensions. Through preliminary similarity calculations, the system retrieves a batch of candidate cases from the fault mode knowledge base; these cases share similarities with the current fault in certain dimensions.

[0333] For the candidate case set, the system applies a graph re-ranking algorithm to further optimize the accuracy of similarity calculation. Graph re-ranking is a post-processing technique that uses structural information within the candidate set to adjust the initial similarity ranking. The core idea is that "good candidates should be similar to each other," meaning that if two cases are similar to the query, they are likely to be similar to each other as well.

[0334] The algorithm first constructs a local similarity graph in the candidate set, where each node is a candidate case and edges reflect the similarity relationships between cases. Then, the algorithm re-evaluates the similarity of each candidate case by analyzing the structural characteristics of the graph. Specifically, this can be implemented using methods such as random walks, label propagation, or variations of PageRank, iteratively calculating and updating the importance score of each node.

[0335] After reordering, the system is able to more accurately identify historical cases and failure modes most similar to the current failure. The system extracts key information from these similar cases, including failure cause analysis, solutions, recovery time estimates, and best practice recommendations, forming structured decision support information.

[0336] The system also analyzes the common features and differences among multiple similar cases, extracting more universally applicable handling patterns and precautions. This "pattern-level" decision support is more general and adaptable than single-case references, and can better guide the handling of current faults.

[0337] Through this graph reordering-based similar case matching mechanism, the system can quickly and accurately find the most relevant reference cases in a massive amount of historical cases, significantly improving the speed and accuracy of fault diagnosis and providing strong support for disaster recovery verification and recovery decisions.

[0338] This invention provides a data disaster recovery verification method and system based on AI technology. Through the deep integration and innovative application of multi-dimensional AI technologies, it solves the technical problems of low verification efficiency and insufficient intelligence in existing disaster recovery verification technologies. The system uses a variational autoencoder to construct a multi-dimensional feature representation of data, employs a reinforcement learning policy network to dynamically generate fault scenarios that conform to real business distributions, performs intelligent analysis through transfer learning and causal inference models, and finally achieves cross-data center policy synchronization and performance optimization through a federated learning framework. This invention significantly improves the efficiency and accuracy of disaster recovery verification, reducing verification time by 60-80% and increasing verification efficiency by 3-5 times compared to traditional methods. It also controls the prediction errors of RTO and RPO to within 5%, providing strong protection for enterprise data security and business continuity.

[0339] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0340] This disclosure also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps described in the above method embodiments. The storage medium may be a volatile or non-volatile computer-readable storage medium.

[0341] In addition, this disclosure also provides a computer program product, on which a computer program is stored. When the computer program is run by a processor, it executes the steps provided in any of the above embodiments of this disclosure. For details, please refer to the above method embodiments, which will not be repeated here.

[0342] The aforementioned computer program product can be implemented through hardware, software, or a combination thereof. In one optional embodiment, the computer program product is specifically embodied in a computer storage medium, which can be a volatile or non-volatile computer-readable storage medium. In another optional embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0343] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices and apparatuses described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. In the several embodiments provided in this disclosure, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection may be through some communication interfaces; the indirect coupling or communication connection of devices or units may be electrical, mechanical, or other forms.

[0344] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0345] In addition, the functional units in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0346] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0347] Finally, it should be noted that the above-described embodiments are merely specific implementations of this disclosure, used to illustrate the technical solutions of this disclosure, and not to limit it. The protection scope of this disclosure is not limited thereto. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this disclosure; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be covered within the protection scope of this disclosure. Therefore, the protection scope of this disclosure should be determined by the protection scope of the claims.

Claims

1. A data disaster recovery verification method based on AI technology, characterized in that, include: The backup data and historical fault data from multiple data centers are acquired, and the backup data are subjected to feature extraction and semantic analysis using a variational autoencoder to construct a multidimensional feature representation that includes metadata profiles and business semantic tags. Based on the multidimensional feature representation and the historical fault data, a reinforcement learning strategy network with an Actor-Critic dual-network structure is used to learn fault modes and dynamically generate scenario-based fault injection instructions that conform to the real business distribution. Based on the scenario-based fault injection instruction, a directed acyclic graph of fault events is constructed and the timing arrangement is optimized using time-optimal directed q-analysis technology. Scenario-based fault injection is then performed and verification data is obtained. The verification data is feature-encoded using a pre-trained Transformer model, and intelligent analysis is performed using transfer learning techniques and causal inference models to output an interpretable verification report that includes assessments of data integrity, consistency, and recoverability. Based on the performance metrics and verification results in the interpretability verification report, the model parameters of each data center node are aggregated and updated through the federated learning framework, thereby achieving policy synchronization and performance optimization across data centers.

2. The method according to claim 1, characterized in that, A variational autoencoder is used to perform feature extraction and semantic analysis on the backup data, constructing a multi-dimensional feature representation that includes metadata profiles and business semantic tags, including: Based on the backup data, a variational autoencoder is used to compress the high-dimensional original backup data into a low-dimensional latent space to obtain the latent feature representation of the data. The latent feature representation is then used to reconstruct the data using a decoder and calculate the reconstruction error, thereby learning the intrinsic structure and semantic features of the data. Based on the aforementioned intrinsic structure and semantic features, a feature vector indexing system is established using an extensible Trie construction algorithm. The branching factor and tree depth are dynamically adjusted according to the data scale and query pattern to generate an efficient feature retrieval structure. For the relationships between the backup data, a data relationship graph is constructed using graph embedding technology and vector mapping is performed using the Node2Vec algorithm to form business semantic tags containing topology information; By integrating the latent feature representation, the feature retrieval structure, and the business semantic label, a unified feature representation framework is constructed to obtain the multidimensional feature representation.

3. The method according to claim 1, characterized in that, Construct a multidimensional feature representation that includes metadata profiles and business semantic tags, including: Based on the basic attributes and statistical characteristics of the backup data, a metadata profile containing data attributes, statistical characteristics, and quality indicators is generated through data quality analysis and importance assessment. The business meaning of the backup data is understood and classified using natural language processing technology and domain knowledge base, and business semantic tags reflecting the business meaning and importance of the data are constructed. Based on the dependencies and references between the backup data, graph theory algorithms are used to analyze data association patterns and calculate the scope of influence, and an association graph describing the data dependencies and scope of influence is established. The metadata profile, the business semantic tags, and the association graph are uniformly encoded and standardized using a feature fusion algorithm to form the multidimensional feature representation.

4. The method according to claim 1, characterized in that, A reinforcement learning strategy network with an Actor-Critic dual-network structure is used for fault mode learning to dynamically generate scenario-based fault injection instructions that conform to the distribution of real business operations, including: Based on the fault types, frequency of occurrence, and scope of impact in the historical fault data, a fault knowledge base containing the correlation between fault occurrence patterns and business characteristics is constructed through data mining and pattern recognition to obtain prior knowledge of fault patterns. Using prior knowledge of the aforementioned fault modes and the current system state, a strategy decision is made through an Actor network, including fault type selection, impact scope determination, and injection timing control, to generate an initial fault injection strategy. For the initial fault injection strategy, the Critic network analysis is used to verify the effectiveness, business impact and recovery performance after the strategy is executed, and the strategy value assessment results are output. Based on the strategy value evaluation results, the parameters of the Actor network and Critic network are updated through the strategy gradient algorithm to optimize the fault injection strategy generation capability. Based on the optimized network parameters and current verification requirements, the Monte Carlo tree search algorithm is used to explore the optimal fault combination scheme and generate the scenario-based fault injection instruction.

5. The method according to claim 1, characterized in that, A directed acyclic graph of fault events is constructed using time-optimal directed q-analysis techniques, and the timing arrangement is optimized. Contextualized fault injection is then performed, and verification data is obtained, including: Based on the fault events in the scenario-based fault injection instructions, the causal relationships and temporal constraints between the events are analyzed, and a directed acyclic graph with fault events as nodes and dependencies as edges is constructed. For the directed acyclic graph, the q-learning algorithm is used to learn the verification effect of fault combinations under different time windows, and the optimal timing arrangement that maximizes verification coverage and minimizes business impact is calculated. According to the optimal timing arrangement, the fault injection operation is executed sequentially in chronological order, and the system response is monitored in real time to collect verification data including system state changes, recovery process and performance indicators.

6. The method according to claim 1, characterized in that, Feature encoding is performed using a pre-trained Transformer model, and intelligent analysis is conducted using transfer learning techniques and causal inference models. The output is an interpretable verification report containing assessments of data integrity, consistency, and recoverability, including: Based on the verification requirements, data features, and fault scenario information in the verification data, the self-attention mechanism of the pre-trained Transformer model is used to encode multimodal information and generate a unified vector representation. Using the unified vector representation, feature extraction is performed on a base model pre-trained on a large-scale disaster recovery scenario dataset, and fine-tuning techniques are used to adjust parameters for the current specific verification scenario to obtain an adaptively enhanced feature representation. Based on the adaptive enhancement feature representation, a structured causal model is used to analyze the causal relationship between data integrity, consistency and recoverability indicators, and to identify the root causes and critical paths affecting disaster recovery effectiveness. Based on the critical path and the current verification scenario, the graph reordering approximate nearest neighbor search technique is used to match similar historical verification cases to obtain verification decision support information. By integrating the root cause, the critical path, and the verification decision support information, an interpretable verification report containing verification conclusions, key indicators, reasoning processes, and key evidence is generated based on an attention mechanism, thus obtaining the interpretable verification report.

7. The method according to claim 1, characterized in that, The federated learning framework is used to aggregate and update model parameters across data center nodes, enabling cross-data center policy synchronization and performance optimization, including: At each data center node, a verification result collector and performance monitoring component are deployed to collect performance metrics such as verification accuracy, execution efficiency, resource consumption, and fault detection rate in real time, thereby obtaining local training data. Based on the local training data, each node independently performs incremental training of the validation model and calculates the gradient update of the model parameters to generate local gradient information. The local gradient information is securely processed using homomorphic encryption and differential privacy techniques, and then uploaded to the central coordinator via a secure aggregation protocol. In the central coordinator, the FedAvg algorithm is used to perform weighted aggregation based on the data quality, number of training samples and historical contribution of each node to generate updated global model parameters. The updated global model parameters are securely distributed to each data center node and merged with the local model to achieve cross-data center policy synchronization and performance optimization.

8. The method according to claim 1, characterized in that, It also includes integrating scalable Trie building algorithms to improve data retrieval efficiency: Based on the access frequency, size, and type characteristics of backup data, a multi-level index structure and an adaptive sharding strategy are designed using an extensible Trie construction algorithm to build a dynamically optimized data index system. Based on the data scale and memory resource status, the branching factor and depth of the Trie tree are dynamically adjusted, and Bloom filters and caching mechanisms are integrated to achieve millisecond-level data location and retrieval.

9. The method according to claim 1, characterized in that, It also includes using graph reordering approximate nearest neighbor search techniques for fault mode identification: Historical failure cases and verification results are represented as high-dimensional feature vectors, and a similarity calculation framework based on graph structure is constructed to establish a failure mode knowledge base. Based on the fault mode knowledge base and current fault characteristics, the accuracy and efficiency of similarity calculation are optimized by graph reordering algorithm, which can quickly identify fault scenarios with similar patterns and provide decision support.

10. A data disaster recovery verification system based on AI technology, characterized in that, It includes a processor and a memory, wherein the processor is configured to perform the method according to any one of claims 1-9.

Images