Artificial intelligence-based flash particle test data processing method and device

Abstract

This application discloses an artificial intelligence-based method and device for processing Flash particle test data. The method includes: collecting particle response behavior data from Flash storage particle test tasks and extracting time-series correlation logic information; creating a dynamic reasoning model based on an adapted behavior sequence analysis algorithm to mine the evolutionary patterns under different instruction requests and generate hidden evolutionary trajectories; deriving the degradation trend prediction results through particle performance evolution characteristics, converting trend features into semantic knowledge nodes and establishing a semantic mapping relationship with failure types; and generating a structured test knowledge graph by combining test log data. This application breaks through the limitations of traditional single feature judgment, provides a new dimension for Flash particle lifetime prediction, and breaks down the information barrier between test data and design optimization, enabling efficient transformation from data to decision-making.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and device for processing Flash particle test data based on artificial intelligence. Background Technology

[0002] As a core component in data storage, the performance stability and lifespan of Flash memory chips directly determine the quality and reliability of storage products. With the continuous development of storage technology, the structure of Flash chips is becoming increasingly complex, and the requirements for their testing are becoming more stringent. Currently, Flash chip testing mainly focuses on electrical characteristics, storage performance, and reliability. Devices such as voltage sensors, current sensors, and temperature sensors are used to collect response data of the chips at different testing stages, such as voltage fluctuations during erase / write commands and current changes under read requests. To extract useful information from the test data, traditional techniques often employ single-feature analysis methods, such as monitoring changes in a fixed voltage or current threshold to determine if the chip has a failure risk. These methods can, to some extent, detect the basic performance of the chips, but their limitations become increasingly apparent as the complexity of the chips increases. Furthermore, to improve testing efficiency and accuracy, some technologies have introduced simple time-series analysis methods to attempt to determine trends in continuously collected test data. However, these methods often remain at the level of analyzing surface-level data trends and fail to delve into the underlying logical relationships within the data.

[0003] In the application of Flash particle test data, traditional technologies often store test data in log format. Testers need to manually analyze and organize the log data to extract information related to particle failure. To facilitate data management and retrieval, some technologies attempt to build simple databases for structured storage of test data. However, these databases can only achieve simple classification and retrieval of data and cannot establish deep relationships between data, resulting in the test data's value not being fully realized.

[0004] With the widespread application of artificial intelligence (AI) technology in various fields, some research has begun to explore its application in Flash particle testing. The aim is to use machine learning algorithms to analyze test data and predict particle failure. However, existing research largely focuses on the application of single algorithms, failing to form a complete technical framework, and exhibits significant shortcomings in data correlation logic mining and knowledge transformation. It is evident that current technologies struggle to effectively uncover the temporal correlation logic of particle response behavior in Flash particle testing, making it impossible to pinpoint the degradation trend of particle stability in advance. Furthermore, information barriers exist between test data and design optimization, making it difficult for fragmented test data to directly guide product iteration. Summary of the Invention

[0005] This application provides an artificial intelligence-based method and device for processing Flash particle test data.

[0006] One embodiment of this application provides an artificial intelligence-based method for processing Flash particle test data, applied to computer equipment, the method comprising:

[0007] Collect particle response behavior data generated during the execution of Flash storage particle test tasks, and extract the temporal correlation logic information of particle response behavior based on the particle response behavior data;

[0008] Based on the behavior sequence analysis algorithm adapted to the Flash storage particle test task, and the particle response behavior characteristics of different test stages in the time-series correlation logic information, a dynamic reasoning model based on behavior sequence is created.

[0009] The particle response behavior data corresponding to different instruction requests in the Flash storage particle test task are input into the dynamic inference model for time series analysis to explore the evolution law of particle response behavior under different instruction requests and generate hidden evolution trajectories corresponding to different instruction requests.

[0010] By analyzing the particle performance evolution characteristics contained in the hidden evolution trajectory, the occurrence trend and impact range of particle test failure events are deduced, and the degradation trend prediction results associated with the particle test failure events are obtained.

[0011] Each trend feature in the decline trend prediction result is converted into a corresponding semantic knowledge node, and a semantic mapping relationship between trend features and failure types is established by combining each semantic knowledge node and particle test failure events.

[0012] The test log data generated during the execution of the particle test task is called in the Flash storage. Test status information related to the particle response behavior and the particle test failure event is extracted from the test log data. Based on the test status information and the semantic mapping relationship, association matching is performed, and a structured test knowledge graph is generated using the association matching results.

[0013] One embodiment of this application provides a computer device, including:

[0014] A processor; a storage device having a computer program stored thereon; a network interface for providing network communication functions; when the computer program is executed by the processor, the processor enables the processor to implement any of the aforementioned artificial intelligence-based Flash particle test data processing methods.

[0015] One embodiment of this application provides a readable storage medium on which a program or instruction is stored. When the program or instruction is executed by a processor, it implements the steps of the artificial intelligence-based Flash particle test data processing method.

[0016] This application embodiment collects particle response behavior data generated during the execution of Flash storage particle testing tasks and extracts temporal correlation logic information. Combined with a behavior sequence analysis algorithm adapted to the testing task, a dynamic reasoning model based on behavior sequences is created. This achieves in-depth mining of the evolutionary patterns of particle response behavior under different instruction requests and accurate generation of hidden evolutionary trajectories. Furthermore, by inferring the degradation trend prediction results of particle test failure events through particle performance evolution characteristics in the hidden evolutionary trajectories, the trend features are transformed into semantic knowledge nodes and a semantic mapping relationship with failure types is established. Finally, a structured test knowledge graph is generated by combining test log data. This design breaks through the limitations of single-feature judgment in traditional Flash particle testing. It constructs a complete technical link from data collection and evolutionary reasoning to knowledge graph construction, with temporal correlation logic as the core. This enables early positioning of Flash particle stability degradation trends, providing a new qualitative judgment dimension for lifetime prediction. Simultaneously, it transforms scattered qualitative test data into structured knowledge that can directly guide product iteration, breaking down the information barrier between test data and design optimization, and improving the efficiency and accuracy of the transformation from test data to decision-making. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating an artificial intelligence-based Flash particle test data processing method provided in an embodiment of this application.

[0019] Figure 2 This is a schematic diagram of the basic structure of a computer device provided in an embodiment of this application.

[0020] Figure 3 This is a functional block diagram of a Flash particle test data processing device provided in an embodiment of this application. Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0022] Please see Figure 1 , Figure 1 This is a flowchart of a Flash particle test data processing method based on artificial intelligence provided in an embodiment of this application. The method can be executed by a computer device or by a computer device and a server. The method may include steps 110-160.

[0023] Step 110: Collect particle response behavior data generated during the execution of the Flash storage particle test task, and extract the temporal correlation logic information of the particle response behavior based on the particle response behavior data.

[0024] In this embodiment, the computer device collects particle response behavior data generated during the execution of the Flash memory particle testing task. This data originates from multiple sensors set up in the Flash memory particle testing task, including voltage sensors, current sensors, and temperature sensors. Each sensor collects data at fixed time intervals, forming particle response behavior data containing timestamps, sensor identifiers, and collected values. For example, the voltage sensor collects the operating voltage of the Flash memory particle at fixed time intervals, the current sensor collects the operating current of the Flash memory particle at fixed time intervals, and the temperature sensor collects the surface temperature of the Flash memory particle at fixed time intervals. Next, the collected particle response behavior data is preprocessed to remove outliers and missing values. Then, a time series correlation analysis algorithm is used to analyze the preprocessed data to extract the temporal correlation logic information of the particle response behavior. This temporal correlation logic information includes the chronological order of data collected by different sensors and the correlation of data changes. For example, when the voltage value collected by the voltage sensor changes, the current value collected by the current sensor will change accordingly after a certain period of time. This chronological order and the correlation of data changes are extracted as the temporal correlation logic information.

[0025] Step 120: Based on the behavior sequence analysis algorithm adapted to the Flash storage particle test task, and the particle response behavior characteristics of different test stages in the time-series correlation logic information, create a dynamic reasoning model based on behavior sequence.

[0026] In this embodiment, the computer device determines a behavioral sequence analysis algorithm adapted to the Flash storage particle testing task. This algorithm is a Long Short-Term Memory (LSTM) network algorithm, which can effectively process time-series data. Based on the particle response behavior characteristics of different testing stages in the time-series correlation logic information, the input and output of the dynamic inference model are determined. For example, in the initial stage of the Flash storage particle testing task, the particle response behavior characteristics are mainly characterized by stable changes in voltage and current; in the middle stage of the testing task, the particle response behavior characteristics are mainly characterized by fluctuating changes in voltage and current; and in the final stage of the testing task, the particle response behavior characteristics are mainly characterized by decreasing changes in voltage and current. The computer device uses these particle response behavior characteristics of different testing stages as input features of the dynamic inference model and uses the probability of particle test failure events as the output of the dynamic inference model. Next, the network structure of the dynamic inference model is constructed. This model includes an input layer, a hidden layer, and an output layer. The number of neurons in the input layer is the same as the number of input features. The hidden layer adopts a LSM network structure, and the output layer adopts a fully connected layer structure. The dynamic inference model is then trained using particle response behavior data and corresponding particle test failure event data collected from historical flash storage particle testing tasks. The training process employs a stochastic gradient descent algorithm to adjust the model's weights and biases, minimizing the error between the model's output and the actual probability of particle test failure events. After training, the computer saves the parameters of the dynamic inference model, completing the creation of the dynamic inference model.

[0027] Step 130: Input the particle response behavior data corresponding to different instruction requests in the Flash storage particle test task into the dynamic inference model for time series analysis, explore the evolution law of particle response behavior under different instruction requests, and generate hidden evolution trajectories corresponding to different instruction requests.

[0028] Step 131: Determine the interval parameters for dividing the time window based on the time sequence association logic information, and divide the particle response behavior data corresponding to different instruction requests in the Flash storage particle test task into time windows according to the interval parameters to obtain several data windows that correspond one-to-one with the instruction requests. Each data window contains particle response behavior data within a continuous time period during the execution of the corresponding instruction request.

[0029] In this embodiment, the computer device determines the interval parameters for dividing the time window based on timing association logic information. These interval parameters include the length of the time window and the sliding step size. For example, based on the frequency of change of granular response behavior data in the timing association logic information, the length of the time window is determined to be a fixed time length, and the sliding step size is also a fixed time step size. Next, the granular response behavior data corresponding to different instruction requests is divided into time windows according to the determined interval parameters. For example, for the granular response behavior data corresponding to instruction request A, starting from the execution start time of instruction request A, several data windows are sequentially divided with the time window length as the interval and the sliding step size as the moving distance. Each data window contains granular response behavior data within the corresponding time period. Similarly, the granular response behavior data corresponding to instruction request B is divided into time windows using the same method to obtain several data windows. Each data window corresponds one-to-one with an instruction request and contains granular response behavior data within a continuous time period during the execution of that instruction request.

[0030] Step 132: Use a latent variable model to extract features from the particle response behavior data within each data window, and generate a hidden state sequence corresponding to each data window. The hidden state sequence is used to characterize the intrinsic change features of the particle response behavior within the corresponding data window.

[0031] In this embodiment, the computer device employs a latent variable model (HMM) to extract features from the particle response behavior data within each data window. This HMM is capable of extracting latent state sequences from observed data. First, the computer device standardizes the particle response behavior data within each data window, converting it into standardized data with a mean of 0 and a variance of 1. Then, the standardized data is input into the HMM. The model learns the probability distribution of the data and extracts the latent state sequence corresponding to each data window. For example, for a particle response behavior data window containing voltage, current, and temperature values, the HMM learns the changing patterns of these data and extracts a sequence containing multiple latent states. Each latent state corresponds to an intrinsic change characteristic of the particle response behavior, such as a stable state, a fluctuating state, or a decreasing state. The generated latent state sequence is used to characterize the intrinsic change characteristics of the particle response behavior within the corresponding data window.

[0032] Step 133: Compare the behavior patterns of the hidden state sequences corresponding to different instruction requests, align the hidden state sequences with similar change features during the comparison, calculate the similarity parameter between the aligned hidden state sequences, cluster all hidden state sequences based on the similarity parameter, and perform time-series segmentation on the clustered hidden state sequences according to the chronological order to obtain several time-series segments.

[0033] In this embodiment, the computer device compares the behavioral patterns of hidden state sequences corresponding to different instruction requests. A dynamic time warping algorithm is used to calculate the distance between different hidden state sequences; the smaller the distance, the more similar the behavioral patterns of the hidden state sequences. Then, hidden state sequences with similar changing characteristics are aligned during the comparison process, adjusting sequences of different lengths to sequences of the same length. Next, a similarity parameter is calculated between the aligned hidden state sequences using a cosine similarity algorithm to obtain the similarity value between each hidden state sequence and other hidden state sequences. Then, based on the similarity parameter, the computer device performs event sequence clustering on all hidden state sequences, using a hierarchical clustering algorithm to group hidden state sequences with high similarity into one class. Finally, the computer device performs temporal segmentation on the clustered hidden state sequences according to their chronological order, obtaining several temporal segments, each containing clustered hidden state sequences within a certain time range.

[0034] Step 134: Extract the state transition probability of the hidden state sequence within each time segment. The state transition probability characterizes the evolutionary stability of the particle response behavior within the corresponding time segment. Input the state transition probability into the dynamic inference model. The dynamic inference model integrates the evolutionary features of each time segment with the time series analysis capability, and mines the evolutionary rules of particle response behavior under different command requests. Based on the evolutionary rules, the evolutionary path of the hidden evolutionary trajectory corresponding to different command requests is extracted, and a hidden evolutionary trajectory corresponding to each command request is generated. The hidden evolutionary trajectory is used to present the dynamic change process of particle response behavior under the corresponding command request.

[0035] In this embodiment, the computer device extracts the state transition probabilities of the hidden state sequence within each time-series segment, and uses a state transition matrix algorithm to calculate the transition probabilities between different hidden states in each hidden state sequence, thus obtaining a state transition matrix. Then, the state transition probabilities characterize the evolutionary stability of the granular response behavior within the corresponding time-series segment; a higher state transition probability indicates a more stable evolution of the granular response behavior, while a lower state transition probability indicates a less stable evolution. Next, the state transition probabilities are input into a dynamic inference model. The model integrates the evolutionary features of each time-series segment using its time-series analysis capabilities to uncover the evolutionary patterns of granular response behavior under different instruction requests. For example, by analyzing the state transition probabilities within different time-series segments, the model finds that when the state transition probability is high, the evolution of the granular response behavior is more stable, while when the state transition probability is low, the evolution of the granular response behavior is less stable. Finally, the computer device extracts the evolution path of the hidden evolution trajectory corresponding to different instruction requests based on the evolution law, and generates a hidden evolution trajectory that corresponds to each instruction request. The trajectory uses time as the horizontal axis and the evolution state of particle response behavior as the vertical axis, which intuitively shows the dynamic change process of particle response behavior under the corresponding instruction request.

[0036] Step 140: By analyzing the particle performance evolution characteristics contained in the hidden evolution trajectory, the occurrence trend and impact range of particle test failure events are deduced to obtain the degradation trend prediction results associated with the particle test failure events.

[0037] Step 141: Extract all evolutionary features related to particle performance from the hidden evolutionary trajectory, classify and organize the extracted evolutionary features according to performance type, and construct a corresponding performance index sequence. The performance index sequence is used to characterize the dynamic changes in particle performance during the execution of the Flash storage particle test task.

[0038] In this embodiment, the computer device extracts all evolutionary features related to particle performance from the hidden evolutionary trajectory. These features include voltage evolution features, current evolution features, and temperature evolution features. For example, voltage evolution features include voltage change trends, change rates, and fluctuation amplitudes; current evolution features include current change trends, change rates, and fluctuation amplitudes; and temperature evolution features include temperature change trends, change rates, and fluctuation amplitudes. Next, the extracted evolutionary features are categorized and organized according to performance type: voltage evolution features are classified as voltage performance type, current evolution features as current performance type, and temperature evolution features as temperature performance type. Then, corresponding performance index sequences are constructed, each corresponding to a performance type. For example, the voltage performance index sequence includes information such as voltage change trends, change rates, and fluctuation amplitudes at different time points; the current performance index sequence includes information such as current change trends, change rates, and fluctuation amplitudes at different time points; and the temperature performance index sequence includes information such as temperature change trends, change rates, and fluctuation amplitudes at different time points. The constructed performance index sequences are used to characterize the dynamic changes in particle performance during the execution of the Flash storage particle testing task.

[0039] Step 142: Use a time series prediction algorithm to perform trend analysis on the preprocessed performance index sequence, predict the change trend of particle performance within the target time period, and obtain the performance index prediction trend data.

[0040] In this embodiment, the computer device preprocesses the performance index sequence to remove outliers and missing values. Then, a time series prediction algorithm is used to perform trend analysis on the preprocessed performance index sequence. This time series prediction algorithm is an autoregressive integral moving average model, capable of predicting time series data. The computer device inputs the preprocessed performance index sequence into the autoregressive integral moving average model. By learning the historical variation patterns of the data, the model predicts the changing trend of particle performance within a target time period. For example, the model predicts that within a future time period, the voltage performance index sequence, the current performance index sequence, and the temperature performance index sequence will continue to rise. The obtained performance index prediction trend data includes the predicted values ​​and prediction intervals for each performance index within the target time period.

[0041] Step 143: Call the causal inference model, input the performance index prediction trend data and the correlation data of particle test failure events into the causal inference model, and let the causal inference model analyze the linkage between particle performance evolution characteristics and particle test failure events, identify the precursor features that lead to the occurrence of particle test failure events, classify and filter the precursor features, and obtain the key precursor features that directly affect the occurrence of particle test failure events.

[0042] In this embodiment, the computer device invokes a causal inference model, which is a Bayesian network model capable of analyzing causal relationships between variables. Then, the correlation data between performance indicator prediction trend data and particle test failure events is input into the causal inference model. The correlation data for particle test failure events includes information such as the occurrence time, cause, and scope of impact of historical particle test failure events. By analyzing the correlation data between performance indicator prediction trend data and particle test failure events, the model identifies precursor features that lead to particle test failure events, such as voltage values ​​exceeding normal ranges, current values ​​exceeding normal ranges, and temperature values ​​exceeding normal ranges. Next, the precursor features are classified and filtered, removing those indirectly affecting particle test failure events to obtain key precursor features that directly influence their occurrence. For example, voltage values ​​exceeding normal ranges, current values ​​exceeding normal ranges, and temperature values ​​exceeding normal ranges are all key precursor features leading to particle test failure events.

[0043] Step 144: Use survival analysis to quantify the key precursor features, combine historical data on particle performance changes to estimate the failure time of particle test failure events, and calculate the probability of particle test failure events occurring in different time periods.

[0044] In this embodiment, the computer device employs a survival analysis method, specifically the Kaplan-Meier method, to quantify key precursor features, enabling estimation of event occurrence time. The computer device uses key precursor features as input variables and the occurrence of particle test failure events as output events. It then uses the Kaplan-Meier method to quantify the key precursor features, obtaining a survival function corresponding to each key precursor feature. Next, it combines historical particle performance data to estimate the failure time of particle test failure events. For example, based on the duration of voltage exceeding the normal range and historical particle test failure event occurrence times, the failure time is estimated to be a future time period. Then, the probability of particle test failure events occurring in different time periods is calculated. For instance, the probability of particle test failure events occurring in the first future time period is calculated as a fixed value, the probability of particle test failure events occurring in the second future time period is calculated as a fixed value, and the probability of particle test failure events occurring in the third future time period is calculated as a fixed value, and so on.

[0045] Step 145: Use a graph neural network to fuse the performance index prediction trend data, causal inference results and failure time estimation data, and quantify the particle performance degradation to obtain a quantified value of the degree of particle performance degradation.

[0046] In this embodiment, the computer device utilizes a graph neural network to fuse performance indicator prediction trend data, causal inference results, and failure time estimation data. This graph neural network is a graph convolutional network, capable of processing graph-structured data. The computer device converts the performance indicator prediction trend data, causal inference results, and failure time estimation data into graph-structured data, where nodes represent different features and edges represent the relationships between features. The graph-structured data is then input into the graph convolutional network, and the model learns the features and relationships within the graph-structured data to fuse the data. Next, a quantization algorithm is used to quantify the granular performance degradation of the fused data, obtaining a quantized value for the degree of granular performance degradation. This value is a number between 0 and 1; a larger value indicates a more severe degree of performance degradation, and a smaller value indicates a less severe degree of performance degradation.

[0047] Step 146: Call the impact diffusion simulation algorithm, combine the degradation degree quantification value to simulate the impact range of the particle test failure event on the surrounding Flash memory particles and the entire Flash memory particle test task, and conduct a risk ripple assessment on the impact range to obtain the risk ripple assessment result.

[0048] In this embodiment, the computer device invokes an impact diffusion simulation algorithm, which is a cellular automaton algorithm capable of simulating the diffusion process of an event. Then, combined with a decay quantification value, the initial state, transition rules, and boundary conditions of the cellular automaton are set to simulate the impact range of a particle testing failure event on surrounding Flash memory particles and the entire Flash memory particle testing task. For example, the layout of the Flash memory particle testing task is divided into multiple cells, each corresponding to one Flash memory particle or one testing area. The initial state is set as follows: the cell where the particle testing failure event occurs is in a failed state, and other cells are in a normal state. The transition rule is set as follows: when a cell's adjacent cells are in a failed state, that cell transitions to a failed state with a certain probability. The boundary condition is set as follows: the boundary cells of the testing task are not affected by external factors. Next, a risk spillover assessment is performed on the simulated impact range. The assessment indicators include the number of affected Flash memory particles, the area of ​​the affected testing area, and the duration of the affected testing task. Finally, the computer device obtains the risk spillover assessment result, which includes information such as the size of the impact range and the risk level.

[0049] Step 147: Combining the performance index prediction trend data, failure time estimation results, degradation degree quantification value and risk ripple assessment results, deduce and analyze the occurrence trend and impact range of particle test failure events, and integrate all derivation and analysis results to obtain the degradation trend prediction results associated with the particle test failure events.

[0050] In this embodiment, the computer device combines performance indicator prediction trend data, failure time estimation results, degradation degree quantification values, and risk ripple assessment results to deduce and analyze the occurrence trend and impact range of particle testing failure events. For example, based on the performance indicator prediction trend data, the occurrence trend of particle testing failure events is predicted to gradually increase; based on the failure time estimation results, the occurrence time of particle testing failure events is predicted to be some time in the future; based on the degradation degree quantification value, the degradation degree of particle performance is predicted to be severe; and based on the risk ripple assessment results, the impact range of particle testing failure events is predicted to be large. Then, all the derivation and analysis results are integrated to obtain the degradation trend prediction results associated with particle testing failure events. This result includes information such as the occurrence time, probability of occurrence, impact range, and degradation degree of particle testing failure events.

[0051] Step 150: Convert each trend feature in the decline trend prediction result into a corresponding semantic knowledge node, and establish a semantic mapping relationship between trend features and failure types by combining each semantic knowledge node and particle test failure event.

[0052] Step 151: Perform attribute vector decomposition on each trend feature in the recession trend prediction result, decomposing each trend feature into multiple attribute vectors of different dimensions, with each attribute vector corresponding to a target attribute of the trend feature.

[0053] In this embodiment, the computer device analyzes the trend features in the degradation trend prediction results and determines the target attributes of each trend feature. For example, the target attributes of the voltage rise trend feature include voltage change trend, voltage change rate, and voltage fluctuation amplitude; the target attributes of the current rise trend feature include current change trend, current change rate, and current fluctuation amplitude; and the target attributes of the temperature rise trend feature include temperature change trend, temperature change rate, and temperature fluctuation amplitude. Next, an attribute vector decomposition algorithm is used to decompose each trend feature into multiple attribute vectors of different dimensions, each corresponding to a target attribute of the trend feature. For example, the voltage rise trend feature is decomposed into a voltage change trend attribute vector, a voltage change rate attribute vector, and a voltage fluctuation amplitude attribute vector; the current rise trend feature is decomposed into a current change trend attribute vector, a current change rate attribute vector, and a current fluctuation amplitude attribute vector; and the temperature rise trend feature is decomposed into a temperature change trend attribute vector, a temperature change rate attribute vector, and a temperature fluctuation amplitude attribute vector.

[0054] Step 152: Use a knowledge representation learning algorithm to perform semantic space embedding on the decomposed attribute vectors, map each attribute vector to a preset semantic space, and generate initial node embeddings corresponding to each trend feature. The initial node embeddings are used to represent the position and semantic features of the trend features in the semantic space.

[0055] In this embodiment, the computer device employs a knowledge representation learning algorithm, specifically the TransE algorithm, to perform semantic space embedding on the decomposed attribute vectors. This algorithm maps entities and relationships to a low-dimensional vector space. Each attribute vector is then treated as an entity, and the relationships between attribute vectors are treated as relations. The TransE algorithm is used to perform semantic space embedding on the attribute vectors, generating initial node embeddings corresponding to each trend feature. For example, the voltage change trend attribute vector is treated as an entity, and the relationship between the voltage change trend attribute vector and the current change trend attribute vector is treated as a relation. The TransE algorithm is used to map these entities and relations to a preset semantic space, generating initial node embeddings corresponding to the voltage increase trend feature, the current increase trend feature, and the temperature increase trend feature, etc. The generated initial node embeddings are used to represent the position and semantic features of the trend features in the semantic space; the closer the positions, the more similar the semantics of the trend features.

[0056] Step 153: Optimize all initial node embeddings through graph attention mechanism to strengthen the semantic association between the initial node embeddings corresponding to each trend feature. Combine semantic association learning to extract the embedding paths between the initial node embeddings. Construct an entity relationship graph based on the embedding paths. The entities in the entity relationship graph are the initial node embeddings corresponding to each trend feature, and the relationships are the semantic associations between the node embeddings.

[0057] In this embodiment, the computer device optimizes all initial node embeddings using a graph attention mechanism, which assigns different attention weights based on the semantic association between nodes. Then, the attention weights between each initial node embedding and other initial node embeddings are calculated, and the initial node embeddings are weighted and summed according to these attention weights to obtain the optimized node embeddings. Next, semantic association learning is used to extract embedding paths between the initial node embeddings. These embedding paths are the paths formed by semantic associations between the initial node embeddings. For example, the initial node embedding corresponding to the voltage rising trend feature and the initial node embedding corresponding to the current rising trend feature are connected by the semantic association "voltage change leads to current change," forming an embedding path. Similarly, the initial node embedding corresponding to the current rising trend feature and the initial node embedding corresponding to the temperature rising trend feature are connected by the semantic association "current change leads to temperature change," forming an embedding path. Finally, the computer device constructs an entity relationship graph based on the embedding paths. The entities in the entity relationship graph are the initial node embeddings corresponding to each trend feature, and the relationships are the semantic associations between the node embeddings.

[0058] Step 154: Establish a mapping rule base, which contains association rules between trend features and failure types.

[0059] In this embodiment, the computer device establishes a mapping rule base, which contains association rules between trend characteristics and failure types. For example, association rules such as "voltage rising trend corresponds to overvoltage failure type," "current rising trend corresponds to overcurrent failure type," and "temperature rising trend corresponds to overtemperature failure type" are included. These association rules are derived from experience and knowledge accumulated in historical Flash memory chip testing tasks. The computer device establishes these association rules by analyzing and summarizing historical data and stores them in the mapping rule base.

[0060] Step 155: Based on the mapping rule base, determine the correspondence between each trend feature and failure type, and use the feature dimension mapping algorithm to map the attribute vector of each trend feature to the corresponding failure type feature space, thereby completing the conversion of each trend feature to semantic knowledge node, so that each trend feature corresponds to a unique semantic knowledge node.

[0061] In this embodiment, the computer device determines the correspondence between each trend feature and failure type based on a mapping rule base. For example, according to the rule "voltage rise trend feature corresponds to overvoltage failure type" in the mapping rule base, the correspondence between voltage rise trend feature and overvoltage failure type is determined; according to the rule "current rise trend feature corresponds to overcurrent failure type" in the mapping rule base, the correspondence between current rise trend feature and overcurrent failure type is determined; and according to the rule "temperature rise trend feature corresponds to overtemperature failure type" in the mapping rule base, the correspondence between temperature rise trend feature and overtemperature failure type is determined. Next, a feature dimension mapping algorithm is used to map the attribute vectors of each trend feature to the corresponding failure type feature space. This algorithm is a principal component analysis algorithm, capable of mapping high-dimensional feature vectors to low-dimensional feature spaces. For example, the attribute vector of voltage rise trend feature is mapped to the overvoltage failure type feature space, the attribute vector of current rise trend feature is mapped to the overcurrent failure type feature space, and the attribute vector of temperature rise trend feature is mapped to the overtemperature failure type feature space. Finally, the computer device completes the conversion of each trend feature into a semantic knowledge node, so that each trend feature corresponds to a unique semantic knowledge node. The semantic knowledge node contains information such as the name, description, and corresponding failure type of the trend feature.

[0062] Step 156: Associate the transformed semantic knowledge nodes with the entities in the entity relationship graph, and establish a semantic mapping relationship between trend features and failure types by combining the embedded paths in the entity relationship graph. The semantic mapping relationship is used to characterize the corresponding association between each trend feature and different failure types.

[0063] Step 1561: Based on the correspondence between entities in the entity relationship graph and the semantic knowledge nodes, construct an auxiliary association graph between the semantic knowledge nodes and entities in the entity relationship graph.

[0064] In this embodiment, the computer device analyzes the correspondence between entities and semantic knowledge nodes in the entity relationship graph. For example, the initial node embedding corresponding to the voltage rise trend feature in the entity relationship graph corresponds to the semantic knowledge node corresponding to the voltage rise trend feature; the initial node embedding corresponding to the current rise trend feature in the entity relationship graph corresponds to the semantic knowledge node corresponding to the current rise trend feature; and the initial node embedding corresponding to the temperature rise trend feature in the entity relationship graph corresponds to the semantic knowledge node corresponding to the temperature rise trend feature. Next, based on these correspondences, an auxiliary association graph of semantic knowledge nodes and entities in the entity relationship graph is constructed. The nodes in the auxiliary association graph include both semantic knowledge nodes and entities in the entity relationship graph, and the edges represent the correspondence between semantic knowledge nodes and entities.

[0065] Step 1562: Calculate the semantic connectivity between the semantic knowledge node and the entity in the entity relationship graph according to the node connection order and path hop count indicated by the embedded path, and assign an initial association weight to each association edge in the auxiliary association graph based on the semantic connectivity.

[0066] In this embodiment, the computer device calculates the semantic connectivity between semantic knowledge nodes and entities in the entity relationship graph based on the node connection order and path hop count indicated by the embedding path. For example, the node connection order of the embedding path is: initial node embedding corresponding to the voltage rising trend feature -- initial node embedding corresponding to the current rising trend feature -- initial node embedding corresponding to the temperature rising trend feature, and the path hop count is 2. The computer device uses a path connectivity algorithm to calculate the semantic connectivity between semantic knowledge nodes and entities in the entity relationship graph. Semantic connectivity is inversely proportional to the path hop count; the smaller the path hop count, the higher the semantic connectivity; the larger the path hop count, the lower the semantic connectivity. Then, based on the semantic connectivity, an initial association weight is assigned to each association edge in the auxiliary association graph; the higher the semantic connectivity, the larger the initial association weight; the lower the semantic connectivity, the smaller the initial association weight.

[0067] Step 1563: Call the node representation learning algorithm to vectorize all nodes in the auxiliary association graph, generate the semantic knowledge node and the node representation vector of the entity in the entity relationship graph, and use the node representation vector to calculate the representation similarity between the semantic knowledge node and the entity in the entity relationship graph.

[0068] In this embodiment, the computer device invokes a node representation learning algorithm to vectorize all nodes in the auxiliary association graph. This algorithm is the DeepWalk algorithm, which can convert nodes in the graph into low-dimensional vectors. Then, using the auxiliary association graph as input, the DeepWalk algorithm performs a random walk on all nodes in the auxiliary association graph to generate a node sequence. The Word2Vec algorithm is then used to train the node sequence, generating semantic knowledge nodes and node representation vectors for entities in the entity relationship graph. Next, the representation similarity between the semantic knowledge nodes and entities in the entity relationship graph is calculated using the node representation vectors. This similarity is calculated using the cosine similarity algorithm to obtain the representation similarity value between each semantic knowledge node and each entity.

[0069] Step 1564: Dynamically adjust the initial association weights based on the representation similarity to obtain an updated association weight set. The updated association weight set is used to quantify the degree of association between the semantic knowledge nodes and entities in the entity relationship graph.

[0070] In this embodiment, the computer device dynamically adjusts the initial association weights based on representation similarity. The adjustment rule is that the higher the representation similarity, the greater the increase in the initial association weight; and the lower the representation similarity, the greater the decrease in the initial association weight. For example, if the initial association weight between a semantic knowledge node and an entity is a fixed value, and the representation similarity is also a fixed value, the initial association weight is multiplied by a coefficient related to representation similarity according to the adjustment rule to obtain the updated association weight. Next, an updated set of association weights is obtained. This set contains the updated association weights between all semantic knowledge nodes and entities in the entity relationship graph. This set is used to quantify the degree of association between semantic knowledge nodes and entities in the entity relationship graph; a larger association weight indicates a higher degree of association, and a smaller association weight indicates a lower degree of association.

[0071] Step 1565: Optimize the topology of the auxiliary association graph according to the updated association weight set, remove the association edges corresponding to association weights below a preset threshold in the updated association weight set, and retain the association edges corresponding to association weights above a preset threshold in the updated association weight set, to obtain the optimized auxiliary association graph.

[0072] In this embodiment, the computer device optimizes the topology of the auxiliary association graph based on the updated association weight set, with a preset threshold set as a fixed value, for example, 0.5. Then, it removes association edges corresponding to association weights below the preset threshold from the updated association weight set, while retaining association edges corresponding to association weights above the preset threshold. For example, an association edge with a weight of 0.4 is removed, while an association edge with a weight of 0.6 is retained. Finally, the computer device obtains an optimized auxiliary association graph where all association edges have association weights above the preset threshold, resulting in a simpler and more accurate topology.

[0073] Step 1566: Based on the association edges retained by the optimized auxiliary association graph, determine the stable node connection relationship between the semantic knowledge node and the entity in the entity relationship graph.

[0074] In this embodiment, the computer device analyzes the retained association edges in the optimized auxiliary association graph. These association edges connect semantic knowledge nodes and entities in the entity relationship graph with a high degree of affinity. Then, based on these association edges, stable node connection relationships between semantic knowledge nodes and entities in the entity relationship graph are determined. For example, there is a stable node connection relationship between the semantic knowledge node "voltage rise trend" and the entity "initial node embedding corresponding to voltage rise trend feature" in the entity relationship graph; there is a stable node connection relationship between the semantic knowledge node "current rise trend" and the entity "initial node embedding corresponding to current rise trend feature" in the entity relationship graph; and there is a stable node connection relationship between the semantic knowledge node "temperature rise trend" and the entity "initial node embedding corresponding to temperature rise trend feature" in the entity relationship graph.

[0075] Step 1567: Calculate the semantic relevance of the stable node connection relationship through a bidirectional attention mechanism, extract the forward attention score of the semantic knowledge node pointing to the entity in the entity relationship graph and the backward attention score of the entity in the entity relationship graph pointing to the semantic knowledge node, and perform weighted fusion of the forward attention score and the backward attention score to obtain the comprehensive semantic relevance.

[0076] In this embodiment, the computer device calculates the semantic relevance of stable node connections using a bidirectional attention mechanism. This mechanism can simultaneously consider both forward and backward attention between nodes. Then, it extracts the forward attention score from semantic knowledge nodes pointing to entities in the entity relationship graph, and the backward attention score from entities in the entity relationship graph pointing to semantic knowledge nodes. Both forward and backward attention scores are obtained using an attention weighting method. Next, the forward and backward attention scores are weighted and fused, with fixed weighting coefficients, for example, 0.5 for both forward and backward attention scores, to obtain the comprehensive semantic relevance.

[0077] Step 1568: Couple the comprehensive semantic association degree with the path weight in the embedded path to generate the fused semantic association value between the semantic knowledge node and the entity in the entity relationship graph.

[0078] In this embodiment, the computer device couples the overall semantic relevance with the path weights in the embedded path. The path weights are the sum of the weights of each edge in the embedded path. For example, the path weights of the embedded path are fixed values, and the overall semantic relevance is also fixed. Multiplying the overall semantic relevance by the path weights yields the fused semantic relevance value. The generated fused semantic relevance value is used to quantify the degree of fused semantic relevance between semantic knowledge nodes and entities in the entity relationship graph. A larger fused semantic relevance value indicates a higher degree of fused semantic relevance; a smaller fused semantic relevance value indicates a lower degree of fused semantic relevance.

[0079] Step 1569: Using the fused semantic association value as a bridge, the trend features represented by the semantic knowledge nodes are linked semantically to the failure types associated with the entities in the entity relationship graph, thereby completing the construction and solidification of the semantic mapping relationship between trend features and failure types.

[0080] In this embodiment, the computer device uses a fused semantic association value as a bridge to semantically link the trend features represented by semantic knowledge nodes with the failure types associated with entities in the entity relationship graph. For example, the trend feature represented by the semantic knowledge node "voltage rise trend" is a voltage rise trend, and the failure type associated with the entity "initial node embedding corresponding to the voltage rise trend feature" in the entity relationship graph is an overvoltage failure type. The fused semantic association value is a fixed value. When the fused semantic association value is higher than a preset threshold, the voltage rise trend feature and the overvoltage failure type are semantically linked. Next, the semantic mapping relationship between trend features and failure types is constructed and solidified, and the constructed semantic mapping relationship is stored in the database.

[0081] Step 160: Call the test log data generated during the execution of the particle test task stored in the Flash, extract the test status information related to the particle response behavior and the particle test failure event from the test log data, perform association matching based on the test status information and the semantic mapping relationship, and generate a structured test knowledge graph using the association matching results.

[0082] Step 161: Organize the test status information related to the particle response behavior and the particle test failure event extracted from the test log data, sort the organized test status information according to the chronological order of the events, and construct a structured event stream. Each event node in the structured event stream corresponds to a test status information, and each event node contains the event occurrence time, event type, and related particle response behavior data.

[0083] In this embodiment, the computer device organizes the test status information related to particle response behavior and particle test failure events extracted from the test log data, removing duplicate and irrelevant information. Then, the organized test status information is sorted according to the chronological order of the events. For example, voltage test events, current test events, temperature test events, and particle test failure events in the test status information are sorted according to their occurrence time. Next, a structured event stream is constructed. Each event node in the structured event stream corresponds to a piece of test status information, and each event node contains the event occurrence time, event type, and related particle response behavior data. For example, the event node "Voltage Test Event" includes an event occurrence time of a certain period, an event type of voltage test, and related particle response behavior data of voltage value; the event node "Current Test Event" includes an event occurrence time of a certain period, an event type of current test, and related particle response behavior data of current value; the event node "Temperature Test Event" includes an event occurrence time of a certain period, an event type of temperature test, and related particle response behavior data of temperature value; the event node "Particle Test Failure Event" includes an event occurrence time of a certain period, an event type of particle test failure, and related particle response behavior data such as voltage value, current value, and temperature value.

[0084] Step 162: Use a graph matching algorithm to associate and match event nodes in the structured event stream with semantic knowledge nodes in the semantic mapping relationship. Introduce a similarity measurement algorithm to calculate the similarity between each event node and each semantic knowledge node. Based on the similarity, determine the degree of association between event nodes and semantic knowledge nodes. Align event nodes with an association degree that reaches a preset threshold with semantic knowledge nodes to complete the initial association matching between event nodes and semantic knowledge nodes.

[0085] In this embodiment, the computer device employs a graph matching algorithm to associate and match event nodes in the structured event stream with semantic knowledge nodes in the semantic mapping relationship. This algorithm is the Hungarian algorithm, which can find the optimal match between two graphs. Then, a similarity measurement algorithm is introduced to calculate the similarity between each event node and each semantic knowledge node. This algorithm is the cosine similarity algorithm, which calculates the similarity between the feature vectors of the event node and the feature vectors of the semantic knowledge node. For example, the feature vector of the event node "voltage test event" includes the event occurrence time, event type, and related particle response behavior data, while the feature vector of the semantic knowledge node "voltage upward trend" includes the name, description, and corresponding failure type of the trend feature. The similarity between these two feature vectors is calculated. Next, the degree of association between the event node and the semantic knowledge node is determined based on the similarity; the higher the similarity, the higher the degree of association; the lower the similarity, the lower the degree of association. A preset threshold is a fixed value, for example, 0.6. Then, event nodes whose correlation reaches a preset threshold are aligned with semantic knowledge nodes. For example, the event node "voltage test event" is aligned with the semantic knowledge node "voltage rise trend," the event node "current test event" with the semantic knowledge node "current rise trend," the event node "temperature test event" with the semantic knowledge node "temperature rise trend," and the event node "particle test failure event" with the semantic knowledge nodes "overvoltage failure type," "overcurrent failure type," and "overtemperature failure type." Finally, the computer device completes the initial correlation matching between event nodes and semantic knowledge nodes.

[0086] Step 163: Merge the event relationships in the structured event stream with the embedded paths in the semantic mapping relationship, merge duplicate relationships in the fusion process to achieve relationship fusion, and obtain the fused set of nodes and relationships.

[0087] Step 1631: Analyze the event relationships formed between adjacent event nodes in the structured event flow, extract the event relationship type and event relationship strength of the event relationship, and perform vectorization conversion on the event relationship type and event relationship strength to generate the event relationship vector corresponding to each event relationship in the structured event flow.

[0088] In this embodiment, the computer device parses the event relationships formed between adjacent event nodes in the structured event stream. For example, the event relationship between adjacent event nodes "voltage test event" and "current test event" is "voltage change leads to current change," the event relationship between adjacent event nodes "current test event" and "temperature test event" is "current change leads to temperature change," and the event relationship between adjacent event nodes "temperature test event" and "particle test failure event" is "temperature change leads to particle test failure." Next, the event relationship type and event relationship strength are extracted. The event relationship type is the name of the event relationship, such as "voltage change leads to current change," "current change leads to temperature change," and "temperature change leads to particle test failure." The event relationship strength is the degree of strength of the event relationship; for example, the event relationship strength of "voltage change leads to current change" is a fixed value, the event relationship strength of "current change leads to temperature change" is a fixed value, and the event relationship strength of "temperature change leads to particle test failure" is a fixed value. Then, the event relationship type and event relationship strength are vectorized. The one-hot encoding algorithm is used to convert the event relationship type into a vector, and the normalization algorithm is used to convert the event relationship strength into a vector. Then, these two vectors are concatenated to generate the event relationship vector corresponding to each event relationship in the structured event flow.

[0089] Step 1632: Parse the embedded paths in the semantic mapping relationship, extract the path relationship type and path relationship strength between the semantic knowledge nodes connected by the embedded paths, and perform vectorization conversion on the path relationship type and the path relationship strength to generate the path relationship vector corresponding to each embedded path in the semantic mapping relationship.

[0090] In this embodiment, the computer device parses the embedding paths in the semantic mapping relationship, such as the embedding paths "initial node embedding corresponding to voltage rise trend feature -- initial node embedding corresponding to current rise trend feature", "initial node embedding corresponding to current rise trend feature -- initial node embedding corresponding to temperature rise trend feature", "initial node embedding corresponding to temperature rise trend feature -- overvoltage failure type node", "initial node embedding corresponding to temperature rise trend feature -- overcurrent failure type node", "initial node embedding corresponding to temperature rise trend feature -- overtemperature failure type node", etc. Next, the path relationship types and strengths between the semantic knowledge nodes connected by the embedded path are extracted. The path relationship type is the name of the relationship between the semantic knowledge nodes connected by the embedded path, such as "voltage change leads to current change", "current change leads to temperature change", "temperature change leads to voltage overload failure", "temperature change leads to current overload failure", "temperature change leads to temperature overload failure", etc. The path relationship strength is the degree of strength of the relationship between the semantic knowledge nodes connected by the embedded path. For example, the path relationship strength of "voltage change leads to current change" is a fixed value, the path relationship strength of "current change leads to temperature change" is a fixed value, and the path relationship strength of "temperature change leads to voltage overload failure" is a fixed value, etc. Then, the path relationship types and strengths are vectorized. A one-hot encoding algorithm is used to convert the path relationship type into a vector, and a normalization algorithm is used to convert the path relationship strength into a vector. These two vectors are then concatenated to generate the path relationship vector corresponding to each embedded path in the semantic mapping relationship.

[0091] Step 1633: Use a relation alignment algorithm to calculate the similarity between the event relation vector and the path relation vector, measure the vector distance between the event relation vector and the path relation vector in the relation type dimension and the relation strength dimension, and determine the semantic similarity between the event relation and the embedded path based on the vector distance.

[0092] In this embodiment, the computer device uses a relation alignment algorithm to calculate the similarity between event relation vectors and path relation vectors. This algorithm is a cosine similarity algorithm, capable of calculating the similarity between two vectors. Then, the vector distance between the event relation vector and the path relation vector is measured in both the relation type and relation strength dimensions, calculated using the Euclidean distance algorithm. Next, the semantic similarity between the event relation and the embedded path is determined based on the vector distance; a smaller vector distance indicates higher semantic similarity, and a larger vector distance indicates lower semantic similarity. For example, a vector distance of 0.2 between the event relation vector and the path relation vector indicates a semantic similarity of 0.8; a vector distance of 0.5 between the event relation vector and the path relation vector indicates a semantic similarity of 0.5.

[0093] Step 1634: Perform relation matching between the event relation and the embedding path based on the semantic similarity, mark the event relation and the embedding path that reach the preset alignment threshold as relation pairs to be fused, compare the relation attributes of the event relation and the embedding path in the relation pairs to be fused, and identify the overlapping and non-overlapping parts of the relation attributes between the event relation and the embedding path in the relation pairs to be fused.

[0094] In this embodiment, the computer device performs relation matching between event relationships and embedding paths based on semantic similarity, with a preset alignment threshold set to a fixed value, for example, 0.7. Then, event relationships and embedding paths whose semantic similarity reaches the preset alignment threshold are marked as relation pairs to be fused. For example, the semantic similarity between the event relationship "voltage change leads to current change" and the embedding path "initial node embedding corresponding to voltage upward trend feature -- initial node embedding corresponding to current upward trend feature" is 0.8, reaching the preset alignment threshold, and is thus marked as a relation pair to be fused. Next, the relation attributes of the event relationships and embedding paths in the relation pairs to be fused are compared. Relationship attributes include relation type, relation strength, and relation direction. For example, comparing the relation attributes of the event relationship "voltage change leads to current change" and the embedding path "initial node embedding corresponding to voltage upward trend feature -- initial node embedding corresponding to current upward trend feature," the relation type is "voltage change leads to current change" in both cases, the relation strength is a fixed value in both cases, and the relation direction is from voltage to current in both cases. The overlapping parts of the relation attributes are identified as relation type, relation strength, and relation direction, while the non-overlapping parts are identified as none.

[0095] Step 1635: Perform attribute merging and deduplication processing on the overlapping relation attributes to generate a fused relation attribute set. The fused relation attribute set inherits the relation attributes of the event relationship and the embedded path in the overlapping part. Perform attribute supplementation and expansion processing on the non-overlapping relation attributes, and add the non-overlapping relation attributes as new attributes to the fused relation attribute set to complete the relation fusion of the relation pairs to be fused, and obtain the fused relation corresponding to the relation pairs to be fused.

[0096] In this embodiment, the computer device performs attribute merging and deduplication on the overlapping relational attributes. For example, it merges the relational type, relational strength, and relational direction of the event relation and the embedded path in the overlapping part, removes duplicate attributes, and generates a fused relational attribute set. The fused relational attribute set inherits the relational attributes of the event relation and the embedded path in the overlapping part. Next, it performs attribute supplementation and expansion on the non-overlapping relational attributes. For example, it adds the relational attributes of the event relation and the embedded path in the non-overlapping part as new attributes to the fused relational attribute set. Finally, the computer device completes the relation fusion of the relation pairs to be fused, obtaining the fused relation corresponding to the relation pairs to be fused. The fused relation contains the fused relational attribute set.

[0097] Step 1636: For event relationships and embedding paths whose semantic similarity does not reach the preset alignment threshold, retain the event relationship and the embedding path as independent relationships in the set of relationships to be merged.

[0098] In this embodiment, the computer device identifies event relationships and embedding paths whose semantic similarity does not reach a preset alignment threshold. For example, the semantic similarity between the event relationship "voltage change leads to current change" and the embedding path "initial node embedding corresponding to current upward trend feature -- initial node embedding corresponding to temperature upward trend feature" is 0.6, which does not reach the preset alignment threshold. These event relationships and embedding paths are then retained as independent relationships in the set of relationships to be fused. The set of relationships to be fused contains event relationships and embedding paths whose semantic similarity does not reach the preset alignment threshold.

[0099] Step 1637: Integrate all the fused relations and all the independent relations to form a fused relation set; extract all nodes connected by the relations in the fused relation set, the nodes including event nodes in the structured event stream and semantic knowledge nodes in the semantic mapping relations, and combine the nodes with the fused relation set to generate the fused node and relation set.

[0100] In this embodiment, the computer device integrates all fused relationships and all independent relationships to form a fused relationship set. Then, it extracts all nodes connected to the relationships in the fused relationship set. These nodes include event nodes from the structured event stream and semantic knowledge nodes from the semantic mapping relationship. For example, the fused relationship "voltage change leads to current change" is connected to the event node "voltage test event" from the structured event stream and the semantic knowledge node "initial node embedding corresponding to current upward trend feature" from the semantic mapping relationship. Finally, the computer device combines the nodes with the fused relationship set to generate a fused node-relationship set.

[0101] Step 164: Construct an initial knowledge graph structure based on the fused set of nodes and relationships, update the topology of the initial knowledge graph structure, and optimize the node connection method of the knowledge graph.

[0102] In this embodiment, the computer device constructs an initial knowledge graph structure based on the fused set of nodes and relationships. The nodes in the initial knowledge graph structure are the fused nodes, and the edges are the fused relationships. Then, the initial knowledge graph structure is updated topologically to optimize the node connection method. For example, the positions of the nodes are adjusted to make the layout of the knowledge graph more reasonable; the length and direction of the edges are adjusted to make the structure of the knowledge graph clearer.

[0103] Step 165: Extract newly added event nodes related to particle test failure events from the test status information, align the newly added event nodes, and integrate the aligned new nodes into the initial knowledge graph structure to realize knowledge graph evolution based on the number of nodes and their relationships.

[0104] In this embodiment, the computer device extracts newly added event nodes related to particle test failure events from the test status information. For example, the newly added event node "Particle Test Failure Event 2" includes an event occurrence time of a certain period, an event type of particle test failure, and related particle response behavior data such as voltage, current, and temperature values. Next, the newly added event nodes are aligned using a node alignment algorithm with nodes in the initial knowledge graph structure. For example, the newly added event node "Particle Test Failure Event 2" is aligned with the node "Particle Test Failure Event" in the initial knowledge graph structure. Then, the aligned new nodes are integrated into the initial knowledge graph structure, realizing knowledge graph evolution based on the number of nodes and their relationships.

[0105] Step 166: During the evolution of the knowledge graph, the association matching results are continuously integrated into the initial knowledge graph structure, and the node attributes and relationships in the initial knowledge graph structure are dynamically updated to generate a structured test knowledge graph.

[0106] Step 1661: Set a graph evolution cycle, monitor the changes in the association matching results in each graph evolution cycle, collect the newly added matching parts and attribute update parts of the association matching results relative to the previous graph evolution cycle, integrate the newly added matching parts and attribute update parts into an incremental matching result set, perform structural analysis on the incremental matching result set, and separate the newly added nodes, node attribute update information and newly added relationships in the incremental matching result set.

[0107] In this embodiment, the computer device sets a graph evolution cycle, for example, a fixed time period. Then, within each graph evolution cycle, it monitors the changes in the association matching results, collecting the newly added matching parts and attribute updates relative to the previous graph evolution cycle. For example, the newly added matching parts are the association matching results between newly added event nodes and semantic knowledge nodes, and the attribute updates are the updated information of node attributes. Next, the newly added matching parts and attribute updates are integrated into an incremental matching result set. The incremental matching result set is then structurally analyzed to separate the newly added nodes, node attribute update information, and newly added relationships. For example, the newly added node is "particle test failure event 3," the node attribute update information is that the influence range of "particle test failure event" has expanded, and the newly added relationship is the association between "particle test failure event 3" and "overheating failure type."

[0108] Step 1662: For the newly added node, calculate the topological connection probability between the newly added node and the existing nodes in the initial knowledge graph structure, insert the newly added node into the corresponding topological position in the initial knowledge graph structure according to the topological connection probability, and establish the initial connection relationship between the newly added node and the existing nodes in the initial knowledge graph structure.

[0109] In this embodiment, for a newly added node, the computer device calculates the probability of topological connections between the new node and existing nodes in the initial knowledge graph structure. A topological connection probability algorithm is used for calculation, and the probability of topological connections is related to factors such as semantic similarity and distance between nodes. For example, the probability of a topological connection between the new node "Particle Test Failure Event 3" and the existing node "Particle Test Failure Event" in the initial knowledge graph structure is a fixed value, as is the probability of a topological connection between the new node and the existing node "Overheat Failure Type". Next, based on the topological connection probability, the new node is inserted into the corresponding topological position in the initial knowledge graph structure. For example, the new node "Particle Test Failure Event 3" is inserted into a position adjacent to both "Particle Test Failure Event" and "Overheat Failure Type". Then, initial connections are established between the new node and existing nodes in the initial knowledge graph structure. For example, connections are established between the new node "Particle Test Failure Event 3" and the existing node "Particle Test Failure Event", and between the new node "Particle Test Failure Event 3" and the existing node "Overheat Failure Type".

[0110] Step 1663: For the node attribute update information, locate the target node in the initial knowledge graph structure pointed to by the node attribute update information, merge the attribute content in the node attribute update information with the existing attributes of the target node, and if there is an attribute conflict between the attribute content in the node attribute update information and the existing attributes of the target node, then select and merge the conflicting attributes according to the attribute timestamp or attribute confidence, and complete the dynamic update of the attributes of the target node.

[0111] In this embodiment, the computer device locates the target node in the initial knowledge graph structure pointed to by the node attribute update information. For example, the target node pointed to by the node attribute update information is "particle testing failure event". Next, the attribute content in the node attribute update information is merged with the existing attributes of the target node. For example, if the attribute content in the node attribute update information is "expanded impact range" and the existing attribute of the target node is "impact range is a fixed value", these two attributes are merged. Then, if there is an attribute conflict between the attribute content in the node attribute update information and the existing attribute of the target node, for example, if the attribute content in the node attribute update information is "impact range is a fixed value 1" and the existing attribute of the target node is "impact range is a fixed value 2", the conflicting attributes are selected and merged based on the attribute timestamp or attribute confidence level. For example, based on the attribute timestamp, the more recent attribute content is retained; based on the attribute confidence level, the attribute content with higher confidence is retained. Finally, the computer device completes the dynamic update of the target node's attributes.

[0112] Step 1664: For the newly added relationship, verify whether the nodes connected by the newly added relationship already exist in the initial knowledge graph structure. If all the nodes connected by the newly added relationship already exist, add the newly added relationship as a new edge to the initial knowledge graph structure. If there are nodes among the nodes connected by the newly added relationship that have not been added to the initial knowledge graph structure, prioritize processing the nodes that have not been added as newly added nodes before adding the newly added relationship.

[0113] In this embodiment, for a newly added relationship, the computer device verifies whether the nodes connected to the newly added relationship already exist in the initial knowledge graph structure. For example, the newly added relationship "Particle Test Failure Event 3" and "Overheat Failure Type" connects the nodes "Particle Test Failure Event 3" and "Overheat Failure Type," where "Particle Test Failure Event 3" is a newly added node, and "Overheat Failure Type" already exists in the initial knowledge graph structure. Next, if all nodes connected to the newly added relationship already exist, the newly added relationship is added to the initial knowledge graph structure as a new edge; if any nodes connected to the newly added relationship are not yet added to the initial knowledge graph structure, these unadded nodes are processed as newly added nodes before the newly added relationship is added. For example, for the newly added relationship "Particle Test Failure Event 3" and "Overheat Failure Type," "Particle Test Failure Event 3" is processed as a newly added node first, and then the newly added relationship is added.

[0114] Step 1665: After completing the insertion of new nodes, merging of node attribute update information, and addition of new relationships, the overall topological connectivity of the initial knowledge graph structure is calibrated, and the presence of isolated nodes or broken relationship paths in the initial knowledge graph structure is detected. The connection relationships between the detected isolated nodes and their neighboring nodes are supplemented, and the broken relationship paths are repaired to maintain the topological connectivity of the initial knowledge graph structure.

[0115] In this embodiment, after inserting new nodes, merging node attribute update information, and adding new relationships, the computer device calibrates the overall topological connectivity of the initial knowledge graph structure. Then, it detects whether there are isolated nodes or broken relationship paths in the initial knowledge graph structure. Isolated nodes are nodes without connections to other nodes, and broken relationship paths are those with missing edges. For example, an isolated node "Particle Test Failure Event 4" is detected, and the broken relationship path is the missing edge "Temperature Test Event - Particle Test Failure Event" in the path "Voltage Test Event - Current Test Event - Temperature Test Event - Particle Test Failure Event". Next, the connected relationships between the detected isolated nodes and their neighboring nodes are supplemented; for example, a connection between the isolated node "Particle Test Failure Event 4" and "Overheating Failure Type" is added. Then, the broken relationship paths are repaired; for example, the edge "Temperature Test Event - Particle Test Failure Event" is added to the broken relationship path "Voltage Test Event - Current Test Event - Temperature Test Event - Particle Test Failure Event". Finally, the computer device maintains the topological connectivity of the initial knowledge graph structure.

[0116] Step 1666: Use the initial knowledge graph structure after topological connection calibration as the output graph of the current graph evolution cycle, and use the output graph as the input knowledge graph structure of the next graph evolution cycle. Iterate through the graph evolution cycle until the association matching result no longer changes or the preset graph evolution termination condition is reached. Use the stable knowledge graph structure obtained after the final iteration as the structured test knowledge graph.

[0117] In this embodiment, the computer device uses the initial knowledge graph structure, after topology connection calibration, as the output graph for the current graph evolution cycle. Then, the output graph is used as the input knowledge graph structure for the next graph evolution cycle, iteratively executing the graph evolution cycle. This continues until the association matching results no longer change or a preset graph evolution termination condition is reached; for example, the preset graph evolution termination condition is that the number of graph evolution cycle executions reaches a fixed value. Finally, the computer device uses the stable knowledge graph structure obtained after the final iteration as the structured test knowledge graph.

[0118] Optionally, the method further includes:

[0119] Step 170: Receive an optimization trigger instruction for the Flash storage particle test task, and based on the optimization trigger instruction, locate the set of key test stages that are directly related to particle test failure events and the particle response behavior patterns corresponding to the set of key test stages from the structured test knowledge graph.

[0120] In this embodiment, the computer device receives an optimized trigger instruction for a Flash memory chip testing task. This instruction is generated automatically by the user or the system. Next, based on the optimized trigger instruction, the system locates a set of key test stages directly related to chip test failure events and the corresponding chip response behavior patterns from a structured test knowledge graph. For example, the set of key test stages directly related to chip test failure events located from the structured test knowledge graph is "voltage test stage," "current test stage," and "temperature test stage," and the corresponding chip response behavior pattern is "voltage rise -- current rise -- temperature rise -- chip test failure."

[0121] Step 171: Based on the node connection relationship of the particle response behavior pattern in the structured test knowledge graph, construct an instruction influence network with test instructions as nodes and the particle performance influence intensity between instructions as edges. Traverse the instruction influence network, calculate the network centrality index of each test instruction node in the key test stage set, and adjust the network centrality index by weighting it in combination with the occurrence probability of the particle test failure event to generate an instruction adjustment priority sequence for each test instruction node.

[0122] In this embodiment, the computer device constructs an instruction influence network based on the node connection relationships of the granular response behavior patterns in a structured test knowledge graph. This network uses test instructions as nodes and the intensity of granular performance influence between instructions as edges. For example, the test instruction nodes are "voltage test instruction," "current test instruction," and "temperature test instruction," with the intensity of granular performance influence between instructions being a fixed value. Then, the instruction influence network is traversed, and the network centrality index of each test instruction node within the key test stage set is calculated. The network centrality index includes degree centrality, betweenness centrality, and proximity centrality. For example, the degree centrality, betweenness centrality, and proximity centrality of the "voltage test instruction" are calculated as fixed values; the degree centrality, betweenness centrality, and proximity centrality of the "current test instruction" are calculated as fixed values; and the degree centrality, betweenness centrality, and proximity centrality of the "temperature test instruction" are calculated as fixed values. Then, the network centrality index is weighted and adjusted based on the probability of granular test failure events. The probability of granular test failure events is a fixed value, and the weighting adjustment method is to multiply the network centrality index by the probability of granular test failure events. Finally, the computer equipment generates an instruction adjustment priority sequence for each test instruction node, and the instruction adjustment priority sequence is sorted from high to low according to the weighted and adjusted network centrality index.

[0123] Step 172: Based on the instruction priority sequence adjustment, rearrange the execution order of the original test instruction sequence in the Flash memory chip test task and remove redundant instructions to generate an optimized test instruction sequence. Inject the optimized test instruction sequence into the execution flow of the Flash memory chip test task. Use the structured test knowledge graph as feedback to dynamically adjust the execution parameters of the test instructions in the optimized test instruction sequence until the probability of the chip test failure event is lower than a preset threshold, and output the final optimized test instruction sequence.

[0124] In this embodiment, the computer device rearranges the execution order of the original test instruction sequence in the Flash memory chip testing task and removes redundant instructions based on the instruction priority sequence adjustment. For example, if the original test instruction sequence is "voltage test instruction -- current test instruction -- temperature test instruction", the instruction priority sequence adjustment becomes "temperature test instruction -- current test instruction -- voltage test instruction", rearranging the execution order of the original test instruction sequence to "temperature test instruction -- current test instruction -- voltage test instruction", and removing redundant instructions. Next, an optimized test instruction sequence is generated and injected into the execution flow of the Flash memory chip testing task. Then, the execution parameters of the test instructions in the optimized test instruction sequence are dynamically adjusted based on a structured test knowledge graph, such as adjusting the execution time interval and number of executions of the temperature test instruction. Finally, the computer device outputs the final optimized test instruction sequence when the probability of a chip test failure event falls below a preset threshold.

[0125] Optionally, the method further includes:

[0126] Step 180: Obtain multiple historical structured test knowledge graphs generated in the historical Flash storage particle test task, extract the node set and the relationship set between the nodes contained in the historical structured test knowledge graph, obtain the node feature set and the relationship feature set, fuse and encode the node feature set and the relationship feature set to generate the graph feature vector corresponding to each historical structured test knowledge graph.

[0127] In this embodiment, the computer device acquires multiple historical structured test knowledge graphs generated in historical Flash storage particle testing tasks. These graphs originate from the execution records of historical Flash storage particle testing tasks. Next, the set of nodes and the set of relationships between nodes within the historical structured test knowledge graphs are extracted. The node set includes entities from historical test state information and nodes from historical semantic mapping relationships. The relationship set includes the associations between entities from historical test state information and nodes from historical semantic mapping relationships. Then, a set of node features and a set of relationship features are obtained. The node feature set includes features such as node name, description, and type. The relationship feature set includes features such as relationship type, strength, and direction. Finally, the computer device fuses and encodes the node feature set and the relationship feature set, using a fusion encoding algorithm to convert the node feature set and the relationship feature set into graph feature vectors, generating a graph feature vector corresponding to each historical structured test knowledge graph.

[0128] Step 181: Calculate the graph similarity between the graph feature vector of the structured test knowledge graph and the graph feature vector of each historical structured test knowledge graph. Based on the graph similarity, select a set of candidate historical graphs with similarity higher than a preset threshold. Perform pattern clustering analysis on the graph feature vectors in the candidate historical graph set to identify repetitive abnormal pattern clusters across historical Flash storage particle test tasks. Extract the common node features and common relationship paths corresponding to the repetitive abnormal pattern clusters.

[0129] In this embodiment, the computer device calculates the graph similarity between the graph feature vector of the structured test knowledge graph and the graph feature vectors of each historical structured test knowledge graph. The graph similarity is calculated using the cosine similarity algorithm. Next, a set of candidate historical graphs with similarity scores higher than a preset threshold is selected based on the graph similarity. The preset threshold is a fixed value, for example, 0.8. Then, pattern clustering analysis is performed on the graph feature vectors in the candidate historical graph set, using a clustering algorithm to group historical structured test knowledge graphs with similar graph feature vectors into one class. Next, repetitive abnormal pattern clusters across historical Flash storage particle test tasks are identified. These repetitive abnormal pattern clusters are abnormal patterns that appear repeatedly in multiple historical structured test knowledge graphs. Finally, the computer device extracts the common node features and common relationship paths corresponding to the repetitive abnormal pattern clusters. The common node features are the common features of nodes in the repetitive abnormal pattern cluster, and the common relationship paths are the common paths of relationships in the repetitive abnormal pattern cluster.

[0130] Step 182: Map the common node features and common relationship paths back to the topology of the structured test knowledge graph, locate the potential anomalous subgraph structure in the structured test knowledge graph that matches the repetitive anomalous pattern cluster, construct the association network between the potential anomalous subgraph structure and the repetitive anomalous pattern cluster, analyze the migration path and evolution trend of the repetitive anomalous pattern cluster in the association network, and generate a cross-particle anomalous pattern migration report.

[0131] In this embodiment, the computer device maps common node features and common relationship paths back to the topology of the structured test knowledge graph. For example, it maps the common node feature "voltage increase trend" and the common relationship path "voltage increase trend -- current increase trend -- temperature increase trend -- particle test failure" back to the topology of the structured test knowledge graph. Next, it locates potential anomalous subgraph structures in the structured test knowledge graph that match the repetitive anomalous pattern clusters. These potential anomalous subgraph structures are subgraph structures in the structured test knowledge graph similar to the repetitive anomalous pattern clusters. Then, it constructs an association network between the potential anomalous subgraph structures and the repetitive anomalous pattern clusters. The nodes in the association network represent the potential anomalous subgraph structures and the repetitive anomalous pattern clusters, and the edges represent the association relationships between them. Finally, it analyzes the migration path and evolution trend of the repetitive anomalous pattern clusters in the association network. The migration path represents the propagation path of the repetitive anomalous pattern clusters in the association network, and the evolution trend represents the changing trend of the repetitive anomalous pattern clusters. Finally, the computer device generates an anomaly pattern migration report across particles, which includes information such as the potential anomaly subgraph structure, recurring anomaly pattern clusters, migration paths, and evolution trends.

[0132] It is important to note that the technical steps in the embodiments of this application encompass multiple complex stages, including data acquisition and time-series analysis, and introduce parameters from various sources and dimensions, such as time-series data and performance indicators. These parameters have different physical meanings and dimensions, such as time units, ratio units, and abstract metrics. Directly performing mathematical operations without considering dimensional uniformity can lead to a mismatch in the physical meaning of the calculations, affecting the accuracy and interpretability of the results. Those skilled in the art can overcome such problems through adaptive normalization. Normalization can convert parameters with different dimensions to a unified scale or dimensionless space. Common methods include min-max scaling and Z-score normalization. This process eliminates dimensional differences, ensuring that subsequent calculations can be performed on a consistent numerical basis, thus guaranteeing the feasibility of the solution. Adaptive normalization is a standard practice in data preprocessing, a conventional technique in the field, and is common knowledge in areas such as artificial intelligence. Multiple mature methods can be selected as needed, and these methods are integrated into various open-source tools for direct use, allowing technicians to implement them without creative effort. Normalization is crucial for ensuring the robustness of algorithms and the reliability of results. It smooths out differences across different scales, facilitates the smooth progress of each stage, improves model convergence speed and generalization ability, and reduces the risk of dimensional errors. Those skilled in the art can naturally integrate it into the process, ensuring the logical coherence and feasibility of the overall technical solution, and achieving efficient processing and intelligent analysis of Flash storage particle test data.

[0133] This application embodiment collects particle response behavior data generated during the execution of Flash storage particle testing tasks and extracts temporal correlation logic information. Combined with a behavior sequence analysis algorithm adapted to the testing task, a dynamic reasoning model based on behavior sequences is created. This achieves in-depth mining of the evolutionary patterns of particle response behavior under different instruction requests and accurate generation of hidden evolutionary trajectories. Furthermore, by inferring the degradation trend prediction results of particle test failure events through particle performance evolution characteristics in the hidden evolutionary trajectories, the trend features are transformed into semantic knowledge nodes and a semantic mapping relationship with failure types is established. Finally, a structured test knowledge graph is generated by combining test log data. This design breaks through the limitations of single-feature judgment in traditional Flash particle testing. It constructs a complete technical link from data collection and evolutionary reasoning to knowledge graph construction, with temporal correlation logic as the core. This enables early positioning of Flash particle stability degradation trends, providing a new qualitative judgment dimension for lifetime prediction. Simultaneously, it transforms scattered qualitative test data into structured knowledge that can directly guide product iteration, breaking down the information barrier between test data and design optimization, and improving the efficiency and accuracy of the transformation from test data to decision-making.

[0134] Please see Figure 2 The figure is a schematic diagram of the basic structure of a computer device 200 provided in an embodiment of this application. The computer device 200 includes: a processor 201; a storage device 202 on which a computer program 2020 is stored; and a network interface 203 for providing network communication functions. When the computer program 2020 is executed by the processor 201, the processor 201 implements any of the aforementioned artificial intelligence-based Flash particle test data processing methods.

[0135] Please see Figure 3 This application provides a functional block diagram of a Flash particle test data processing device, which includes:

[0136] The inference model creation module is used to collect particle response behavior data generated during the execution of Flash storage particle test tasks, extract the temporal correlation logic information of particle response behavior based on the particle response behavior data, and create a dynamic inference model based on behavior sequence based on the behavior sequence analysis algorithm adapted to the Flash storage particle test task and the particle response behavior characteristics of different test stages in the temporal correlation logic information.

[0137] The evolution trajectory generation module is used to input the particle response behavior data corresponding to different instruction requests in the Flash storage particle test task into the dynamic inference model for time series analysis, to explore the evolution law of particle response behavior under different instruction requests, and to generate hidden evolution trajectories corresponding to different instruction requests.

[0138] The degradation trend derivation module is used to deduce and analyze the occurrence trend and impact range of particle test failure events through the particle performance evolution characteristics contained in the hidden evolution trajectory, and obtain the degradation trend prediction results associated with the particle test failure events.

[0139] The semantic mapping establishment module is used to convert each trend feature in the decline trend prediction result into a corresponding semantic knowledge node, and to establish a semantic mapping relationship between trend features and failure types by combining each semantic knowledge node and particle test failure events.

[0140] The knowledge graph generation module is used to call the test log data generated during the execution of the particle test task stored in the Flash, extract the test status information related to the particle response behavior and the particle test failure event from the test log data, perform association matching based on the test status information and the semantic mapping relationship, and generate a structured test knowledge graph using the association matching results.

[0141] Based on the above, a readable storage medium is provided, on which a program or instructions are stored, and when the program or instructions are executed by a processor, the steps of the above method are implemented.

[0142] Furthermore, it should be noted that this application also provides a computer program product, which may include a computer program that can be stored in a computer-readable storage medium. The processor of a computer device reads the computer program from the computer-readable storage medium, and the processor can execute the computer program, causing the computer device to perform the aforementioned... Figure 1 The methods described in the corresponding embodiments are already known, and therefore will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated. For technical details not disclosed in the computer program product embodiments related to this application, please refer to the description of the method embodiments of this application.

[0143] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems or apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant parts can be referred to the method section.

Claims

1. A method for processing Flash particle test data based on artificial intelligence, characterized in that, The method includes: Collect particle response behavior data generated during the execution of Flash storage particle test tasks, and extract the temporal correlation logic information of particle response behavior based on the particle response behavior data; Based on the behavior sequence analysis algorithm adapted to the Flash storage particle test task, and the particle response behavior characteristics of different test stages in the time-series correlation logic information, a dynamic reasoning model based on behavior sequence is created. The particle response behavior data corresponding to different instruction requests in the Flash storage particle test task are input into the dynamic inference model for time series analysis to uncover the evolutionary patterns of particle response behavior under different instruction requests and generate hidden evolutionary trajectories corresponding to different instruction requests. Based on the time series correlation logic information, the interval parameters for time window division are determined. The particle response behavior data corresponding to different instruction requests in the Flash storage particle test task are divided into time windows according to the interval parameters, resulting in several data windows that correspond one-to-one with each instruction request. Each data window contains particle response behavior data within a continuous time period during the execution of the corresponding instruction request. A latent variable model is used to extract features from the particle response behavior data within each data window, generating a hidden state sequence corresponding to each data window. The hidden state sequence is used to characterize the intrinsic change features of particle response behavior within the corresponding data window. Behavioral patterns are compared between the hidden state sequences corresponding to different instruction requests. The method involves aligning hidden state sequences with similar change characteristics during the comparison process, calculating the similarity parameter between the aligned hidden state sequences, clustering all hidden state sequences based on the similarity parameter, and then temporally segmenting the clustered hidden state sequences according to their chronological order to obtain several temporal segments. The method extracts the state transition probability of the hidden state sequences within each temporal segment, using this probability to characterize the evolutionary stability of the particle response behavior within the corresponding temporal segment. The state transition probability is input into the dynamic inference model, which, combined with temporal analysis capabilities, integrates the evolutionary features of each temporal segment to uncover the evolutionary patterns of particle response behavior under different command requests. Based on these evolutionary patterns, the method extracts the evolutionary paths of the hidden evolutionary trajectories corresponding to different command requests, generating a hidden evolutionary trajectory that corresponds one-to-one with each command request. This hidden evolutionary trajectory is used to present the dynamic change process of the particle response behavior under the corresponding command request. By analyzing the particle performance evolution characteristics contained in the hidden evolution trajectory, the occurrence trend and impact range of particle test failure events are deduced, and the degradation trend prediction results associated with the particle test failure events are obtained. Each trend feature in the decline trend prediction result is converted into a corresponding semantic knowledge node, and a semantic mapping relationship between trend features and failure types is established by combining each semantic knowledge node and particle test failure events. The test log data generated during the execution of the particle test task is called in the Flash storage. Test status information related to the particle response behavior and the particle test failure event is extracted from the test log data. Based on the test status information and the semantic mapping relationship, association matching is performed, and a structured test knowledge graph is generated using the association matching results.

2. The method according to claim 1, characterized in that, The process of deriving and analyzing the occurrence trend and impact range of particle test failure events through the particle performance evolution characteristics contained in the hidden evolution trajectory, and obtaining the degradation trend prediction results associated with the particle test failure events, includes: Extract all evolutionary features related to particle performance from the hidden evolutionary trajectory, classify and organize the extracted evolutionary features according to performance type, and construct a corresponding performance index sequence. The performance index sequence is used to characterize the dynamic changes in particle performance during the execution of the Flash storage particle test task. A time series prediction algorithm is used to perform trend analysis on the preprocessed performance index sequence to predict the change trend of particle performance within the target time period, and to obtain performance index prediction trend data. The causal inference model is invoked, and the correlation data between the predicted trend data of the performance index and the particle test failure event is input into the causal inference model. The causal inference model analyzes the linkage between the particle performance evolution characteristics and the particle test failure event, identifies the precursor features that lead to the occurrence of the particle test failure event, classifies and filters the precursor features, and obtains the key precursor features that directly affect the occurrence of the particle test failure event. Survival analysis was used to quantify the key precursor features, and combined with historical data on particle performance changes, the failure time of particle test failure events was estimated, and the probability of particle test failure events occurring in different time periods was calculated. The graph neural network is used to fuse the performance index prediction trend data, causal inference results and failure time estimation data and to quantify the particle performance degradation to obtain a quantified value of the degree of particle performance degradation. The impact diffusion simulation algorithm is invoked, and the impact range of the particle test failure event on the surrounding Flash memory particles and the entire Flash memory particle test task is simulated after the decay degree quantification value is combined. The risk ripple assessment of the impact range is then performed to obtain the risk ripple assessment result. By combining the performance index prediction trend data, failure time estimation results, degradation degree quantification value and risk ripple assessment results, the occurrence trend and impact range of particle test failure events are deduced and analyzed, and all deduction and analysis results are integrated to obtain the degradation trend prediction results associated with the particle test failure events.

3. The method according to claim 1, characterized in that, The step of converting each trend feature in the decline trend prediction result into a corresponding semantic knowledge node, and establishing a semantic mapping relationship between trend features and failure types by combining each semantic knowledge node and particle test failure events, includes: The trend features in the decline trend prediction results are decomposed into attribute vectors, and each trend feature is decomposed into multiple attribute vectors of different dimensions. Each attribute vector corresponds to a target attribute of the trend feature. A knowledge representation learning algorithm is used to embed the decomposed attribute vectors into a semantic space, mapping each attribute vector to a preset semantic space to generate initial node embeddings corresponding to each trend feature. The initial node embeddings are used to characterize the position and semantic features of the trend features in the semantic space. The graph attention mechanism is used to optimize all initial node embeddings, strengthen the semantic association between the initial node embeddings corresponding to each trend feature, and extract the embedding path between each initial node embedding by combining semantic association learning. Based on the embedding path, an entity relationship graph is constructed. The entities in the entity relationship graph are the initial node embeddings corresponding to each trend feature, and the relationship is the semantic association between each node embedding. Establish a mapping rule base, which contains association rules between trend characteristics and failure types; Based on the mapping rule base, the correspondence between each trend feature and failure type is determined. The feature dimension mapping algorithm is used to map the attribute vector of each trend feature to the corresponding failure type feature space, and the transformation of each trend feature to semantic knowledge node is completed, so that each trend feature corresponds to a unique semantic knowledge node. The transformed semantic knowledge nodes are associated with entities in the entity relationship graph. By combining the embedded paths in the entity relationship graph, a semantic mapping relationship between trend features and failure types is established. The semantic mapping relationship is used to characterize the corresponding association between each trend feature and different failure types.

4. The method according to claim 3, characterized in that, The step of associating the transformed semantic knowledge nodes with entities in the entity relationship graph, and establishing a semantic mapping relationship between trend features and failure types by combining the embedded paths in the entity relationship graph, includes: Based on the correspondence between entities in the entity relationship graph and semantic knowledge nodes, an auxiliary association graph between semantic knowledge nodes and entities in the entity relationship graph is constructed. Based on the node connection order and path hop count indicated by the embedded path, calculate the semantic connectivity between the semantic knowledge node and the entity in the entity relationship graph, and assign an initial association weight to each association edge in the auxiliary association graph based on the semantic connectivity. The node representation learning algorithm is invoked to vectorize all nodes in the auxiliary association graph, generating the semantic knowledge node and the node representation vector of the entity in the entity relationship graph. The representation similarity between the semantic knowledge node and the entity in the entity relationship graph is calculated using the node representation vector. The initial association weights are dynamically adjusted based on the representation similarity to obtain an updated set of association weights. The updated set of association weights is used to quantify the degree of association between the semantic knowledge nodes and entities in the entity relationship graph. The topology of the auxiliary association graph is optimized based on the updated association weight set. The association edges corresponding to association weights below a preset threshold in the updated association weight set are removed, while the association edges corresponding to association weights above a preset threshold in the updated association weight set are retained, resulting in an optimized auxiliary association graph. Based on the association edges retained by the optimized auxiliary association graph, a stable node connection relationship between the semantic knowledge node and the entity in the entity relationship graph is determined. The semantic relevance of the stable node connection relationship is calculated by using a bidirectional attention mechanism. The forward attention score of the semantic knowledge node pointing to the entity in the entity relationship graph and the backward attention score of the entity in the entity relationship graph pointing to the semantic knowledge node are extracted. The forward attention score and the backward attention score are weighted and fused to obtain the comprehensive semantic relevance. The comprehensive semantic association degree is coupled with the path weight in the embedded path to generate the fused semantic association value between the semantic knowledge node and the entity in the entity relationship graph; Using the fused semantic association value as a bridge, the trend features represented by the semantic knowledge nodes are linked semantically to the failure types associated with entities in the entity relationship graph, thereby completing the construction and solidification of the semantic mapping relationship between trend features and failure types.

5. The method according to claim 1, characterized in that, The step of performing association matching based on the test state information and the semantic mapping relationship, and generating a structured test knowledge graph using the association matching results, includes: The test status information related to the particle response behavior and the particle test failure event extracted from the test log data is organized, and the organized test status information is sorted according to the chronological order of the events to construct a structured event flow. Each event node in the structured event flow corresponds to a test status information, and each event node contains the event occurrence time, event type, and related particle response behavior data. A graph matching algorithm is used to associate and match event nodes in the structured event stream with semantic knowledge nodes in the semantic mapping relationship. A similarity measurement algorithm is introduced to calculate the similarity between each event node and each semantic knowledge node. Based on the similarity, the degree of association between event nodes and semantic knowledge nodes is determined. Event nodes with an association degree that reaches a preset threshold are aligned with semantic knowledge nodes to complete the initial association matching between event nodes and semantic knowledge nodes. The event relationships in the structured event stream are fused with the embedded paths in the semantic mapping relationship. Duplicate relationships existing in the fusion process are merged to achieve relationship fusion, resulting in a fused set of nodes and relationships. An initial knowledge graph structure is constructed based on the fused set of nodes and relationships. The topology of the initial knowledge graph structure is then updated to optimize the node connection method of the knowledge graph. New event nodes related to particle test failure events are extracted from the test status information. The new event nodes are aligned and then integrated into the initial knowledge graph structure to realize knowledge graph evolution based on the number of nodes and their relationships. During the evolution of the knowledge graph, the association matching results are continuously integrated into the initial knowledge graph structure, and the node attributes and relationships in the initial knowledge graph structure are dynamically updated to generate a structured test knowledge graph.

6. The method according to claim 5, characterized in that, The process of fusing the event relationships in the structured event stream with the embedded paths in the semantic mapping relationship, merging duplicate relationships during the fusion process to achieve relationship fusion, and obtaining a fused set of nodes and relationships includes: The event relationships formed between adjacent event nodes in the structured event flow are analyzed, the event relationship type and event relationship strength of the event relationship are extracted, and the event relationship type and event relationship strength are vectorized to generate the event relationship vector corresponding to each event relationship in the structured event flow. The embedded paths in the semantic mapping relationship are parsed, and the path relationship types and path relationship strengths between the semantic knowledge nodes connected by the embedded paths are extracted. The path relationship types and path relationship strengths are vectorized to generate a path relationship vector corresponding to each embedded path in the semantic mapping relationship. A relationship alignment algorithm is used to calculate the similarity between the event relationship vector and the path relationship vector, and the vector distance between the event relationship vector and the path relationship vector is measured in terms of relationship type dimension and relationship strength dimension. Based on the vector distance, the semantic similarity between the event relationship and the embedded path is determined. Based on the semantic similarity, the event relationship and the embedding path are matched. Event relationships and embedding paths with semantic similarity reaching a preset alignment threshold are marked as relationship pairs to be fused. The relationship attributes of the event relationships and embedding paths in the relationship pairs to be fused are compared to identify the overlapping and non-overlapping parts of the relationship attributes between the event relationships and embedding paths in the relationship pairs to be fused. The overlapping relational attributes are merged and deduplicated to generate a fused relational attribute set, which inherits the relational attributes of the event relationship and the embedded path in the overlapping part. The non-overlapping relational attributes are supplemented and expanded, and the non-overlapping relational attributes are added as new attributes to the fused relational attribute set to complete the relational fusion of the relational pair to be fused, and obtain the fused relation corresponding to the relational pair to be fused. For event relationships and embedding paths whose semantic similarity does not reach the preset alignment threshold, the event relationships and embedding paths are retained as independent relationships in the set of relationships to be fused. Integrate all the fused relationships and all the independent relationships to form a fused relationship set; extract all nodes connected by the relationships in the fused relationship set, the nodes including event nodes in the structured event stream and semantic knowledge nodes in the semantic mapping relationship, and combine the nodes with the fused relationship set to generate the fused node and relationship set.

7. The method according to claim 5, characterized in that, During the knowledge graph evolution process, the association matching results are continuously integrated into the initial knowledge graph structure, and the node attributes and relationships in the initial knowledge graph structure are dynamically updated to generate a structured test knowledge graph, including: A graph evolution cycle is set, and the changes in the association matching results are monitored in each graph evolution cycle. The newly added matching part and attribute update part of the association matching results relative to the previous graph evolution cycle are collected. The newly added matching part and attribute update part are integrated into an incremental matching result set. The incremental matching result set is structurally parsed to separate the newly added nodes, node attribute update information and newly added relationships in the incremental matching result set. For the newly added node, calculate the probability of topological connection between the newly added node and existing nodes in the initial knowledge graph structure, insert the newly added node into the corresponding topological position in the initial knowledge graph structure according to the probability of topological connection, and establish the initial connection relationship between the newly added node and existing nodes in the initial knowledge graph structure. For the node attribute update information, locate the target node in the initial knowledge graph structure pointed to by the node attribute update information, merge the attribute content in the node attribute update information with the existing attributes of the target node, and if there is an attribute conflict between the attribute content in the node attribute update information and the existing attributes of the target node, then select and merge the conflicting attributes according to the attribute timestamp or attribute confidence, and complete the dynamic update of the attributes of the target node. For the newly added relationship, verify whether the nodes connected by the newly added relationship already exist in the initial knowledge graph structure. If all the nodes connected by the newly added relationship already exist, then add the newly added relationship as a new edge to the initial knowledge graph structure. If there are nodes among the nodes connected by the newly added relationship that have not been added to the initial knowledge graph structure, then prioritize processing the nodes that have not been added as newly added nodes before adding the newly added relationship. After inserting new nodes, merging node attribute update information, and adding new relationships, the overall topological connectivity of the initial knowledge graph structure is calibrated. The system detects whether there are isolated nodes or broken relationship paths in the initial knowledge graph structure. For the isolated nodes detected, the system supplements their connection relationships with neighboring nodes. For the broken relationship paths detected, the system repairs the path to maintain the topological connectivity of the initial knowledge graph structure. The initial knowledge graph structure after topological connection calibration is used as the output graph of the current graph evolution cycle, and the output graph is used as the input knowledge graph structure of the next graph evolution cycle. The graph evolution cycle is iteratively executed until the association matching result no longer changes or the preset graph evolution termination condition is reached. The stable knowledge graph structure obtained after the final iteration is used as the structured test knowledge graph.

8. A computer device, characterized in that, include: A processor; a storage device having a computer program stored thereon; a network interface for providing network communication functions; when the computer program is executed by the processor, the processor enables the processor to implement the artificial intelligence-based Flash particle test data processing method as described in any one of claims 1-7.

9. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions, which, when executed by a processor, implement the artificial intelligence-based Flash particle test data processing method as described in any one of claims 1-7.

Images