An automated testing method based on knowledge base management

Abstract

This invention discloses an automated testing method based on knowledge base management. The method first acquires telemetry data streams from the target device in real time and calculates the gradient of its higher-order statistical moments to detect phase transition characteristics of the system transitioning from a steady state to an unstable state. When a phase transition characteristic is detected, the posterior probability of a specific failure mode is updated based on a Bayesian network risk model. If the posterior probability exceeds a dynamic risk threshold, relevant diagnostic test scripts are automatically retrieved and matched from an intelligently managed knowledge base. This knowledge base uses natural language processing and graph neural network technology to intelligently classify and associate historical test problems. Finally, the diagnostic scripts are injected into the target device for execution to proactively verify potential faults, and the knowledge base and risk model are updated based on the execution results. This invention achieves a shift from passive response to proactive prediction and diagnosis, significantly improving fault predictability, testing efficiency, and system reliability.

Description

Technical Field

[0001] This invention relates to the fields of automated testing, knowledge base management, artificial intelligence, and predictive maintenance, and in particular to an artificial intelligence-based test script management and predictive fault diagnosis method and system. Background Technology

[0002] In the development and maintenance cycle of modern software and embedded systems, automated testing plays a crucial role, serving as a key link in ensuring product quality and accelerating iteration efficiency. However, the current field of automated testing generally faces the severe challenge of inefficient script management. Test scripts are typically stored in a fragmented and isolated manner in the local working environments of individual test engineers or in scattered code repositories, severely lacking a unified, centralized version control and standardized management mechanism.

[0003] This decentralized management model directly leads to several core problems. First, the reusability of test scripts is extremely low. Due to the lack of a shared and retrieval platform, test engineers from different projects often need to repeatedly write test scripts with highly similar functional logic, such as CRUD verification for databases and boundary condition testing for user authentication interfaces. This repetitive work not only consumes a lot of valuable human resources but also slows down the overall testing progress. Second, version consistency is difficult to guarantee. In the process of rapid product iteration, changes in underlying interfaces or business logic require synchronous updates to related test scripts. Under the decentralized management model, it is impossible to effectively track and distribute the latest version of scripts, and testers may still be testing based on outdated scripts, resulting in a significant reduction in the validity and reliability of test results. Finally, testing knowledge and experience are difficult to accumulate and pass on. Defects, solutions, efficient test cases, and SQL query statements discovered during the testing process constitute valuable testing assets for the team. However, these assets are often lost with personnel turnover or project termination, failing to form a sustainable knowledge system for accumulation and utilization. New members need to spend a lot of time exploring and repeating problems already solved by predecessors.

[0004] Meanwhile, traditional system monitoring and troubleshooting methods are typically reactive and reactive. They mostly rely on post-event analysis of system logs or setting fixed performance alarm thresholds (such as CPU utilization exceeding 90%). The limitation of these methods is that by the time an alarm is triggered, system performance has often already significantly deteriorated or even failed, leaving very limited reaction time for operations and testing personnel. They lack the ability to proactively diagnose and provide early warnings by analyzing subtle changes in system operating status before a failure actually occurs.

[0005] Therefore, how to build a unified system for managing and efficiently reusing test scripts, accumulating test knowledge assets, and combining predictive maintenance technology to achieve a shift from passive response to proactive prediction and diagnosis is a pressing technical challenge in the field of automated testing. Summary of the Invention

[0006] Technical problems to be solved This invention aims to address the problems of fragmented management, low reusability, and difficulty in version tracing in existing automated test scripts, as well as the difficulty in accumulating and utilizing test knowledge assets. Furthermore, this invention aims to overcome the passive and lagging shortcomings of traditional fault diagnosis methods, providing a proactive testing method capable of predictive diagnosis before system failures occur.

[0007] Technical solution To achieve the above objectives, this invention provides an automated testing method based on knowledge base management. The core of this method lies in constructing a closed-loop system for proactive diagnosis, from predictive risk perception to intelligent knowledge-driven approaches. The method includes the following steps: First, the telemetry data stream of the target device is acquired in real time. The telemetry data stream is continuous performance index data reflecting the operating status of the device.

[0008] Next, the gradient of at least one higher-order statistical moment of the telemetry data stream is calculated to detect phase transition characteristics that characterize the transition of the target device's operating state from a steady state to an unstable state. This step aims to detect early signs of system instability by capturing dynamic changes in data distribution, rather than relying on static thresholds.

[0009] When the phase transition feature is detected, the posterior probability of the specific failure mode associated with the phase transition feature is updated based on a pre-defined Bayesian network risk model. This model quantifies the probabilistic relationship between observed anomalous system behavior and potential fundamental failures.

[0010] Then, it is determined whether the updated posterior probability exceeds a dynamically set risk threshold.

[0011] If the posterior probability exceeds the dynamic risk threshold, it indicates that the probability of a specific failure mode occurring has increased significantly. At this time, the system will automatically retrieve and match diagnostic test scripts that are strongly correlated with the specific failure mode from a knowledge base that stores historical test problems and solutions.

[0012] Finally, the matched diagnostic test script is injected into the target device and executed. By running targeted tests, potential faults are proactively verified, thereby identifying the root cause of the problem before the fault fully erupts.

[0013] Beneficial effects Compared with existing technologies, the technical solution provided by this invention brings significant benefits by introducing a phase transition detection and dynamic risk assessment model based on higher-order statistical moments, combined with intelligent knowledge base management: Significantly improved fault predictability: This invention detects the "phase transition" of the system state by calculating the gradient of changes in higher-order statistical moments such as variance and skewness of the telemetry data stream in real time. This method is extremely sensitive to the early transition characteristics of the system from stable to unstable. Compared with traditional alarm methods that rely on fixed thresholds or log keyword matching, it can advance the warning time of potential system failure risks (e.g., failure risks in firmware online upgrade (FOTA) scenarios) by 30% to 50%, gaining valuable time windows for preventive intervention.

[0014] Revolutionary Improvements in Testing Targeting and Efficiency: This invention constructs a closed-loop adaptive testing system encompassing "perception-decision-execution." It abandons the inefficient model of large-scale, indiscriminate regression testing when risks are unknown, instead employing a "surgical" approach of precise diagnosis. The system automatically injects the most relevant diagnostic script only when the risk model indicates a high probability of failure for a specific device. This on-demand, precise testing method significantly reduces resource consumption (such as computation and network bandwidth) and interference with normal business operations during the testing process.

[0015] Enhanced System Reliability and Knowledge Asset Value: By proactively diagnosing and intervening before failures actually occur (e.g., pausing upgrades to high-risk equipment groups or switching to more conservative execution strategies), large-scale failure events can be effectively avoided, reducing the overall failure rate by an order of magnitude, thereby significantly enhancing system stability and reliability. Simultaneously, the entire closed-loop process of "prediction-diagnosis-verification-feedback" continuously updates and optimizes the knowledge base and risk model, enabling test assets not only to be accumulated but also to continuously increase in value through self-learning, forming a sustainably evolving intelligent testing system. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of the intelligent classification and processing of historical test problems provided in this embodiment of the invention.

[0018] Figure 2 This is a flowchart of the standardized SQL template storage and optimization process provided in this embodiment of the invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1 This embodiment provides a complete implementation process for an automated testing method based on knowledge base management. This method integrates predictive risk perception, dynamic decision-making, and intelligent knowledge base management to form a proactive closed-loop system for fault diagnosis and verification.

[0021] The method can be implemented by an automated test management platform deployed on a server, which communicates with one or more target devices via a network. The target devices can be embedded devices, servers, virtual machines, or any computing unit capable of generating telemetry data.

[0022] Step 1: Real-time telemetry data acquisition The system first collects telemetry data streams from the target device in real time through a data acquisition module. This module is implemented by deploying a lightweight agent on the target device or utilizing the device's own monitoring interfaces (such as SNMP or JMX). The collected telemetry data streams are multi-dimensional time-series data, specifically including but not limited to device performance metrics selected from at least one of the following: CPU utilization, memory usage, disk I / O rate, network throughput, network retransmission rate, application response latency, etc. This data is collected at a preset time frequency (e.g., once per second) and transmitted to the test management platform for subsequent processing. This step provides the raw data foundation for subsequent risk analysis.

[0023] Step 2: Phase transition detection based on higher-order statistical moments The test management platform has a built-in phase transition detection engine. The core responsibility of this engine is to analyze the received telemetry data stream to detect early "phase transition" characteristics as the system transitions from a steady state to an unstable state. Unlike traditional methods that set fixed thresholds (such as CPU utilization > 90%), this engine does not focus on the absolute value of the indicator, but rather on the dynamic changes in its statistical distribution.

[0024] Specifically, the phase transition detection engine calculates the first moment (mean), second central moment (variance), and third central moment (skewness) of each dimension of the data stream in real time within a continuous, sliding time window (e.g., a window size of 5 minutes).

[0025] Mean (μ): Reflects the average level of the data stream.

[0026] Variance (σ²): Reflects the dispersion or volatility of the data stream. When a system transitions from a steady state to an unstable state, its internal random disturbances intensify, typically manifested as a significant increase in variance.

[0027] Skewness reflects the symmetry of data distribution. In a stable system, the performance metrics distribution typically approximates a normal distribution, with skewness close to zero. When anomalies occur in the system, the symmetry of the distribution is often disrupted, causing the skewness to deviate significantly in the positive or negative direction.

[0028] Phase transition detection is specifically achieved by monitoring a positive surge in the variance gradient and a deviation of the skewness gradient from near zero to a non-zero value. When the engine detects a sharp positive pulse in the variance gradient within a short period, coupled with a significant deviation in the skewness gradient, the simultaneous occurrence of these two events is identified as a high-confidence phase transition signal. This detection method can capture early signs that the system's internal dynamic equilibrium has been disrupted before macroscopic indicators exceed conventional thresholds, thus achieving proactive perception.

[0029] Step 3: Dynamic Risk Assessment and Decision-Making Once the phase transition detection engine detects a phase transition feature, it immediately transmits the event information to the risk assessment module. This module uses a pre-defined Bayesian network risk model to quantitatively assess the potential failure risk.

[0030] This Bayesian network predefines conditional probabilistic relationships between "phase transition features" (as evidence nodes) and various "specific failure modes" (as query nodes, such as "flash write failure," "database connection pool exhaustion," and "memory leak"). These relationships can be initialized using domain expert knowledge and trained and optimized using historical failure data.

[0031] Upon receiving new evidence of phase transition characteristics, the risk assessment module uses a Bayesian inference algorithm to update the posterior probability of each node in the network that exhibits a specific failure mode. This posterior probability intuitively represents the likelihood of a specific failure occurring given the observed abnormal behavior of the current system.

[0032] The system then compares the updated posterior probability with a dynamic risk threshold. This threshold is not static but adaptively adjusted based on the target device's business criticality level and historical failure frequency. Specifically, the dynamic risk threshold is inversely proportional to both the business criticality level and the historical failure frequency. For example, for devices carrying core business functions and historically prone to failure, the risk threshold will be set lower, thereby enabling a more sensitive early warning strategy. This adaptive mechanism ensures that testing resources are prioritized for the highest-risk components.

[0033] Step 4: Intelligent knowledge base-driven script retrieval and injection If the posterior probability of a specific failure mode exceeds its corresponding dynamic risk threshold, the system initiates an active diagnostic process. At this time, the risk assessment module sends a retrieval request to the knowledge base management module, which includes the identifier of the identified high-risk failure mode (e.g., "concurrent deadlock").

[0034] The knowledge base management module is one of the core components of this invention, responsible for storing, managing, and providing test assets. This knowledge base employs a distributed storage architecture to handle the storage of massive amounts of test data and the demands of high-concurrency access. Specifically, a NoSQL database (such as MongoDB) is used to store unstructured raw test data and logs, such as detailed issue description documents, screenshots, and stack trace information. Simultaneously, a relational database (such as PostgreSQL) is used to store structured metadata and index information for historical test issues, such as issue IDs, category tags, and associated diagnostic script IDs. This hybrid storage architecture balances flexibility and query efficiency.

[0035] To achieve intelligent retrieval, the historical test questions in the knowledge base have undergone sophisticated automated processing. This is primarily accomplished by the intelligent classification module, and its processing flow can be found in [link to documentation]. Figure 1 As shown.

[0036] Text Preprocessing and Feature Extraction: When a new historical test question (including the question description text) is entered, the system first performs text preprocessing, including filtering special characters, segmenting words using a domain-specific dictionary, stemming, and stop word filtering. Then, a hybrid model combining TF-IDF and Word2Vec is used for feature extraction. TF-IDF is used to capture local keyword features of the text, and its calculation formula is as follows: in, It is the frequency of word i in document j. N is the number of documents containing word i, and N is the total number of documents. Word2Vec uses a model pre-trained on a large corpus of test domains to convert words into dense vectors that capture semantic information. The document vector is then obtained by a weighted average of the word vectors it contains. in is the 300-dimensional word vector of word k, and n is the number of words in the document.

[0037] The final text feature is a concatenation of the TF-IDF weight vector and the Word2Vec semantic vector, which takes into account both keyword information and deep semantics.

[0038] Hierarchical Intelligent Classification: After feature extraction, the system employs a hierarchical classification architecture to classify the problem. The first layer uses a Support Vector Machine (SVM) classifier for coarse-grained domain partitioning (e.g., dividing into broad categories such as "database," "interface," and "UI"). SVM classifies by constructing a hyperplane, and its decision function is: Where K(x_i, x) is the kernel function, and in this embodiment, a radial basis function (RBF) kernel adjusted for the text features of the test question is used. For the text feature distribution of the test question, the parameter γ is set to 0.5 in this embodiment.

[0039] The second layer employs a Bi-LSTM+Attention network integrated with an attention mechanism for fine-grained intent recognition and root cause classification for each coarse-grained category (e.g., within the broad category of "database," further subdivided into subcategories such as "slow query," "concurrency deadlock," and "index failure"). BiLSTM captures the bi-directional dependencies within text sequences, while the attention mechanism allows the model to focus on the words most important for classification. Its weight calculation conforms to the Softmax function form. Wherein is the alignment score between the hidden state at time step t and a trainable context vector.

[0040] Knowledge Graph Construction: After classification, the system automatically constructs the relationships between historical test questions to form a knowledge graph based on graph neural network algorithms. Specifically, the GraphSAGE algorithm is used to generate feature vectors that aggregate neighbor information for each classified question node (and associated solution and script nodes). The aggregation and update process can be represented as follows: in : The feature vector of node v at layer k (the updated representation after layer k) σ(⋅): Activation function (such as ReLU, sigmoid, etc.), used to introduce nonlinear transformations. The learnable weight matrix of the k-th layer, used to perform a linear transformation on the concatenated features. CONCAT(⋅,⋅): Concatenates two vectors according to their dimensions to form a higher-dimensional vector. : The feature vector of node v in the k-1 layer (the feature vector of the previous layer itself) : The feature aggregation result of the set of neighboring nodes N(v) of node v at the k-1 level (usually a vector after aggregation operations such as the mean, sum, or maximum value of the neighboring features). In this way, potential relationships between problems and between problems and solutions can be uncovered (for example, two problems with different appearances may be caused by the same underlying reason). The effectiveness of these relationships (edges) can then be learned and evaluated using an embedding model similar to TransE, with the following scoring function: Where h is the embedding vector of the head entity, representing the starting entity of the triple in the knowledge graph; r is the embedding vector of the relation, representing the association between the head entity and the tail entity; and t is the embedding vector of the tail entity, representing the entity in the triple that is connected to the head entity through a relation. L1 norm (Manhattan distance) calculates the sum of the absolute values ​​of vector differences; The score of the triple (h, r, t) indicates that the triple (h, r, t) is more likely to be real and valid in the knowledge graph.

[0041] When a retrieval request arrives, the system uses the aforementioned classification tags and knowledge graph to quickly locate diagnostic test scripts that are directly related to or semantically similar to high-risk failure modes (such as "concurrency deadlock") and returns them.

[0042] In addition, the knowledge base stores a large number of standardized SQL templates, which are managed by the SQL template management module. The processing flow is detailed in [link to relevant documentation]. Figure 2 As shown.

[0043] Standardized Template Generation: When a tester submits a commonly used test SQL statement, this module first parses it using parsing tools such as ANTLR4 to construct an Abstract Syntax Tree (AST). By traversing the AST, it can identify the variable parts in the SQL statement (such as values ​​in the WHERE condition, the numerical value of LIMIT, etc.) and parameterize these variables to generate a standardized SQL template, such as SELECT. The expression `FROM ${table} WHERE ${column} = ${value} LIMIT ${limit}` is used to parse the parameterized data. Simultaneously, metadata descriptions, such as type and value range constraints, are generated for each parameter. This AST-based parsing method ensures the accuracy and robustness of the parameterization process.

[0044] To achieve efficient template storage and comparison, the system further calculates the structural hash value of the abstract syntax tree. The calculation formula is as follows: Where `nodeTypei` is the node type encoding, `depthi` is the node depth, `A` is the base of a prime number (e.g., 31), and `(mod264)` is the modulo operation, limiting the hash value to a 64-bit integer. This hash value serves as a unique fingerprint of the SQL structure, used to quickly detect the existence of historical templates with the same structure.

[0045] Intelligent Performance Optimization: This module also features performance optimization suggestions. Based on historical SQL execution data associated with each SQL template (such as execution time, number of rows scanned, and whether indexes were used), it constructs a cost model to predict query performance. This model can be a linear model. Here, α, β, and γ are weight coefficients learned through regression analysis. Furthermore, the system employs a reinforcement learning algorithm, treating the database's execution planner as the environment and the cost model as part of the reward function. An agent explores different query rewriting or index addition strategies (as the action space) to proactively recommend performance optimization suggestions that reduce query costs to the user.

[0046] The reward function R is defined as: Wherein, exec_time is the estimated or actual execution time of the query, and taking the reciprocal indicates that the higher the efficiency, the greater the reward; index_size is the storage space occupied by the index; λ is the weight coefficient, used to balance the trade-off between query efficiency and storage resource consumption.

[0047] For complex queries, heuristic algorithms such as differential evolution can be used for optimization exploration. The mutation operation can be represented as: in, Represents the feature vector of the query plan; : No. g +1 generation i Candidate solutions for each individual (new solutions generated after mutation); , , : No. g Three different individuals were randomly selected from the generation ( r 1, r 2, r 3 are all distinct and not equal to i (random index), where As basis vectors, is the difference vector; F=0.8 is the scaling factor (usually a constant between 0 and 1), used to control the influence of the difference vector and determine the search step size of the mutation operation.

[0048] Version Management: To track the evolution of SQL templates, the system also implements version management functionality. When a template is modified, an improved Myers difference algorithm is used to efficiently detect the differences between the old and new versions of the AST. Its core function is to find the path D with the minimum edit distance. Where N and M are the version lengths, and k is the edit path. D ( N , M ): Function value, representing the maximum number (or maximum sequence number) of a path that meets a specific condition. Number is k The path (which can be understood as a sequence of nodes or coordinates from the starting point to a certain point); : Indicates that the endpoint of the path is at coordinates ( x , y (or a certain one) x , y (The node is identified by )); max{⋅}: retrieves the maximum value, i.e., starts from all nodes whose endpoints are ( x , y In the path, select the one with the largest number. k。

[0049] The cosine similarity between the embedding vectors of the two versions of the AST is calculated to assess the degree of semantic change. sim(vi,vj): represents the similarity between two ASTs (corresponding to vi and vj respectively, which may be AST representations of code snippets, query statements, etc.), with a value ranging from [-1, 1]. The closer the value is to 1, the higher the similarity.

[0050] hASTi and hASTj: These are the embedding vectors of AST_i and AST_j, respectively (the AST structure is transformed into a low-dimensional dense vector by the model to capture the structural and semantic features of the AST).

[0051] hASTi and hASTj: The inner product (dot product) of two embedded vectors, used to measure the consistency of the vectors in direction.

[0052] hASTi and hASTj are the L2 norms (modulus lengths) of two embedding vectors, used to normalize the inner product result.

[0053] This makes version tracking and change review more intuitive and efficient.

[0054] Step 5: Active verification and feedback loop Finally, the system retrieves diagnostic test scripts from the knowledge base and injects them into the target devices identified as high-risk via the device control channel, triggering their execution. For example, if the identified risk is "database concurrency deadlock," the injected script might be a test case simulating high-concurrency transaction conflicts.

[0055] After the script executes, its results (success, failure, performance data, logs, etc.) are sent back to the test management platform. Upon receiving the execution results, the system performs the following two key feedback actions: Update the knowledge base: If the execution result successfully reproduces or verifies the potential fault, the system will store this diagnostic process (phase transition characteristics -> failure mode -> diagnostic script -> verification success) as a new success case in the knowledge base, thereby enriching the test assets and making future diagnoses more accurate.

[0056] Updating the risk model: Simultaneously, the validation results are used to update the Bayesian network risk model. Specifically, when the diagnostic script successfully validates a potential failure, the system strengthens (increases weights) the correlation between the observed phase transition feature and the validated specific failure mode within the Bayesian network. This continuous learning and optimization based on practical results makes the risk assessment model increasingly accurate over time.

[0057] Through the above steps, this embodiment constructs a complete automated testing loop from risk prediction, intelligent decision-making, knowledge-driven diagnosis to feedback learning, which greatly improves the initiative, accuracy and efficiency of testing.

[0058] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An automated testing method based on knowledge base management, characterized in that, include: Real-time acquisition of telemetry data streams from target devices; Calculate the gradient of at least one higher-order statistical moment of the telemetry data stream to detect phase transition characteristics that characterize the transition of the target device's operating state from a steady state to an unstable state; When the phase transition feature is detected, the posterior probability of the specific failure mode associated with the phase transition feature is updated based on the preset Bayesian network risk model. Determine whether the posterior probability exceeds the dynamic risk threshold; If the posterior probability exceeds the dynamic risk threshold, then a diagnostic test script related to the specific failure mode is automatically retrieved and matched from a knowledge base storing historical test questions. The diagnostic test script is injected into the target device and executed to proactively verify potential faults.

2. The method according to claim 1, characterized in that, The calculation of the gradient of at least one higher-order statistical moment of the telemetry data stream specifically includes: Within a continuous time window, the first moment (mean), second central moment (variance), and third central moment (skewness) of the telemetry data stream are calculated. The detection of the phase transition characteristics is achieved by monitoring the positive surge of the variance gradient and the deviation of the skewness gradient from near zero to non-zero values. The positive surge and deviation together characterize the disruption of the symmetry of the stable distribution of the system.

3. The method according to claim 1, characterized in that, The dynamic risk threshold is adaptively adjusted based on the business criticality level and historical failure frequency of the target equipment; specifically, the dynamic risk threshold is inversely proportional to the business criticality level and inversely proportional to the historical failure frequency, thereby adopting a more sensitive early warning strategy for equipment with high criticality or high failure rate.

4. The method according to claim 1, characterized in that, The historical test questions in the knowledge base are classified hierarchically using natural language processing technology. The classification includes: A support vector machine (SVM) classifier is used to perform coarse-grained domain segmentation of the problem description text; For each coarse-grained domain, a bidirectional long short-term memory network (BiLSTM+Attention) with an integrated attention mechanism is used for fine-grained intent recognition and root cause classification.

5. The method according to claim 4, characterized in that, The method further includes: After completing the hierarchical intelligent classification, the associations between the historical test questions are automatically constructed based on the graph neural network algorithm to form a knowledge graph. Specifically, the GraphSAGE algorithm is used to generate feature vectors for each classified question node, and by learning the aggregation method of neighboring nodes, potential associations between questions and between questions and solutions are mined and established.

6. The method according to claim 1, characterized in that, The knowledge base also stores standardized SQL templates generated by parsing the abstract syntax tree (AST) of the original SQL statement; The generation process includes: parsing the original SQL statement to construct an abstract syntax tree, traversing the abstract syntax tree to identify and extract the variable parts in the query, and parameterizing the variable parts to form the standardized SQL template.

7. The method according to claim 6, characterized in that, The method further includes: Based on historical SQL execution data associated with the standardized SQL template, a cost model for predicting query performance is constructed; and, using a reinforcement learning algorithm, the cost model is used as an environment to explore different query rewriting or index addition strategies through an agent, so as to proactively recommend performance optimization suggestions that can reduce query costs to the user.

8. The method according to claim 1, characterized in that, The telemetry data stream includes device performance metrics selected from at least one of the following: CPU utilization, memory usage, disk I / O rate, network throughput, and application response latency.

9. The method according to claim 1, characterized in that, The method further includes: Receive the execution results of the diagnostic test script on the target device; The knowledge base and the Bayesian network risk model are updated based on the execution results, wherein when the execution results successfully verify a potential fault, the association weight between the detected phase transition feature and the specific failure mode in the Bayesian network is enhanced.

10. The method according to claim 1, characterized in that, The knowledge base adopts a distributed storage architecture, in which a NoSQL database is used to store unstructured raw test data and logs, and a relational database is used to store the metadata and index information of historical test questions that have been structured.

Images