Integrated circuit testing method and system
Patent Information
- Application Number
- CN202411396966.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-09
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-10-09
AI Technical Summary
Existing integrated circuit testing methods are inefficient and prone to errors, making it difficult to effectively apply artificial intelligence technology to automatically generate test cases and accurately locate fault areas.
We construct a knowledge graph for integrated circuit testing, extract circuit topology and functional modules using graph neural networks, generate test cases through multi-task learning, and locate fault areas by combining a simulation engine with multiple fault models and spectrum analysis methods, thereby achieving hardware and software data fusion and simulation testing.
It significantly improves the efficiency and accuracy of test case generation, enhances the comprehensiveness, precision, and reliability of fault diagnosis, and reduces omissions and errors in the testing process.
Smart Images

Figure CN118884191B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to integrated circuit testing technology, and in particular to an integrated circuit testing method and system. Background Art
[0002] Integrated circuit testing is a key link in ensuring chip quality. Traditional testing methods mainly rely on manually designed test vectors and manually analyzed test results, which is inefficient and error-prone when facing modern complex integrated circuits. In recent years, with the development of artificial intelligence technology, especially the rise of technologies such as graph neural networks and knowledge graphs, new opportunities have been brought to the field of integrated circuit testing. However, how to effectively apply these advanced technologies to actual integrated circuit testing, especially in automatically generating test cases and accurately locating fault areas, still faces huge challenges. The present invention aims to solve these problems and proposes an integrated circuit testing method and system. Summary of the invention
[0003] The embodiments of the present invention provide an integrated circuit testing method and system, which can solve the problems in the prior art.
[0004] According to a first aspect of the embodiments of the present invention,
[0005] A method for testing an integrated circuit is provided, comprising:
[0006] Construct a knowledge graph in the field of integrated circuit testing, extract the circuit topology from the schematic diagram of the circuit to be tested based on the knowledge graph in the field of integrated circuit testing, identify the functional modules in the circuit, extract the input and output characteristics of the functional modules, form a circuit knowledge representation, perform graph embedding learning on the circuit knowledge representation, embed the nodes and edges in the circuit diagram into a low-dimensional vector space through a graph neural network, extract the structure and function representation of the circuit using a message passing mechanism, input the extracted structure and function representation as a feature vector into a pre-built test case generation model, and automatically generate test cases that meet the test requirements and constraints;
[0007] Input the generated test case into a simulation engine that supports multiple fault models, set the fault type, injection position and trigger condition through the simulation engine, dynamically modify the circuit model behavior to simulate the impact of the fault on the normal operation of the circuit, generate simulation data, extract the feature representation of the simulation data, calculate the similarity between the simulation data and the reference data based on the extracted feature representation of the simulation data, and pre-locate the first candidate fault area in the simulation data;
[0008] Perform hardware testing on the integrated circuit, collect hardware test response data, align the collected hardware test response data with the simulation data, establish a mapping relationship between the software and hardware data, and based on the mapping relationship, use a spectral analysis method to calculate the difference between the hardware test data and the simulation data, locate the second candidate fault area in the hardware test data, combine the first candidate fault area and the second candidate fault area, determine the final fault area, and output fault diagnosis suggestions.
[0009] In an optional embodiment,
[0010] The nodes and edges in the circuit diagram are embedded into a low-dimensional vector space through a graph neural network, and the structure and function representation of the circuit are extracted using a message passing mechanism. The extracted structure and function representation are input as feature vectors into a pre-built test case generation model to automatically generate test cases that meet the test requirements and constraints, including:
[0011] The circuit is converted into a graph structure representation, where nodes represent circuit components and edges represent the connection relationship between components. Each node in the circuit is represented as a node feature and aggregated with neighbor node information. A message passing mechanism is used to enable each node to interact with its neighboring nodes to generate an embedded representation of the node. Based on the embedded representation of the node, a global representation feature of the circuit is formed, and a shared layer for circuit feature extraction is constructed.
[0012] On top of the shared layer, a task-specific layer is constructed, wherein for the fault detection task, a classification network based on a graph neural network is used to process the global representation features of the circuit; for the performance verification task, a regression network is used to process the global representation features of the circuit; for the power consumption optimization task, a convolutional neural network or a recursive neural network is used to process the global representation features of the circuit to obtain the output results of each task;
[0013] According to the loss function of each task, a multi-task loss function is constructed, and the multi-task loss function is optimized at the same time to obtain the loss value of each task, and the loss gradient of each task is calculated according to the loss value of each task, and the weight of each task is dynamically adjusted based on the loss gradient, and the multi-task loss function is updated according to the adjusted task weight, and the shared layer and the task-specific layer are jointly trained;
[0014] In the process of the joint training, a multi-task alternating optimization strategy is adopted, different tasks are set as dominant tasks respectively, the loss function of the dominant task is optimized, and the parameters of the task-specific layer corresponding to the shared layer and the dominant task are updated until all tasks are trained in sequence, thereby obtaining a trained test case generation model;
[0015] In the encoder of the trained test case generation model, a graph attention layer is used to aggregate the node features output by the shared layer to capture the global structural information of the circuit. At the same time, in the decoder, decoding is performed according to the output of the encoder and the test requirements to generate multiple candidate test case sequences. The candidate test case sequences are evaluated and screened through preset screening rules and scoring functions to obtain test cases that meet the test requirements and constraints.
[0016] In an optional embodiment,
[0017] The candidate test case sequences are evaluated and screened by using preset screening rules and scoring functions to obtain test cases that meet the test requirements and constraints, including:
[0018] The test case screening process is modeled as a Markov decision process, the state space is defined as the candidate test case sequence to be screened, the action space is the screening decision of the test case, the reward function is the quality score of the test case, and a transition probability function is constructed to describe the state transition. A deep neural network is constructed as a value function network and a policy network. The value function network is used to estimate the state-action value function, and the policy network is used to generate the screening decision.
[0019] Initialize the screening rules and scoring function as the initial strategy, create a priority experience playback buffer, and set the priority sampling weight, screen the candidate test case sequence according to the current state and the initial strategy, obtain the screening decision and immediate reward, and calculate the test case coverage and fault detection rate;
[0020] The screening decision, the immediate reward, the test case coverage and the fault detection rate are stored in a priority experience replay buffer, and the sample priority is updated according to the temporal difference error, and importance sampling is performed from the priority experience replay buffer according to the priority sampling weight to obtain an experience data batch;
[0021] Calculate the target Q value using the state, action, reward and next state information in the experience data batch, and update the value function network by minimizing the mean square error between the predicted Q value and the target Q value; at the same time, calculate the policy gradient using the policy gradient method based on the experience data batch and the updated value function network, and update the policy network according to the policy gradient;
[0022] Based on the updated value function network and policy network, combined with the test case coverage and fault detection rate information in the empirical data batch, the parameters of the screening rules and scoring functions are updated, and the iteration is repeated until the preset termination condition is reached. The policy networks of multiple training stages are combined using an ensemble learning method to generate the final screening policy network;
[0023] The final screening strategy network is used to screen the candidate test case sequences, and the test case quality is evaluated based on the test case coverage and fault detection rate. The correlation and complementarity between test cases are considered at the same time, and test cases that meet the test requirements and constraints are obtained.
[0024] In an optional embodiment,
[0025] The generated test cases are input into a simulation engine that supports multiple fault models. The fault type, injection location, and trigger condition are set through the simulation engine. The circuit model behavior is dynamically modified to simulate the impact of the fault on the normal operation of the circuit. Simulation data is generated, and feature representation of the simulation data is extracted. The similarity between the simulation data and the reference data is calculated based on the extracted feature representation of the simulation data. The first candidate fault area in the pre-positioned simulation data includes:
[0026] Setting a fault model library, the fault model library includes predefined fault types, adding parameter faults, delay faults and bridge faults to the fault model library, creating a user-defined fault model interface, and obtaining an extended fault model library;
[0027] Creating a fault injection language and script, and injecting faults in the extended fault model library into a pre-built multi-level circuit model using the fault injection language and script to obtain a fault injection model;
[0028] The generated test cases are input into the simulation engine that supports multiple fault models. Parallel computing, incremental evaluation and event-driven technologies are integrated into the simulation execution engine. The fault type to be injected is selected according to the extended fault model library. The fault injection location is determined in the multi-level circuit model. The fault triggering conditions are set. During the simulation process, when the triggering conditions are met, the fault injection model is simulated by changing the parameter values of the specified components or modules, adjusting the signal propagation delay, adding connections that should not exist between the specified nodes, or modifying the corresponding circuit behavior according to the predefined fault model to obtain simulation data reflecting the impact of the fault.
[0029] Extracting features from the obtained simulation data, processing the extracted features by using feature dimension reduction and transformation methods to generate an optimized feature set, constructing and training a similarity measurement model based on the optimized feature set to obtain a trained similarity measurement model, and calculating the similarity between the simulation data and the reference data by using the trained similarity measurement model;
[0030] A multi-objective optimization model is constructed, which comprehensively considers positioning accuracy, efficiency and coverage factors, and based on the similarity calculation results, identifies the fault area in each level of the multi-level circuit model, prioritizes the identified fault areas, considers the fault impact and the probability of occurrence, uses a heuristic algorithm to search in areas with a priority higher than a preset threshold, combines the search results of different levels, determines the cross-level fault propagation path through inter-level correlation analysis, and generates the first candidate fault area based on the search results and the inter-level correlation.
[0031] In an optional embodiment,
[0032] A multi-objective optimization model is constructed, wherein the multi-objective optimization model comprehensively considers positioning accuracy, efficiency and coverage factors, and identifies the fault area in each level of the multi-level circuit model based on the similarity calculation result. The steps before prioritizing the identified fault areas also include:
[0033] Determine the degree of fault impact based on the fault impact characteristics in the simulation data, and calculate the probability of fault occurrence based on historical fault data and current circuit operation status;
[0034] Fuzzy sets are defined for the fault impact degree and the fault occurrence probability respectively, each fuzzy set is described by a corresponding fuzzy membership function, and a fuzzy rule base is constructed at the same time, wherein the fuzzy rule base contains fuzzy rules describing the relationship between the fault impact degree and the fault occurrence probability;
[0035] Performing fuzzification processing on the fault impact degree and the fault occurrence probability to obtain the membership degree of the fault impact degree and the fault occurrence probability in their respective fuzzy sets, performing fuzzy reasoning based on the fuzzy rule base and the membership degree to obtain the fuzzy output of the comprehensive evaluation result, performing defuzzification processing on the fuzzy output of the comprehensive evaluation result to obtain the comprehensive evaluation score, and determining the search priority according to the comprehensive evaluation score;
[0036] Based on the similarity calculation results and the search priority, the fault area is identified in each level of the multi-level circuit model, and the identified fault areas are prioritized.
[0037] In an optional embodiment,
[0038] Perform hardware testing on the integrated circuit, collect hardware test response data, align the collected hardware test response data with the simulation data, establish a mapping relationship between the software and hardware data, calculate the difference between the hardware test data and the simulation data based on the mapping relationship, locate the second candidate fault area in the hardware test data, combine the first candidate fault area and the second candidate fault area, determine the final fault area, and output fault diagnosis suggestions including:
[0039] Performing hardware testing on the target integrated circuit, collecting test response data under the actual operating environment through the standard test interface in the circuit design, the test response data including the state of the trigger, the level signal of the node and the operating parameters;
[0040] The test response data is mapped one by one with the corresponding signals and nodes in the simulation data through the signal name and node hierarchy to obtain the spatially aligned test response data; the spatially aligned test response data and the simulation data are time-aligned, calibrated through timestamp comparison, and a complete mapping relationship between software and hardware data is established to obtain the target test response data;
[0041] Perform fast Fourier transform on the target test response data and the simulation data to obtain the frequency domain representation of the target test response data and the frequency domain representation of the simulation data, and use the distance metric to calculate the frequency component difference between the target test response data and the simulation data in the frequency domain, map the frequency component difference exceeding the preset threshold back to the time domain, determine the time interval in which the test response data and the simulation data are inconsistent, and obtain the second candidate fault area;
[0042] Calculate the spatial overlap between the second candidate fault area and the first candidate fault area, classify the candidate fault areas into high overlap areas and low overlap areas based on a preset overlap threshold, apply an adaptive clustering algorithm to the high overlap areas to generate fused candidate fault areas, and use a multi-factor weighted scoring model for the low overlap areas to calculate the comprehensive confidence of each area;
[0043] The fused candidate fault areas and the low-overlap areas with the highest comprehensive confidence are selected to construct an optimized set of candidate fault areas. Combined with the topological structure information and functional module division of the target integrated circuit, the graph theory analysis method is used to screen the optimized set of candidate fault areas, determine the final fault area, and output fault diagnosis suggestions.
[0044] In an optional embodiment,
[0045] Apply graph theory analysis methods to screen the optimized candidate fault area set, determine the final fault area, and output fault diagnosis suggestions including:
[0046] Mapping the topological structure of the target integrated circuit into a graph structure, wherein nodes represent functional modules or submodules in the circuit, and edges represent signal connections or functional dependencies between modules;
[0047] The fault confidence and module function information of each node are obtained through circuit testing and simulation analysis, and stored in the graph structure as node attribute values. At the same time, the signal transmission delay and connection strength of each edge are calculated based on the circuit design parameters, and the calculation results are stored in the graph structure as edge weight values.
[0048] Using the graph structure that stores node attribute values and edge weight values, the shortest path algorithm is applied to calculate the shortest path from the node whose fault confidence exceeds the preset threshold to other nodes, and the preliminary fault propagation path is obtained;
[0049] Based on the maximum flow minimum cut algorithm, the minimum cut set of the graph structure is calculated to identify the modules and signal channels that have the greatest impact on the system function. By analyzing the connectivity between nodes and the weight values of the edges, the target path in the circuit is determined, and the nodes on the target path are formed into a target element set.
[0050] The regions in the optimized candidate fault region set are spatially matched with the nodes in the graph structure, and the correspondence between the candidate fault regions and the graph nodes is established. The candidate fault regions with the established correspondence are comprehensively scored, taking into account the fault confidence of the corresponding nodes, whether they are in the fault propagation path set, and whether they belong to the key element set;
[0051] The candidate fault areas are sorted in descending order according to the comprehensive scores, and the candidate fault area with the highest score and the corresponding node in the fault propagation path set is selected as the final fault area. Based on the final fault area, fault propagation path set and key element set, fault diagnosis suggestions are output.
[0052] According to a second aspect of the embodiments of the present invention,
[0053] An integrated circuit testing system is provided, comprising:
[0054] The first unit is used to construct a knowledge graph in the field of integrated circuit testing. According to the knowledge graph in the field of integrated circuit testing, the circuit topology is extracted from the schematic diagram of the circuit to be tested, the functional modules in the circuit are identified, the input and output characteristics of the functional modules are extracted, and the circuit knowledge representation is formed. The circuit knowledge representation is subjected to graph embedding learning, the nodes and edges in the circuit diagram are embedded into a low-dimensional vector space through a graph neural network, the structure and function representation of the circuit are extracted using a message passing mechanism, and the extracted structure and function representation are input as feature vectors into a pre-built test case generation model to automatically generate test cases that meet the test requirements and constraints.
[0055] The second unit is used to input the generated test case into a simulation engine that supports multiple fault models, set the fault type, injection position and trigger condition through the simulation engine, dynamically modify the circuit model behavior to simulate the impact of the fault on the normal operation of the circuit, generate simulation data, extract the feature representation of the simulation data, calculate the similarity between the simulation data and the reference data according to the extracted feature representation of the simulation data, and pre-locate the first candidate fault area in the simulation data;
[0056] The third unit is used to perform hardware testing on the integrated circuit, collect hardware test response data, align the collected hardware test response data with the simulation data, establish a mapping relationship between the software and hardware data, and based on the mapping relationship, use a spectral analysis method to calculate the difference between the hardware test data and the simulation data, locate the second candidate fault area in the hardware test data, combine the first candidate fault area and the second candidate fault area, determine the final fault area, and output a fault diagnosis suggestion.
[0057] According to a third aspect of the embodiments of the present invention,
[0058] An electronic device is provided, comprising:
[0059] processor;
[0060] a memory for storing processor-executable instructions;
[0061] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.
[0062] According to a fourth aspect of the embodiments of the present invention,
[0063] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.
[0064] In this embodiment, by constructing a knowledge graph in the field of integrated circuit testing, extracting the topological structure and functional module information of the circuit, and using graph neural network for graph embedding learning, the complex circuit structure is mapped to a low-dimensional space to capture its structural and functional characteristics. This process enables the test case generation model to automatically generate test cases that meet the test requirements and constraints, greatly reducing the time and energy of manually writing test cases, and significantly improving the efficiency and accuracy of test case generation. Through the simulation engine of the multi-fault model, faults are dynamically injected into the circuit model, and their impact on the normal operation of the circuit is simulated to generate simulation data. By extracting the simulation data features and the reference data for similarity calculation, the first candidate fault area in the simulation data can be efficiently pre-positioned. This simulation engine can not only identify potential faults in the design stage and reduce the workload of later debugging, but also provide a highly flexible fault simulation environment, so that different types of faults can be quickly simulated and analyzed. Combining hardware test data with simulation data, the second candidate fault area in the hardware test data is further located. This fusion analysis overcomes the deviation problem between the simulation model and the actual circuit, and improves the comprehensiveness, accuracy and reliability of fault analysis. The collaborative analysis of software and hardware data fusion and simulation test data is realized, and the efficiency of integrated circuit testing is significantly improved by automatically generating test cases, simulating fault injection and locating fault areas. The technical means of multi-model fusion, difference analysis and automatic fault location in the solution make the fault diagnosis of complex circuits more accurate, thereby improving the test quality and reliability of integrated circuits and reducing potential omissions and errors in the test process. BRIEF DESCRIPTION OF THE DRAWINGS
[0065] Figure 1 A schematic diagram of a flow chart of an integrated circuit testing method according to an embodiment of the present invention;
[0066] Figure 2 It is a schematic diagram of the structure of an integrated circuit testing system according to an embodiment of the present invention. DETAILED DESCRIPTION
[0067] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0068] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.
[0069] Figure 1 FIG. 1 is a flow chart of an integrated circuit testing method according to an embodiment of the present invention. Figure 1 As shown, the method includes:
[0070] S101. Construct a knowledge graph in the field of integrated circuit testing. According to the knowledge graph in the field of integrated circuit testing, extract the circuit topology from the schematic diagram of the circuit to be tested, identify the functional modules in the circuit, extract the input and output characteristics of the functional modules, form a circuit knowledge representation, perform graph embedding learning on the circuit knowledge representation, embed the nodes and edges in the circuit diagram into a low-dimensional vector space through a graph neural network, extract the structure and function representation of the circuit using a message passing mechanism, input the extracted structure and function representation as a feature vector into a pre-built test case generation model, and automatically generate test cases that meet the test requirements and constraints.
[0071] Among them, the knowledge graph in the field of integrated circuit testing contains key concepts, entities and semantic relationships in each link of integrated circuit design, manufacturing, testing, etc. By extracting and integrating knowledge from a large number of professional literature, patents, technical documents, etc., and using ontology construction technology, a hierarchical and semantic knowledge representation system is formed. The concept nodes in the knowledge graph can cover multiple dimensions such as circuit schematics, netlists, layouts, manufacturing processes, test standards, fault models, test algorithms, etc., and the relationships between concepts include composition, attributes, causality, timing and other types.
[0072] After obtaining the schematic diagram of the circuit to be tested, it is necessary to extract the circuit topology information and identify the functional modules. First, the schematic diagram is preprocessed using image processing technology to extract basic graphic elements such as component symbols and connections. Then, according to the concepts and relationships in the integrated circuit knowledge graph, these graphic elements are mapped to nodes and edges of the circuit to form a topological structure representation of the circuit. Next, by analyzing the connectivity, hierarchy and other characteristics of the circuit topology, and combining the concept of functional modules in the knowledge graph, the various functional sub-modules in the circuit are identified, such as operational amplifiers, comparators, filters, etc. For each functional module, its input and output pins and corresponding electrical characteristic parameters are further extracted. Finally, a structured circuit knowledge representation is formed, which contains information such as the circuit topology, functional module division and module characteristics.
[0073] The circuit knowledge representation is converted into a numerical feature vector, and the graph embedding learning method is adopted. By constructing a circuit graph with circuit elements as nodes and connections between elements as edges, the graph neural network model is used to perform end-to-end embedding learning. In the graph neural network, each node and edge is represented by a low-dimensional real-valued vector. Through iterative message passing and aggregation operations, the embedding vectors of nodes and edges are continuously updated so that they can carry the structure and function information of the circuit. Among them, the embedding vector of the node not only encodes the attribute characteristics of the node itself, but also integrates the information of its neighboring nodes; while the embedding vector of the edge describes the direction, weight and other characteristics of the connection between elements.
[0074] Using the circuit structure function representation learned in the previous step, combined with the pre-built test case generation model, test cases that meet the test requirements and constraints are automatically generated. The test case generation model can be constructed in a rule-based, model-based, or a combination of the two. Among them, the rule-based method summarizes a series of rule templates for test case generation based on expert knowledge and experience, such as boundary value testing, equivalence class partitioning, etc.; while the model-based method uses machine learning algorithms to learn the patterns and strategies of test case generation from historical test data. The graph embedding features of the circuit are input into the test case generation model, and through the reasoning and decision-making process, test stimuli and expected responses covering various fault types and fault modes are automatically generated. At the same time, constraints such as test coverage and test cost can also be introduced to optimize the quality and efficiency of test cases. The generated test cases will be used in subsequent simulation and hardware testing links to evaluate the functional correctness and reliability of the circuit.
[0075] In an optional embodiment,
[0076] The nodes and edges in the circuit diagram are embedded into a low-dimensional vector space through a graph neural network, and the structure and function representation of the circuit are extracted using a message passing mechanism. The extracted structure and function representation are input as feature vectors into a pre-built test case generation model to automatically generate test cases that meet the test requirements and constraints, including:
[0077] The circuit is converted into a graph structure representation, where nodes represent circuit components and edges represent the connection relationship between components. Each node in the circuit is represented as a node feature and aggregated with neighbor node information. A message passing mechanism is used to enable each node to interact with its neighboring nodes to generate an embedded representation of the node. Based on the embedded representation of the node, a global representation feature of the circuit is formed, and a shared layer for circuit feature extraction is constructed.
[0078] On top of the shared layer, a task-specific layer is constructed, wherein for the fault detection task, a classification network based on a graph neural network is used to process the global representation features of the circuit; for the performance verification task, a regression network is used to process the global representation features of the circuit; for the power consumption optimization task, a convolutional neural network or a recursive neural network is used to process the global representation features of the circuit to obtain the output results of each task;
[0079] According to the loss function of each task, a multi-task loss function is constructed, and the multi-task loss function is optimized at the same time to obtain the loss value of each task, and the loss gradient of each task is calculated according to the loss value of each task, and the weight of each task is dynamically adjusted based on the loss gradient, and the multi-task loss function is updated according to the adjusted task weight, and the shared layer and the task-specific layer are jointly trained;
[0080] In the process of the joint training, a multi-task alternating optimization strategy is adopted, different tasks are set as dominant tasks respectively, the loss function of the dominant task is optimized, and the parameters of the task-specific layer corresponding to the shared layer and the dominant task are updated until all tasks are trained in sequence, thereby obtaining a trained test case generation model;
[0081] In the encoder of the trained test case generation model, a graph attention layer is used to aggregate the node features output by the shared layer to capture the global structural information of the circuit. At the same time, in the decoder, decoding is performed according to the output of the encoder and the test requirements to generate multiple candidate test case sequences. The candidate test case sequences are evaluated and screened through preset screening rules and scoring functions to obtain test cases that meet the test requirements and constraints.
[0082] Exemplarily, the circuit is first converted into a graph structure representation, where nodes represent circuit components and edges represent the connection relationship between components. For each node, its features are extracted and aggregated with neighbor node information. Each node interacts with its neighbor nodes through a message passing mechanism to generate an embedded representation of the node. Based on the embedded representation of the node, the global representation features of the circuit are formed, and a shared layer for circuit feature extraction is constructed.
[0083] On top of the shared layer, task-specific layers are built for different test tasks. For fault detection tasks, a classification network based on graph neural networks is used to process the global representation features of the circuit; for performance verification tasks, a regression network is used to process the global representation features of the circuit; for power consumption optimization tasks, a convolutional neural network or a recursive neural network is used to process the global representation features of the circuit to obtain the output results of each task.
[0084] In order to achieve joint training of multiple tasks, a multi-task loss function is constructed based on the loss functions of each task, and it is expressed as the weighted sum of the loss functions of each task. At the same time, the multi-task loss function is optimized to obtain the loss value of each task. The loss gradient of each task is calculated based on the loss value of each task, and the weight of each task is dynamically adjusted based on the loss gradient. The multi-task loss function is updated according to the adjusted task weights, and the shared layer and task-specific layer are jointly trained.
[0085] During the joint training process, a multi-task alternating optimization strategy is adopted. Different tasks are set as the dominant tasks in turn, the loss function of the dominant task is optimized, and the parameters of the shared layer and the task-specific layer corresponding to the dominant task are updated until all tasks are trained in sequence, and the trained test case generation model is obtained.
[0086] In the trained test case generation model, the encoder uses the graph attention layer to aggregate the node features of the shared layer output to capture the global structural information of the circuit. The decoder decodes according to the encoder output and test requirements to generate multiple candidate test case sequences. The candidate test case sequences are evaluated and screened through preset heuristic rules and scoring functions to obtain test cases that meet the test requirements and constraints.
[0087] In this embodiment, by converting the circuit into a graph structure representation, the structure and functional characteristics of the circuit are extracted using a graph neural network, the topological structure and device attribute information of the circuit are fully mined, and rich prior knowledge is provided for test case generation. A multi-task learning framework is adopted. On the basis of the shared circuit feature extraction layer, a specific task layer is constructed for different test tasks to achieve joint optimization of multiple test tasks such as fault detection, performance verification and power consumption optimization, thereby improving the efficiency and quality of test case generation. By dynamically adjusting the weights of each task, balancing the importance of different tasks, avoiding a single task dominating the training process, and improving the comprehensiveness and balance of test case generation. The graph attention mechanism is introduced to capture the long-distance dependencies between circuit nodes, enhance the understanding and modeling ability of the test case generation model for the global structure of the circuit, and generate test cases that are more in line with the circuit characteristics. Heuristic rules and scoring functions are used to evaluate and screen the generated candidate test case sequences to ensure that the generated test cases meet the test requirements and constraints, and improve the effectiveness and executability of the test cases.
[0088] In an optional embodiment,
[0089] The candidate test case sequences are evaluated and screened by using preset screening rules and scoring functions to obtain test cases that meet the test requirements and constraints, including:
[0090] The test case screening process is modeled as a Markov decision process, the state space is defined as the candidate test case sequence to be screened, the action space is the screening decision of the test case, the reward function is the quality score of the test case, and a transition probability function is constructed to describe the state transition. A deep neural network is constructed as a value function network and a policy network. The value function network is used to estimate the state-action value function, and the policy network is used to generate the screening decision.
[0091] Initialize the screening rules and scoring function as the initial strategy, create a priority experience playback buffer, and set the priority sampling weight, screen the candidate test case sequence according to the current state and the initial strategy, obtain the screening decision and immediate reward, and calculate the test case coverage and fault detection rate;
[0092] The screening decision, the immediate reward, the test case coverage and the fault detection rate are stored in a priority experience replay buffer, and the sample priority is updated according to the temporal difference error, and importance sampling is performed from the priority experience replay buffer according to the priority sampling weight to obtain an experience data batch;
[0093] Calculate the target Q value using the state, action, reward and next state information in the experience data batch, and update the value function network by minimizing the mean square error between the predicted Q value and the target Q value; at the same time, calculate the policy gradient using the policy gradient method based on the experience data batch and the updated value function network, and update the policy network according to the policy gradient;
[0094] Based on the updated value function network and policy network, combined with the test case coverage and fault detection rate information in the empirical data batch, the parameters of the screening rules and scoring functions are updated, and the iteration is repeated until the preset termination condition is reached. The policy networks of multiple training stages are combined using an ensemble learning method to generate the final screening policy network;
[0095] The final screening strategy network is used to screen the candidate test case sequences, and the test case quality is evaluated based on the test case coverage and fault detection rate. The correlation and complementarity between test cases are considered at the same time, and test cases that meet the test requirements and constraints are obtained.
[0096] Exemplarily, the test case screening process is modeled as a Markov decision process, which can fully utilize the advantages of reinforcement learning, autonomously learn and optimize the test case screening strategy by continuously interacting with the environment and obtaining feedback. In this modeling process, the state space is defined as the sequence of candidate test cases to be screened, representing the various decision-making stages in the screening process; the action space is the screening decision of the test case, representing the screening behavior that the agent can take in each state; the reward function is the quality score of the test case, which is used to guide the agent to learn the optimal screening strategy; the transition probability function describes the transition relationship between states and characterizes the impact of the screening decision on the state transition.
[0097] In the construction of the value function network and the policy network, a deep neural network is used as a function approximator. The value function network is used to estimate the state-action value function, that is, the long-term expected return of taking a specific action in a given state. The network takes the state and action as input and outputs the corresponding Q value estimate. The policy network is used to generate screening decisions, that is, the probability distribution of actions to be taken in a given state. The network takes the state as input and outputs the probability distribution of actions. Through the coordinated optimization of the value function network and the policy network, continuous improvement of the strategy and improvement of the accuracy of value estimation can be achieved.
[0098] In the initial stage of the algorithm, screening rules and scoring functions are used as initial strategies to provide a good starting point for subsequent reinforcement learning. Screening rules can be designed based on prior knowledge, such as the importance and coverage of test cases; scoring functions quantitatively measure the quality of test cases, taking into account indicators such as defect detection capabilities and execution efficiency. At the same time, a priority experience playback buffer is created to store the experience data collected during the interaction between the agent and the environment, and priority sampling weights are set so that more valuable experience can be sampled and used first.
[0099] According to the current state and initial strategy, the candidate test case sequence is screened to obtain the corresponding screening decision and immediate reward. The screening decision indicates whether to keep or remove each test case, and the immediate reward is judged based on the quality of the screening decision, which can be calculated using indicators such as test case coverage and fault detection rate. Information such as screening decisions, immediate rewards, test case coverage, and fault detection rate are stored in the priority experience playback buffer, and the sample priority is updated according to the temporal difference error, so that the experience that is more helpful for strategy optimization can obtain a higher sampling probability.
[0100] Importance sampling is performed from the priority experience replay buffer according to the priority sampling weights to obtain batches of experience data for training and updating the neural network. The state, action, reward, and next state information in the experience data batch are used to calculate the target Q value through the temporal difference learning algorithm, and the mean square error between the predicted Q value and the target Q value is used as the loss function to optimize the parameters of the value function network so that it can more accurately estimate the state-action value function. At the same time, based on the experience data batch and the updated value function network, the policy gradient method is used to calculate the policy gradient, and the parameters of the policy network are updated according to the gradient information, so that the policy network can generate better screening decisions.
[0101] While the policy network is being updated, the parameters of the screening rules and scoring functions are adjusted and optimized based on the test case coverage and fault detection rate in the empirical data batches, so that they can better adapt to the changing test requirements and constraints. The above process is repeated through multiple rounds of iterations until the preset termination conditions are reached, such as convergence judgment or iteration number limit.
[0102] In order to further improve the generalization performance and robustness of the screening strategy, an ensemble learning method is used to combine the policy networks of multiple training stages to generate the final screening strategy network. Ensemble learning can reduce the risk of overfitting and improve the generalization ability of the strategy by comprehensively utilizing the prediction results of multiple learners. Finally, the obtained screening strategy network is used to screen the candidate test case sequence, and the test case quality is evaluated based on indicators such as test case coverage and fault detection rate. At the same time, the correlation and complementarity between test cases are considered to obtain a set of high-quality test cases that meet the test requirements and constraints.
[0103] In this embodiment, through Markov decision process modeling and reinforcement learning optimization, the test case screening strategy is adaptively learned and improved, the limitations of traditional heuristic rules are overcome, and the intelligent level of screening decision is improved. Using deep neural network as an approximator of value function and policy function, it has powerful function fitting and characterization capabilities, can effectively handle high-dimensional and nonlinear test case features, and improves the expression ability and generalization performance of the screening strategy. Combining immediate rewards and long-term returns, through temporal difference learning and policy gradient optimization, the continuous improvement of the strategy is achieved, so that the screening decision can take into account the local and global test effects, and improve the overall quality of the screening results. While optimizing the strategy, the screening rules and scoring functions are dynamically adjusted so that they can adapt to the changing test requirements and constraints, and improve the adaptability and flexibility of the screening method. Fully consider the correlation and complementarity between test cases, select a test case combination with high coverage, strong fault detection capability, and excellent execution efficiency, so as to maximize the test effect and improve the test input-output ratio under limited test resource conditions.
[0104] S102. Input the generated test case into a simulation engine that supports multiple fault models, set the fault type, injection location and trigger condition through the simulation engine, dynamically modify the circuit model behavior to simulate the impact of the fault on the normal operation of the circuit, generate simulation data, extract the feature representation of the simulation data, calculate the similarity between the simulation data and the reference data based on the extracted feature representation of the simulation data, and pre-locate the first candidate fault area in the simulation data.
[0105] Among them, the simulation engine can flexibly set different types of faults, such as struck-at faults, bridging faults, open faults, delay faults, etc., by modifying the parameters and structure of the circuit model. By injecting faults at different locations of the circuit and setting corresponding trigger conditions, such as input excitation, timing constraints, etc., the impact of faults on the normal function of the circuit is simulated. The simulation engine will generate a large amount of simulation data to record the output response of the circuit under fault conditions, such as voltage and current waveforms at key nodes. Feature extraction is performed on the simulation data to obtain high-dimensional feature vectors that characterize the fault behavior, such as statistical signatures, frequency domain features, etc. Then, by calculating the similarity between the simulation data features and the reference data features, candidate areas where faults may exist in the simulation data are pre-located.
[0106] In an optional embodiment,
[0107] The generated test cases are input into a simulation engine that supports multiple fault models. The fault type, injection location, and trigger condition are set through the simulation engine. The circuit model behavior is dynamically modified to simulate the impact of the fault on the normal operation of the circuit. Simulation data is generated, and feature representation of the simulation data is extracted. The similarity between the simulation data and the reference data is calculated based on the extracted feature representation of the simulation data. The first candidate fault area in the pre-positioned simulation data includes:
[0108] Setting a fault model library, the fault model library includes predefined fault types, adding parameter faults, delay faults and bridge faults to the fault model library, creating a user-defined fault model interface, and obtaining an extended fault model library;
[0109] Creating a fault injection language and script, and injecting faults in the extended fault model library into a pre-built multi-level circuit model using the fault injection language and script to obtain a fault injection model;
[0110] The generated test cases are input into the simulation engine that supports multiple fault models. Parallel computing, incremental evaluation and event-driven technologies are integrated into the simulation execution engine. The fault type to be injected is selected according to the extended fault model library. The fault injection location is determined in the multi-level circuit model. The fault triggering conditions are set. During the simulation process, when the triggering conditions are met, the fault injection model is simulated by changing the parameter values of the specified components or modules, adjusting the signal propagation delay, adding connections that should not exist between the specified nodes, or modifying the corresponding circuit behavior according to the predefined fault model to obtain simulation data reflecting the impact of the fault.
[0111] Extracting features from the obtained simulation data, processing the extracted features by using feature dimension reduction and transformation methods to generate an optimized feature set, constructing and training a similarity measurement model based on the optimized feature set to obtain a trained similarity measurement model, and calculating the similarity between the simulation data and the reference data by using the trained similarity measurement model;
[0112] A multi-objective optimization model is constructed, which comprehensively considers positioning accuracy, efficiency and coverage factors, and based on the similarity calculation results, identifies the fault area in each level of the multi-level circuit model, prioritizes the identified fault areas, considers the fault impact and the probability of occurrence, uses a heuristic algorithm to search in areas with a priority higher than a preset threshold, combines the search results of different levels, determines the cross-level fault propagation path through inter-level correlation analysis, and generates the first candidate fault area based on the search results and the inter-level correlation.
[0113] Exemplarily, a fault model library containing predefined fault types is first set up. On the basis of the original library, parameter faults, delay faults, and bridging faults are added, and an interface for user-defined fault models is provided to form an extended fault model library. Parameter faults refer to faults caused by circuit component parameter values deviating from the normal range, delay faults refer to faults caused by signal propagation delays exceeding a predetermined time limit, and bridging faults refer to connections between two nodes that should not be connected. The user-defined fault model interface allows users to flexibly add new fault types and models according to actual needs.
[0114] After building the extended fault model library, this solution creates a fault injection language and scripts to inject the fault models in the library into the pre-built multi-level circuit model. The multi-level circuit model includes models at different abstract levels such as system level, board level, chip level, RTL level and gate level, which can meet the simulation and fault analysis requirements at different stages. The fault injection language and scripts provide a flexible and efficient way to automatically complete the fault injection process and obtain the fault injection model by describing information such as fault type, injection location, trigger conditions, etc.
[0115] After the fault injection is completed, the generated test cases are input into the simulation engine that supports multiple fault models for simulation. The simulation execution engine integrates acceleration technologies such as parallel computing, incremental evaluation, and event-driven, which can significantly improve the simulation efficiency. According to the definition in the extended fault model library, the simulation engine selects the fault type to be injected and determines the fault injection location in the multi-level circuit model. At the same time, the fault trigger condition is set. When the simulation process meets the trigger condition, the fault injection is realized by changing the parameter value of the specified component or module, adjusting the signal propagation delay, adding connections between specified nodes, or modifying the circuit behavior. After the simulation is completed, simulation data reflecting the impact of the fault can be obtained.
[0116] In order to analyze simulation data intelligently, this solution uses feature engineering and machine learning technology. First, extract indicators that reflect fault characteristics from simulation data, such as voltage, current, delay, etc. of key nodes. Then, use feature dimension reduction and transformation methods, such as principal component analysis and independent component analysis, to remove redundant and noise features and generate an optimized feature set. Based on the optimized feature set, a similarity measurement model is constructed, and machine learning algorithms such as support vector machines, decision trees, and Bayesian networks are used for training to obtain a similarity measurement model with excellent performance. Using the trained model, the similarity between the current simulation data and the standard reference data can be calculated.
[0117] Based on the similarity calculation, this scheme constructs a multi-objective optimization model for searching and locating fault areas in each level of the multi-level circuit model. The optimization model comprehensively considers multiple objectives such as positioning accuracy, positioning efficiency and diagnostic coverage, and forms a single optimization objective function through weighted combination. According to the similarity analysis results, the fault areas with higher suspicion are identified in each layer of the circuit model, and the areas are prioritized according to factors such as the degree of fault impact and the probability of occurrence. In order to balance the search time and accuracy, the priority threshold is set, and heuristic search is performed preferentially in high-priority areas, such as genetic algorithm, particle swarm optimization, etc. The search results of different levels are comprehensively analyzed, and the propagation and impact of faults between levels are considered to determine the fault propagation path across levels. Finally, based on the search results and inter-level analysis, candidate areas for fault pre-location are generated.
[0118] The following is a specific example to illustrate the execution process of the above fault pre-location scheme. Suppose we want to pre-locate faults for a multi-level processor circuit, which includes four-level models: system level, board level, chip level, and gate level. First, according to the definition of the extended fault model library, select several typical fault types, such as parameter faults of the processor cache module, delay faults of the bus, bridge faults inside the chip, etc., and use the fault injection script to inject these faults into the processor model of the corresponding level. Then, generate a set of test cases and input them into the simulation engine that supports multiple faults to simulate the behavior of the processor under different faults. During the simulation process, when specific conditions are met, such as cache access failure, excessive bus delay, chip function abnormality, etc., the corresponding fault mechanism is triggered to introduce the fault impact. After the simulation is completed, the key signals of each level are extracted as the original features, and the optimized feature set is obtained through feature dimensionality reduction. Using the pre-trained similarity measurement model, the similarity between the simulation data and the reference data at the system level, board level, chip level, and gate level is calculated respectively. Based on the similarity results, several fault areas with the highest suspicion are identified at each level, such as cache modules, bus interfaces, and key paths inside the chip. By combining the analysis results at different levels, we can infer the cross-level propagation path of the fault within the processor, and finally lock the most suspicious area containing the root cause of the fault to complete the fault pre-location.
[0119] In this embodiment, the constructed extended fault model library supports a variety of common and customized fault types, improving fault coverage and modeling accuracy; fault injection language and scripts realize automated, precise and controllable fault injection, simplifying the fault simulation process; the simulation execution engine integrates acceleration technology, supports multi-fault concurrent simulation, greatly improving simulation efficiency and speed; machine learning technology is used to optimize feature extraction and similarity calculation, enhancing the ability to distinguish fault features and measurement accuracy; the multi-objective optimization model achieves a trade-off between accuracy, efficiency and coverage, improving fault search speed and accuracy; joint analysis and reasoning are performed on multi-level circuit models to achieve global cross-level fault location. It can significantly improve the efficiency, accuracy and intelligence level of fault pre-location under multi-fault models, and provide strong support for reliability design and fault diagnosis of complex circuit systems.
[0120] In an optional embodiment,
[0121] A multi-objective optimization model is constructed, wherein the multi-objective optimization model comprehensively considers positioning accuracy, efficiency and coverage factors, and identifies the fault area in each level of the multi-level circuit model based on the similarity calculation result. The steps before prioritizing the identified fault areas also include:
[0122] Determine the degree of fault impact based on the fault impact characteristics in the simulation data, and calculate the probability of fault occurrence based on historical fault data and current circuit operation status;
[0123] Fuzzy sets are defined for the fault impact degree and the fault occurrence probability respectively, each fuzzy set is described by a corresponding fuzzy membership function, and a fuzzy rule base is constructed at the same time, wherein the fuzzy rule base contains fuzzy rules describing the relationship between the fault impact degree and the fault occurrence probability;
[0124] Performing fuzzification processing on the fault impact degree and the fault occurrence probability to obtain the membership degree of the fault impact degree and the fault occurrence probability in their respective fuzzy sets, performing fuzzy reasoning based on the fuzzy rule base and the membership degree to obtain the fuzzy output of the comprehensive evaluation result, performing defuzzification processing on the fuzzy output of the comprehensive evaluation result to obtain the comprehensive evaluation score, and determining the search priority according to the comprehensive evaluation score;
[0125] Based on the similarity calculation results and the search priority, the fault area is identified in each level of the multi-level circuit model, and the identified fault areas are prioritized.
[0126] For example, first, the impact of the fault on system function and safety is determined by analyzing the fault impact characteristics in the simulation data, such as the performance degradation caused by the fault, the degree of key parameter deviation, etc. At the same time, the probability of the fault occurring is calculated by combining historical fault data and current circuit operation conditions, such as the degree of component aging and environmental stress levels.
[0127] In order to establish a nonlinear mapping relationship between the degree of fault impact and the probability of fault occurrence, fuzzy logic theory is introduced. According to the degree of fault impact and the probability of fault occurrence, corresponding fuzzy sets are defined, such as "minor impact", "medium impact", "serious impact", etc. Each fuzzy set is described by a fuzzy membership function. The fuzzy membership function can measure the degree to which a specific fault impact or fault probability value belongs to each fuzzy set, and its value range is [0, 1]. Commonly used fuzzy membership functions include triangle, trapezoid, Gaussian, etc.
[0128] Based on fuzzy sets, a fuzzy rule base is constructed to characterize the logical relationship between the degree of fault impact and the probability of fault occurrence. Fuzzy rules are usually expressed in the form of "IF-THEN", for example: "IF the degree of fault impact is serious AND the probability of fault occurrence is high THEN the comprehensive evaluation result is high risk". The fuzzy rule base contains a series of such language rules, covering various combinations of fault impact and fault probability.
[0129] In the comprehensive assessment process, the specific values of the fault impact degree and the fault probability are first fuzzified, that is, their membership in each fuzzy set is calculated according to their membership function. Then, based on the fuzzy rule base and the membership value, the fuzzy reasoning algorithm is used for reasoning calculation. Commonly used fuzzy reasoning algorithms include Mamdani reasoning and Sugeno reasoning. The core idea is to obtain the final output result through the trigger strength of each rule. The result of fuzzy reasoning is a fuzzy output, which represents the membership distribution of the comprehensive assessment results at each risk level.
[0130] In order to obtain a clear comprehensive evaluation score, the fuzzy output needs to be defuzzified. The purpose of defuzzification is to convert the fuzzy output into a clear value. Common defuzzification methods include the centroid method and the maximum membership method. According to the comprehensive evaluation score obtained by defuzzification, the search priority of the fault can be determined. The higher the score, the greater the impact of the fault and the higher the possibility of occurrence, and the higher the search priority should be given.
[0131] Finally, in each level of the multi-level circuit model, the fault areas with higher suspicion are identified based on the similarity calculation results and search priority. The similarity calculation evaluates the possibility of failure of each circuit module or component by comparing the difference between the fault characteristics and the standard pattern. The similarity calculation results are combined with the search priority to prioritize the identified fault areas, and prioritize the search and diagnosis of areas with greater impact and more likely to fail.
[0132] In this embodiment, by introducing fuzzy logic theory, a nonlinear mapping relationship between the degree of fault impact and the probability of fault occurrence can be established, which overcomes the limitations of hard threshold division in traditional methods and improves the rationality and accuracy of comprehensive evaluation. By using fuzzy sets and membership functions, the uncertainty and language description of the degree of fault impact and the probability of fault occurrence can be described more flexibly and meticulously, making the evaluation results closer to human thinking habits and judgment methods. Reasoning based on the fuzzy rule base can comprehensively consider different combinations of the degree of fault impact and the probability of fault occurrence, and obtain a comprehensive evaluation result that takes into account multiple factors, providing a more comprehensive and reliable basis for fault search. Combining the comprehensive evaluation results with similarity calculation, the fault area is identified and located at each level of the multi-level circuit model, realizing a top-down, step-by-step refinement fault diagnosis strategy, narrowing the search range and improving the diagnosis efficiency.
[0133] S103. Perform hardware testing on the integrated circuit, collect hardware test response data, align the collected hardware test response data with the simulation data, establish a mapping relationship between the software and hardware data, and based on the mapping relationship, use a spectral analysis method to calculate the difference between the hardware test data and the simulation data, locate the second candidate fault area in the hardware test data, combine the first candidate fault area and the second candidate fault area, determine the final fault area, and output fault diagnosis suggestions.
[0134] Among them, hardware testing uses special test equipment such as ATE (Automatic Test Equipment) to apply test stimuli to chip samples according to the generated test cases, and collects test response data such as voltage and current of each key node. Due to the difference between the simulation environment and the real hardware environment, it is necessary to align the collected hardware test data with the simulation data in time and space. In the time dimension, the two sets of data are aligned on the time axis through algorithms such as Dynamic Time Warping; in the spatial dimension, a one-to-one correspondence is established between the node signals in the hardware test data and the simulation node signals through signal name matching, hierarchical mapping, etc.
[0135] In an optional embodiment,
[0136] Perform hardware testing on the integrated circuit, collect hardware test response data, align the collected hardware test response data with the simulation data, establish a mapping relationship between the software and hardware data, calculate the difference between the hardware test data and the simulation data based on the mapping relationship, locate the second candidate fault area in the hardware test data, combine the first candidate fault area and the second candidate fault area, determine the final fault area, and output fault diagnosis suggestions including:
[0137] Performing hardware testing on the target integrated circuit, collecting test response data under the actual operating environment through the standard test interface in the circuit design, the test response data including the state of the trigger, the level signal of the node and the operating parameters;
[0138] The test response data is mapped one by one with the corresponding signals and nodes in the simulation data through the signal name and node hierarchy to obtain the spatially aligned test response data; the spatially aligned test response data and the simulation data are time-aligned, calibrated through timestamp comparison, and a complete mapping relationship between software and hardware data is established to obtain the target test response data;
[0139] Perform fast Fourier transform on the target test response data and the simulation data to obtain the frequency domain representation of the target test response data and the frequency domain representation of the simulation data, and use the distance metric to calculate the frequency component difference between the target test response data and the simulation data in the frequency domain, map the frequency component difference exceeding the preset threshold back to the time domain, determine the time interval in which the test response data and the simulation data are inconsistent, and obtain the second candidate fault area;
[0140] Calculate the spatial overlap between the second candidate fault area and the first candidate fault area, classify the candidate fault areas into high overlap areas and low overlap areas based on a preset overlap threshold, apply an adaptive clustering algorithm to the high overlap areas to generate fused candidate fault areas, and use a multi-factor weighted scoring model for the low overlap areas to calculate the comprehensive confidence of each area;
[0141] The fused candidate fault areas and the low-overlap areas with the highest comprehensive confidence are selected to construct an optimized set of candidate fault areas. Combined with the topological structure information and functional module division of the target integrated circuit, the graph theory analysis method is used to screen the optimized set of candidate fault areas, determine the final fault area, and output fault diagnosis suggestions.
[0142] For example, during the hardware test process, the test response data of the target integrated circuit in the actual operating environment is collected through the standard test interface reserved during circuit design, such as JTAG, boundary scan, etc. These test response data include the state of the internal trigger of the circuit, the level signal of the key node, and related operating parameters, such as power supply voltage, clock frequency, etc. The collected test response data can truly reflect the operating status and behavioral characteristics of the circuit, and provide an important basis for subsequent fault diagnosis.
[0143] In order to compare and analyze the test response data with the simulation data, it is necessary to first establish a mapping relationship between the two. By parsing the signal names and node hierarchies in the test response data and simulation data, they are matched one by one to achieve spatial alignment. For example, "module A / register B / trigger C" in the test response data and "module A / register B / trigger C" in the simulation data are mapped to the same node. Through spatial alignment, the test response data and simulation data are consistent in logical structure, laying the foundation for subsequent data fusion.
[0144] On the basis of spatial alignment, the test response data and simulation data are further time-aligned. Since the test response data comes from the actual circuit and the simulation data comes from the simulation model, there may be deviations between the timestamps of the two. By comparing the timestamps of the test response data and the simulation data, the time difference between them is calculated, and the time axis of the test response data is calibrated based on this, so that it is synchronized with the simulation data in the time dimension. After time alignment, the test response data and the simulation data are completely matched in both time and space dimensions, forming a complete mapping relationship between software and hardware data, and obtaining the target test response data for fault diagnosis.
[0145] In order to analyze the difference between the test response data and the simulation data from the frequency domain perspective, the target test response data and the simulation data are subjected to fast Fourier transform (FFT). FFT can convert the time domain signal into the frequency domain signal, revealing the frequency components and amplitude characteristics in the data. Through FFT, the frequency domain representation of the target test response data and the frequency domain representation of the simulation data are obtained, which contain the amplitude and phase information of each frequency component.
[0146] In the frequency domain, the difference between the target test response data and the simulation data on each frequency component is calculated using distance measurement methods such as Euclidean distance and Manhattan distance. The larger the difference value, the more significant the inconsistency on the frequency component. By setting a preset difference threshold, the frequency components that exceed the threshold are screened out, which represent significant differences between the test response data and the simulation data in the frequency domain.
[0147] Map the frequency domain difference back to the time domain to determine the time interval where the test response data is inconsistent with the simulation data. Specifically, for each frequency component that exceeds the difference threshold, convert it into a time domain signal through an inverse Fourier transform (IFFT) to locate the time interval corresponding to the frequency component. These time intervals represent the inconsistency between the test response data and the simulation data in the time domain, that is, the time range where the test response data is abnormal, and constitute the second candidate fault area.
[0148] In order to comprehensively consider multiple fault diagnosis information, the spatial position overlap between the second candidate fault area and the first candidate fault area (obtained by other diagnostic methods) is calculated. The spatial position overlap measures the similarity and proximity of the two areas in physical layout. The higher the overlap, the closer the two areas are in space, and the greater the possibility of a fault.
[0149] According to the preset overlap threshold, the candidate fault areas are divided into high overlap areas and low overlap areas. High overlap areas refer to the second candidate fault areas whose overlap with the first candidate fault area exceeds the threshold. They are highly consistent in space and the possibility of failure is very high. Adaptive clustering algorithms such as DBSCAN and OPTICS are applied to high overlap areas to aggregate adjacent areas into a larger fused candidate fault area, reduce the number of areas, and improve the targeted diagnosis.
[0150] For low overlap areas, that is, second candidate fault areas with low overlap with the first candidate fault area, a multi-factor weighted scoring model is used to calculate the comprehensive confidence of each area by comprehensively considering multiple factors such as the area, signal amplitude, and frequency characteristics of the area. The comprehensive confidence quantifies the possibility of each area as the location of the fault. The higher the confidence, the more likely it is to contain the fault point.
[0151] Based on the optimization of candidate fault regions, the fusion candidate fault regions and low-overlap regions with the highest comprehensive confidence are selected to construct an optimized candidate fault region set. This set summarizes fault regions from multiple diagnostic approaches and multiple analysis angles, covers multiple hot spots in the fault space, and provides a comprehensive candidate location for the final fault location.
[0152] In order to further improve the accuracy of fault location, the optimized candidate fault area set is analyzed by graph theory based on the topological structure information and functional module division of the target integrated circuit. The integrated circuit is abstracted as a directed graph, where nodes represent the functional modules of the circuit and edges represent the connection relationship between modules. The candidate fault area can be mapped to certain nodes or edges on the directed graph.
[0153] Through graph theory algorithms, such as the shortest path and critical path, the location and impact range of candidate fault areas in the circuit topology are analyzed. Considering factors such as the importance of the module where the fault area is located, the dependency relationship with other modules, and the signal propagation path, the fault impact and propagation risk of each candidate area are evaluated. Candidate areas with high impact and high propagation risk are more likely to contain critical fault points.
[0154] Based on the results of graph theory analysis, the optimized candidate fault area set is finally screened, and the area with the highest impact and the greatest propagation risk is selected as the final fault area. At the same time, according to the characteristics and circuit functions of the final fault area, the corresponding fault diagnosis suggestions are output, such as suspected fault types, circuit units that need to be checked, further testing and verification directions, etc., to provide guidance for fault location and repair.
[0155] In this embodiment, hardware test data and simulation data are comprehensively utilized to analyze fault characteristics from the perspectives of time domain and frequency domain, thereby improving the comprehensiveness and reliability of fault diagnosis. Space alignment and time alignment technologies are introduced to establish an accurate mapping relationship between test response data and simulation data, providing a basis for data fusion. Frequency domain analysis and difference measurement are used to quickly identify the inconsistent intervals between test response data and simulation data, and locate potential fault time windows. The candidate fault areas obtained by multiple diagnostic methods are comprehensively considered, and the pertinence and accuracy of fault location are improved through overlap analysis and cluster optimization. Graph theory analysis technology is used to evaluate the impact and propagation risk of candidate fault areas in the circuit topology, and the most likely fault location is selected. Fault diagnosis suggestions are output to provide intelligent guidance for fault location and repair, thereby improving diagnostic efficiency and success rate.
[0156] In an optional embodiment,
[0157] Apply graph theory analysis methods to screen the optimized candidate fault area set, determine the final fault area, and output fault diagnosis suggestions including:
[0158] Mapping the topological structure of the target integrated circuit into a graph structure, wherein nodes represent functional modules or submodules in the circuit, and edges represent signal connections or functional dependencies between modules;
[0159] The fault confidence and module function information of each node are obtained through circuit testing and simulation analysis, and stored in the graph structure as node attribute values. At the same time, the signal transmission delay and connection strength of each edge are calculated based on the circuit design parameters, and the calculation results are stored in the graph structure as edge weight values.
[0160] Using the graph structure that stores node attribute values and edge weight values, the shortest path algorithm is applied to calculate the shortest path from the node whose fault confidence exceeds the preset threshold to other nodes, and the preliminary fault propagation path is obtained;
[0161] Based on the maximum flow minimum cut algorithm, the minimum cut set of the graph structure is calculated to identify the modules and signal channels that have the greatest impact on the system function. By analyzing the connectivity between nodes and the weight values of the edges, the target path in the circuit is determined, and the nodes on the target path are formed into a target element set.
[0162] The regions in the optimized candidate fault region set are spatially matched with the nodes in the graph structure, and the correspondence between the candidate fault regions and the graph nodes is established. The candidate fault regions with the established correspondence are comprehensively scored, taking into account the fault confidence of the corresponding nodes, whether they are in the fault propagation path set, and whether they belong to the key element set;
[0163] The candidate fault areas are sorted in descending order according to the comprehensive scores, and the candidate fault area with the highest score and the corresponding node in the fault propagation path set is selected as the final fault area. Based on the final fault area, fault propagation path set and key element set, fault diagnosis suggestions are output.
[0164] Exemplarily, first, the topological structure of the target integrated circuit is mapped into a graph structure model, where the nodes of the graph represent the functional modules or submodules in the circuit, such as processors, memories, interface circuits, etc., and the edges represent the signal connections or functional dependencies between modules. Through this mapping, the physical structure and logical relationships of the circuit are abstracted into concepts in graph theory, which is convenient for subsequent analysis and calculation.
[0165] In the process of building the graph structure, it is necessary to obtain the fault confidence and module function information of each node, which are stored in the graph as the attribute value of the node. The fault confidence reflects the possibility of failure of the module and can be obtained through circuit testing, simulation analysis and other means. The module function information describes the role and behavioral characteristics of the module in the circuit. At the same time, based on circuit design parameters such as interconnect length and drive strength, the signal transmission delay and connection strength of each edge are calculated, and the calculation results are stored in the graph structure as the weight value of the edge. The signal transmission delay indicates the time required for the signal to propagate through the connection, and the connection strength reflects the degree of coupling between the two modules.
[0166] After establishing a graph structure containing node attributes and edge weights, the shortest path algorithm is applied to calculate the shortest path from nodes whose fault confidence exceeds the preset threshold to other nodes. The shortest path algorithm, such as the Dijkstra algorithm, can find the path with the smallest weight from the starting point to the end point in the graph, where the weight comprehensively considers the signal transmission delay and connection strength of the edge. By analyzing the shortest path, a preliminary fault propagation path can be obtained, that is, the process of the fault starting from a high-confidence node and spreading to other modules along the most likely path.
[0167] In order to further identify the key modules and signal channels that have the greatest impact on system functions, the maximum flow minimum cut algorithm is introduced. The maximum flow minimum cut algorithm is a classic algorithm in network flow theory, which is used to find the maximum flow and minimum cut set in the flow network. In the fault diagnosis scenario, by setting the source and sink points in the graph structure, the maximum flow minimum cut algorithm is used to calculate the minimum cut set of the graph. The minimum cut set represents the minimum set of edges that need to be removed to separate the source and sink points. These edges correspond to the key signal channels in the circuit. By analyzing the connectivity between nodes and the weight values of the edges, the critical path in the circuit can be determined, and the nodes on the critical path can be formed into a target element set.
[0168] In the process of optimizing the fault search range, the regions in the candidate fault region set are spatially matched with the nodes in the graph structure to establish a correspondence between the candidate fault regions and the graph nodes. Then, a comprehensive score is given to the candidate fault regions with established correspondence. The comprehensive score takes into account multiple factors, including the fault confidence of the corresponding node, whether it is located in the fault propagation path set, and whether it belongs to the key element set. The higher the fault confidence, the candidate region on the fault propagation path, and the candidate region belonging to the key element set, the higher its comprehensive score.
[0169] Finally, the candidate fault areas are sorted in descending order according to the comprehensive scores, and the candidate fault areas with the highest scores and corresponding nodes in the fault propagation path set are selected as the final fault areas. Based on the determined final fault areas, fault propagation path sets, and key element sets, fault diagnosis suggestions are generated to guide subsequent fault analysis and maintenance work.
[0170] In this embodiment, by mapping the integrated circuit topology to a graph structure, the key components and connection relationships of the circuit are abstracted, providing a convenient mathematical model and calculation framework for subsequent fault analysis. By introducing node attributes and edge weights, the fault probability, functional characteristics and electrical characteristics of the module are comprehensively considered, and the accuracy and comprehensiveness of fault location are improved. By applying the shortest path algorithm, the most likely propagation path of the fault starting from the high-confidence node can be quickly found, which narrows the fault search range and improves the diagnosis efficiency. By using the maximum flow minimum cut algorithm, the key modules and signal channels that have the greatest impact on the system function are automatically identified, providing key focus objects for fault analysis and reducing the workload of blind search. Through the spatial matching and comprehensive scoring of candidate fault areas and graph nodes, the optimal sorting of candidate areas is achieved, ensuring that the most suspicious fault areas are searched and analyzed first, and improving the accuracy and efficiency of diagnosis.
[0171] Figure 2 FIG. 1 is a schematic diagram of the structure of an integrated circuit test system according to an embodiment of the present invention. Figure 2 As shown, the system comprises:
[0172] The first unit is used to construct a knowledge graph in the field of integrated circuit testing. According to the knowledge graph in the field of integrated circuit testing, the circuit topology is extracted from the schematic diagram of the circuit to be tested, the functional modules in the circuit are identified, the input and output characteristics of the functional modules are extracted, and the circuit knowledge representation is formed. The circuit knowledge representation is subjected to graph embedding learning, the nodes and edges in the circuit diagram are embedded into a low-dimensional vector space through a graph neural network, the structure and function representation of the circuit are extracted using a message passing mechanism, and the extracted structure and function representation are input as feature vectors into a pre-built test case generation model to automatically generate test cases that meet the test requirements and constraints.
[0173] The second unit is used to input the generated test case into a simulation engine that supports multiple fault models, set the fault type, injection position and trigger condition through the simulation engine, dynamically modify the circuit model behavior to simulate the impact of the fault on the normal operation of the circuit, generate simulation data, extract the feature representation of the simulation data, calculate the similarity between the simulation data and the reference data according to the extracted feature representation of the simulation data, and pre-locate the first candidate fault area in the simulation data;
[0174] The third unit is used to perform hardware testing on the integrated circuit, collect hardware test response data, align the collected hardware test response data with the simulation data, establish a mapping relationship between the software and hardware data, and based on the mapping relationship, use a spectral analysis method to calculate the difference between the hardware test data and the simulation data, locate the second candidate fault area in the hardware test data, combine the first candidate fault area and the second candidate fault area, determine the final fault area, and output a fault diagnosis suggestion.
[0175] According to a third aspect of the embodiments of the present invention,
[0176] An electronic device is provided, comprising:
[0177] processor;
[0178] a memory for storing processor-executable instructions;
[0179] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.
[0180] According to a fourth aspect of the embodiments of the present invention,
[0181] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.
[0182] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.
[0183] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. An integrated circuit testing method, characterized in that: include: Construct a knowledge graph in the field of integrated circuit testing, extract the circuit topology from the schematic diagram of the circuit to be tested based on the knowledge graph in the field of integrated circuit testing, identify the functional modules in the circuit, extract the input and output characteristics of the functional modules, form a circuit knowledge representation, perform graph embedding learning on the circuit knowledge representation, embed the nodes and edges in the circuit diagram into a low-dimensional vector space through a graph neural network, extract the structure and function representation of the circuit using a message passing mechanism, input the extracted structure and function representation as a feature vector into a pre-built test case generation model, and automatically generate test cases that meet the test requirements and constraints; Input the generated test case into a simulation engine that supports multiple fault models, set the fault type, injection position and trigger condition through the simulation engine, dynamically modify the circuit model behavior to simulate the impact of the fault on the normal operation of the circuit, generate simulation data, extract the feature representation of the simulation data, calculate the similarity between the simulation data and the reference data based on the extracted feature representation of the simulation data, and pre-locate the first candidate fault area in the simulation data; Perform hardware testing on the integrated circuit, collect hardware test response data, align the collected hardware test response data with the simulation data, establish a mapping relationship between the software and hardware data, and based on the mapping relationship, use a spectral analysis method to calculate the difference between the hardware test data and the simulation data, locate the second candidate fault area in the hardware test data, combine the first candidate fault area and the second candidate fault area, determine the final fault area, and output fault diagnosis suggestions.
2. The method according to claim 1, characterized in that The nodes and edges in the circuit diagram are embedded into a low-dimensional vector space through a graph neural network, and the structure and function representation of the circuit are extracted using a message passing mechanism. The extracted structure and function representation are input as feature vectors into a pre-built test case generation model to automatically generate test cases that meet the test requirements and constraints, including: The circuit is converted into a graph structure representation, where nodes represent circuit components and edges represent the connection relationship between components. Each node in the circuit is represented as a node feature and aggregated with neighbor node information. A message passing mechanism is used to enable each node to interact with its neighboring nodes to generate an embedded representation of the node. Based on the embedded representation of the node, a global representation feature of the circuit is formed, and a shared layer for circuit feature extraction is constructed. On top of the shared layer, a task-specific layer is constructed, wherein for the fault detection task, a classification network based on a graph neural network is used to process the global representation features of the circuit; for the performance verification task, a regression network is used to process the global representation features of the circuit; for the power consumption optimization task, a convolutional neural network or a recursive neural network is used to process the global representation features of the circuit to obtain the output results of each task; According to the loss function of each task, a multi-task loss function is constructed, and the multi-task loss function is optimized at the same time to obtain the loss value of each task, and the loss gradient of each task is calculated according to the loss value of each task, and the weight of each task is dynamically adjusted based on the loss gradient, and the multi-task loss function is updated according to the adjusted task weight, and the shared layer and the task-specific layer are jointly trained; In the process of the joint training, a multi-task alternating optimization strategy is adopted, different tasks are set as dominant tasks respectively, the loss function of the dominant task is optimized, and the parameters of the task-specific layer corresponding to the shared layer and the dominant task are updated until all tasks are trained in sequence, thereby obtaining a trained test case generation model; In the encoder of the trained test case generation model, a graph attention layer is used to aggregate the node features output by the shared layer to capture the global structural information of the circuit. At the same time, in the decoder, decoding is performed according to the output of the encoder and the test requirements to generate multiple candidate test case sequences. The candidate test case sequences are evaluated and screened through preset screening rules and scoring functions to obtain test cases that meet the test requirements and constraints.
3. The method according to claim 2, characterized in that The candidate test case sequences are evaluated and screened by using preset screening rules and scoring functions to obtain test cases that meet the test requirements and constraints, including: The test case screening process is modeled as a Markov decision process, the state space is defined as the candidate test case sequence to be screened, the action space is the screening decision of the test case, the reward function is the quality score of the test case, and a transition probability function is constructed to describe the state transition. A deep neural network is constructed as a value function network and a policy network. The value function network is used to estimate the state-action value function, and the policy network is used to generate the screening decision. Initialize the screening rules and scoring function as the initial strategy, create a priority experience playback buffer, and set the priority sampling weight, screen the candidate test case sequence according to the current state and the initial strategy, obtain the screening decision and immediate reward, and calculate the test case coverage and fault detection rate; The screening decision, the immediate reward, the test case coverage and the fault detection rate are stored in a priority experience replay buffer, and the sample priority is updated according to the temporal difference error, and importance sampling is performed from the priority experience replay buffer according to the priority sampling weight to obtain an experience data batch; Calculate the target Q value using the state, action, reward and next state information in the experience data batch, and update the value function network by minimizing the mean square error between the predicted Q value and the target Q value; at the same time, calculate the policy gradient using the policy gradient method based on the experience data batch and the updated value function network, and update the policy network according to the policy gradient; Based on the updated value function network and policy network, combined with the test case coverage and fault detection rate information in the empirical data batch, the parameters of the screening rules and scoring functions are updated, and the iteration is repeated until the preset termination condition is reached. The policy networks of multiple training stages are combined using an ensemble learning method to generate the final screening policy network; The final screening strategy network is used to screen the candidate test case sequences, and the test case quality is evaluated based on the test case coverage and fault detection rate. The correlation and complementarity between test cases are considered at the same time, and test cases that meet the test requirements and constraints are obtained.
4. The method according to claim 1, characterized in that: The generated test cases are input into a simulation engine that supports multiple fault models. The fault type, injection location, and trigger condition are set through the simulation engine. The circuit model behavior is dynamically modified to simulate the impact of the fault on the normal operation of the circuit. Simulation data is generated, and feature representation of the simulation data is extracted. The similarity between the simulation data and the reference data is calculated based on the extracted feature representation of the simulation data. The first candidate fault area in the pre-positioned simulation data includes: Setting a fault model library, the fault model library includes predefined fault types, adding parameter faults, delay faults and bridge faults to the fault model library, creating a user-defined fault model interface, and obtaining an extended fault model library; Creating a fault injection language and script, and injecting faults in the extended fault model library into a pre-built multi-level circuit model using the fault injection language and script to obtain a fault injection model; The generated test cases are input into the simulation engine that supports multiple fault models. Parallel computing, incremental evaluation and event-driven technologies are integrated into the simulation execution engine. The fault type to be injected is selected according to the extended fault model library. The fault injection location is determined in the multi-level circuit model. The fault triggering conditions are set. During the simulation process, when the triggering conditions are met, the fault injection model is simulated by changing the parameter values of the specified components or modules, adjusting the signal propagation delay, adding connections that should not exist between the specified nodes, or modifying the corresponding circuit behavior according to the predefined fault model to obtain simulation data reflecting the impact of the fault. Extracting features from the obtained simulation data, processing the extracted features by using feature dimension reduction and transformation methods to generate an optimized feature set, constructing and training a similarity measurement model based on the optimized feature set to obtain a trained similarity measurement model, and calculating the similarity between the simulation data and the reference data by using the trained similarity measurement model; A multi-objective optimization model is constructed, which comprehensively considers positioning accuracy, efficiency and coverage factors, and based on the similarity calculation results, identifies the fault area in each level of the multi-level circuit model, prioritizes the identified fault areas, considers the fault impact and the probability of occurrence, uses a heuristic algorithm to search in areas with a priority higher than a preset threshold, combines the search results of different levels, determines the cross-level fault propagation path through inter-level correlation analysis, and generates the first candidate fault area based on the search results and the inter-level correlation.
5. The method according to claim 4, characterized in that A multi-objective optimization model is constructed, wherein the multi-objective optimization model comprehensively considers positioning accuracy, efficiency and coverage factors, and identifies the fault area in each level of the multi-level circuit model based on the similarity calculation result. The steps before prioritizing the identified fault areas also include: Determine the degree of fault impact based on the fault impact characteristics in the simulation data, and calculate the probability of fault occurrence based on historical fault data and current circuit operation status; Fuzzy sets are defined for the fault impact degree and the fault occurrence probability respectively, each fuzzy set is described by a corresponding fuzzy membership function, and a fuzzy rule base is constructed at the same time, wherein the fuzzy rule base contains fuzzy rules describing the relationship between the fault impact degree and the fault occurrence probability; Performing fuzzification processing on the fault impact degree and the fault occurrence probability to obtain the membership degree of the fault impact degree and the fault occurrence probability in their respective fuzzy sets, performing fuzzy reasoning based on the fuzzy rule base and the membership degree to obtain the fuzzy output of the comprehensive evaluation result, performing defuzzification processing on the fuzzy output of the comprehensive evaluation result to obtain the comprehensive evaluation score, and determining the search priority according to the comprehensive evaluation score; Based on the similarity calculation results and the search priority, the fault area is identified in each level of the multi-level circuit model, and the identified fault areas are prioritized.
6. The method according to claim 1, characterized in that Perform hardware testing on the integrated circuit, collect hardware test response data, align the collected hardware test response data with the simulation data, establish a mapping relationship between the software and hardware data, calculate the difference between the hardware test data and the simulation data based on the mapping relationship, locate the second candidate fault area in the hardware test data, combine the first candidate fault area and the second candidate fault area, determine the final fault area, and output fault diagnosis suggestions including: Performing hardware testing on the target integrated circuit, collecting test response data under the actual operating environment through the standard test interface in the circuit design, the test response data including the state of the trigger, the level signal of the node and the operating parameters; The test response data is mapped one by one with the corresponding signals and nodes in the simulation data through the signal name and node hierarchy to obtain the spatially aligned test response data; the spatially aligned test response data and the simulation data are time-aligned, calibrated through timestamp comparison, and a complete mapping relationship between software and hardware data is established to obtain the target test response data; Perform fast Fourier transform on the target test response data and the simulation data to obtain the frequency domain representation of the target test response data and the frequency domain representation of the simulation data, and use the distance metric to calculate the frequency component difference between the target test response data and the simulation data in the frequency domain, map the frequency component difference exceeding the preset threshold back to the time domain, determine the time interval in which the test response data and the simulation data are inconsistent, and obtain the second candidate fault area; Calculate the spatial overlap between the second candidate fault area and the first candidate fault area, classify the candidate fault areas into high overlap areas and low overlap areas based on a preset overlap threshold, apply an adaptive clustering algorithm to the high overlap areas to generate fused candidate fault areas, and use a multi-factor weighted scoring model for the low overlap areas to calculate the comprehensive confidence of each area; The fused candidate fault areas and the low-overlap areas with the highest comprehensive confidence are selected to construct an optimized set of candidate fault areas. Combined with the topological structure information and functional module division of the target integrated circuit, the graph theory analysis method is used to screen the optimized set of candidate fault areas, determine the final fault area, and output fault diagnosis suggestions.
7. The method according to claim 6, characterized in that Apply graph theory analysis methods to screen the optimized candidate fault area set, determine the final fault area, and output fault diagnosis suggestions including: Mapping the topological structure of the target integrated circuit into a graph structure, wherein nodes represent functional modules or submodules in the circuit, and edges represent signal connections or functional dependencies between modules; The fault confidence and module function information of each node are obtained through circuit testing and simulation analysis, and stored in the graph structure as node attribute values. At the same time, the signal transmission delay and connection strength of each edge are calculated based on the circuit design parameters, and the calculation results are stored in the graph structure as edge weight values. Using the graph structure that stores node attribute values and edge weight values, the shortest path algorithm is applied to calculate the shortest path from the node whose fault confidence exceeds the preset threshold to other nodes, and the preliminary fault propagation path is obtained; Based on the maximum flow minimum cut algorithm, the minimum cut set of the graph structure is calculated to identify the modules and signal channels that have the greatest impact on the system function. By analyzing the connectivity between nodes and the weight values of the edges, the target path in the circuit is determined, and the nodes on the target path are formed into a target element set. The regions in the optimized candidate fault region set are spatially matched with the nodes in the graph structure, and the correspondence between the candidate fault regions and the graph nodes is established. The candidate fault regions with the established correspondence are comprehensively scored, taking into account the fault confidence of the corresponding nodes, whether they are in the fault propagation path set, and whether they belong to the key element set; The candidate fault areas are sorted in descending order according to the comprehensive scores, and the candidate fault area with the highest score and the corresponding node in the fault propagation path set is selected as the final fault area. Based on the final fault area, fault propagation path set and key element set, fault diagnosis suggestions are output.
8. An integrated circuit testing system, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to construct a knowledge graph in the field of integrated circuit testing. According to the knowledge graph in the field of integrated circuit testing, the circuit topology is extracted from the schematic diagram of the circuit to be tested, the functional modules in the circuit are identified, the input and output characteristics of the functional modules are extracted, and the circuit knowledge representation is formed. The circuit knowledge representation is subjected to graph embedding learning, the nodes and edges in the circuit diagram are embedded into a low-dimensional vector space through a graph neural network, the structure and function representation of the circuit are extracted using a message passing mechanism, and the extracted structure and function representation are input as feature vectors into a pre-built test case generation model to automatically generate test cases that meet the test requirements and constraints. The second unit is used to input the generated test case into a simulation engine that supports multiple fault models, set the fault type, injection position and trigger condition through the simulation engine, dynamically modify the circuit model behavior to simulate the impact of the fault on the normal operation of the circuit, generate simulation data, extract the feature representation of the simulation data, calculate the similarity between the simulation data and the reference data according to the extracted feature representation of the simulation data, and pre-locate the first candidate fault area in the simulation data; The third unit is used to perform hardware testing on the integrated circuit, collect hardware test response data, align the collected hardware test response data with the simulation data, establish a mapping relationship between the software and hardware data, and based on the mapping relationship, use a spectral analysis method to calculate the difference between the hardware test data and the simulation data, locate the second candidate fault area in the hardware test data, combine the first candidate fault area and the second candidate fault area, determine the final fault area, and output a fault diagnosis suggestion.
9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.
10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.
