Reinforcement learning-based integrated circuit test parameter adaptive optimization method and system

Abstract

The application provides a kind of integrated circuit test parameter adaptive optimization method and system based on reinforcement learning, it is related to integrated circuit test technical field, including: collection test parameter and result, extract features and evaluate performance to build test state space.Parameter adjustment distribution is obtained by analyzing state transition in test state space, reward matrix is constructed by combining performance evaluation and calculating state value, and state action mapping table is generated.Candidate actions are obtained by querying mapping table according to current state, expected performance is evaluated and screened, and optimal adjustment instruction is generated after resource constraint correction and issued to test equipment.The application can realize intelligent dynamic optimization of test parameters, and improve test efficiency and coverage.

Description

Technical Field

[0001] This invention relates to the field of integrated circuit testing technology, and in particular to an adaptive optimization method and system for integrated circuit testing parameters based on reinforcement learning. Background Technology

[0002] In the field of integrated circuit testing, the setting of test parameters has a decisive impact on test efficiency, cost, and chip quality assessment. Traditional test parameter optimization methods mainly rely on engineers' experience and fixed test procedures, typically using static parameter settings based on design specifications or fine-tuning based on linear regression models built from historical test data.

[0003] Existing practices can guarantee basic test coverage to a certain extent, but they lack the ability to adapt to dynamic changes during the testing process. Static parameter settings cannot respond to fluctuations in test equipment status, changes in environmental noise, and process deviations of the chip under test itself, resulting in unstable test results and a tendency to overtest or undertest. While linear models based on historical data can be adjusted to a limited extent, the adjustment strategy is rigid and it is difficult to handle the complex nonlinear relationship between test parameters and test performance, making it impossible to achieve dynamic optimal allocation when test resources are limited.

[0004] Static strategies suffer from lag in parameter adjustment when dealing with novel or complex integrated circuits, forcing longer testing cycles. Fixed models cannot fully utilize real-time test feedback information and may continue to use suboptimal parameters, leading to insufficient utilization of test equipment and inadequate test coverage. Therefore, there is an urgent need for a method that can intelligently optimize test parameters to improve test efficiency and resource utilization. Summary of the Invention

[0005] This invention provides a method and system for adaptive optimization of integrated circuit test parameters based on reinforcement learning, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of this invention provides an adaptive optimization method for integrated circuit test parameters based on reinforcement learning, comprising:

[0007] The test parameter sequence and test result sequence of the integrated circuit test equipment are collected. The time domain features and frequency domain features of the test parameter sequence are extracted and concatenated to obtain the test feature vector. The test pass rate and test coverage are calculated based on the test result sequence, and the test performance evaluation value is solved. The test feature vector and the test performance evaluation value are correlated and mapped to establish the test state space.

[0008] The test state space is divided into states to construct a state transition sequence and the test parameter difference between adjacent states is calculated. Based on the test parameter difference, the parameter adjustment magnitude is solved and statistical analysis is performed to obtain the adjustment distribution characteristics. Based on the adjustment distribution characteristics and the test performance evaluation value, a state transition reward matrix is ​​constructed and value propagation calculation is performed to obtain the state value distribution. Based on the state value distribution, the mapping weight from different states to parameter adjustment actions in the test state space is solved to obtain the state action mapping table.

[0009] The current test status of the integrated circuit is obtained, and a set of candidate adjustment actions is obtained by querying the state action mapping table. The expected test performance corresponding to each candidate adjustment action in the set of candidate adjustment actions is calculated and the target adjustment action is selected. The target adjustment action is matched with the resource occupancy status of the current test environment and the target adjustment action is corrected to obtain the optimal parameter adjustment instruction and send it to the test equipment for execution.

[0010] In one alternative implementation,

[0011] The process involves collecting test parameter sequences and test result sequences from integrated circuit testing equipment, extracting time-domain and frequency-domain features from the test parameter sequences and concatenating them to obtain a test feature vector, calculating the test pass rate and test coverage based on the test result sequences, and obtaining a test performance evaluation value. Finally, the test feature vector and the test performance evaluation value are correlated and mapped to establish a test state space, including:

[0012] The test parameter sequence and test result sequence are collected from the integrated circuit test equipment for multiple test cycles. The test parameter sequence is segmented according to time order and the statistics in each segment are calculated to obtain time domain features. The test parameter sequence is subjected to spectrum transformation and the dominant frequency component and energy distribution are extracted to obtain frequency domain features. The time domain features and the frequency domain features are concatenated according to a preset dimension order to obtain a test feature vector.

[0013] The test pass rate is obtained by statistically analyzing the number of samples that passed the test in the test result sequence and the total number of samples, and calculating the ratio. The test coverage rate is obtained by identifying the number of test item types in the test result sequence and combining it with the preset total number of test item types. The test pass rate and the test coverage rate are assigned weights and weighted summed to obtain a comprehensive evaluation value. A timeliness penalty is calculated based on the comprehensive evaluation value and preset timeliness constraints. The test performance evaluation value is obtained by solving based on the timeliness penalty and the comprehensive evaluation value.

[0014] The test feature vectors are clustered to obtain multiple feature clusters and assigned state labels. The correlation between the test feature vectors in each feature cluster and the test performance evaluation value is calculated and a state feature matrix is ​​constructed. Based on the state feature matrix and the test performance evaluation value, state boundary conditions are determined. The state labels and the state boundary conditions are combined to obtain the test state space.

[0015] In one alternative implementation,

[0016] Calculating the correlation between the test feature vectors within each feature cluster and the test performance evaluation value, and constructing a state feature matrix, then determining the state boundary conditions based on the state feature matrix and the test performance evaluation value includes:

[0017] Extract test feature vector sets from different feature clusters and calculate the corresponding center vectors. Calculate the deviation product between the test feature vectors in each feature cluster and the test performance evaluation value and solve for the linear correlation degree. Sort the test feature vectors in each feature cluster and the test performance evaluation value respectively and calculate the cumulative value of the difference in sorting position to obtain the nonlinear correlation degree. Perform a weighted sum of the linear correlation degree and the nonlinear correlation degree to obtain the comprehensive correlation degree.

[0018] The state identifier of each feature cluster and the center vector are arranged in order to form an initial state feature matrix. The comprehensive correlation degree is embedded as a weight parameter into the corresponding state identifier position in the initial state feature matrix, and the elements in the initial state feature matrix are scaled and adjusted to obtain the state feature matrix.

[0019] Extract the test feature vector distribution corresponding to each state identifier from the state feature matrix and calculate the feature distribution range. Statistically analyze the numerical distribution of the test performance evaluation value under different state identifiers and calculate the value range of the test performance evaluation value. Perform cross-matching between the feature distribution range and the value range of the test performance evaluation value and identify overlapping and interval regions. Based on the overlapping and interval regions, determine the critical conditions for state switching to obtain the state boundary conditions.

[0020] In one alternative implementation,

[0021] The test state space is divided into states to construct a state transition sequence, and the test parameter differences between adjacent states are calculated. Based on the test parameter differences, the parameter adjustment magnitude is solved, and statistical analysis is performed to obtain the adjustment distribution characteristics, including:

[0022] Based on the pre-acquired state boundary conditions, the test feature vector in the test state space is divided into multiple state regions and assigned a state region identifier. The jump records between the state region identifiers in the pre-stored historical test data are extracted and arranged to obtain a state transition trajectory. The repeated state region identifier pairs in the state transition trajectory are counted and the transition frequency is calculated. Based on the transition frequency, a high-frequency transition path is determined and the state region identifiers are connected according to the transition direction to obtain a state transition sequence.

[0023] The test feature vectors corresponding to adjacent state region identifiers are extracted from the state transition sequence, and the numerical differences of each dimension are calculated to obtain the test parameter differences. The test performance evaluation values ​​corresponding to adjacent state region identifiers are extracted, and the numerical differences are calculated to obtain the change in test performance evaluation values. The ratio between the difference of each dimension in the test parameter differences and the change in test performance evaluation values ​​is calculated, and the influence weight of each dimension is obtained by scaling. The absolute value of the difference of each dimension in the test parameter differences is calculated, and the influence weights are combined to perform a weighted summation to obtain the parameter adjustment range.

[0024] The parameter adjustment ranges are summarized, sorted, and divided into intervals. The number of samples in each interval is counted to obtain the frequency distribution. The central tendency and dispersion measures of the frequency distribution are calculated, and the adjustment distribution characteristics are constructed.

[0025] In one alternative implementation,

[0026] Based on the adjustment distribution characteristics and the test performance evaluation value, a state transition reward matrix is ​​constructed and value propagation calculation is performed to obtain the state value distribution. Based on the state value distribution, the mapping weights from different states to parameter adjustment actions in the test state space are solved to obtain a state action mapping table, including:

[0027] The central tendency measure in the adjustment distribution characteristics is used as the expected adjustment range. The expected return value is calculated based on the expected adjustment range and the test performance evaluation value. The adjustment risk value is calculated based on the pre-acquired dispersion measure and combined with the expected return value to obtain the state transition reward value. All state region identifier pairs in the test state space are traversed and the corresponding state transition reward values ​​are filled to construct the state transition reward matrix.

[0028] Initialize the value estimate of each state region identifier, extract the state transition reward value corresponding to the target state region identifier from the state transition reward matrix and combine it with the set of subsequent state region identifiers, calculate the value estimate of different subsequent state region identifiers in the set of subsequent state region identifiers and solve for the cumulative value by combining the corresponding state transition reward value, update the value estimate of the target state region identifier based on the cumulative value, and repeat the update until the value estimate converges to obtain the state value distribution.

[0029] The estimated value of each state region identifier is extracted from the state value distribution, and the correlation between the value and the preset parameter adjustment action is calculated to obtain the mapping weight. A state action mapping table is constructed based on the mapping weight.

[0030] In one alternative implementation,

[0031] The current test status of the integrated circuit is obtained, and a set of candidate adjustment actions is obtained based on the state action mapping table. The expected test performance corresponding to each candidate adjustment action in the set of candidate adjustment actions is calculated, and the target adjustment action is selected, including:

[0032] The current test state of the integrated circuit is obtained and time-domain and frequency-domain features are extracted to construct the current state feature vector. The cosine similarity between the current state feature vector and the test feature vector corresponding to each state region identifier in the test state space is calculated. The state region identifier with the highest cosine similarity is selected as the matching state identifier. The mapping weight corresponding to each parameter adjustment action is queried in the state action mapping table. The parameter adjustment actions with the mapping weight greater than the preset weight threshold are extracted to construct a candidate adjustment action set.

[0033] The value estimate corresponding to the matching state identifier is extracted from the state value distribution as the benchmark value. The parameter adjustment amount corresponding to each candidate adjustment action in the candidate adjustment action set is determined and the simulated successor feature vector is calculated. The simulated successor feature vector is matched with the test feature vector corresponding to each state region identifier in the test state space to obtain the predicted successor state region identifier. The value estimate corresponding to each predicted successor state region identifier is extracted from the state value distribution and combined with the benchmark value to calculate the value increment. Based on the value increment and the state transition reward value corresponding to the predicted successor state region identifier, the expected test performance corresponding to each candidate adjustment action is solved, and the candidate adjustment action with the highest expected test performance is taken as the target adjustment action.

[0034] In one alternative implementation,

[0035] Constraint matching is performed based on the target adjustment action and the current resource occupancy status of the test environment, and the target adjustment action is corrected to obtain the optimal parameter adjustment command, which is then sent to the test equipment for execution, including:

[0036] Obtain the current resource usage status of the test environment and the resource load index of the test equipment, count the number of test tasks currently being executed and determine the concurrent task load, calculate the resource availability based on the resource load index and combine it with the concurrent task load to solve for the resource constraint strength;

[0037] Extract parameter adjustment amounts from the target adjustment action and calculate the resource consumption required to execute the target adjustment action. Calculate the resource gap and identify the resource bottleneck type based on the resource consumption and resource availability. Construct a constraint matching matrix based on the resource bottleneck type and the resource constraint strength and determine the constraint satisfaction degree corresponding to the target adjustment action.

[0038] Determine whether the constraint satisfaction is greater than a preset constraint threshold. If it is not greater, reduce the parameter adjustment amount in the target adjustment action based on the resource bottleneck type to obtain a corrected adjustment amount. Calculate the expected test performance corresponding to the corrected adjustment amount and balance it with the resource constraint strength to obtain a corrected target adjustment action. Iterate and verify the corrected target adjustment action and the constraint matching matrix until the constraint satisfaction is greater than the constraint threshold.

[0039] The verified target adjustment action is converted into a device-recognizable instruction format and encapsulated to obtain the optimal parameter adjustment instruction. The optimal parameter adjustment instruction is then sent to the test device for execution via a communication connection.

[0040] A second aspect of this invention provides an adaptive optimization system for integrated circuit test parameters based on reinforcement learning, comprising:

[0041] The test state modeling unit is used to collect the test parameter sequence and test result sequence of the integrated circuit test equipment, extract the time domain features and frequency domain features of the test parameter sequence and concatenate them to obtain the test feature vector, calculate the test pass rate and test coverage based on the test result sequence and solve the test performance evaluation value, and establish a test state space by associating and mapping the test feature vector and the test performance evaluation value.

[0042] The state value mapping unit is used to divide the test state space into states, construct a state transition sequence, calculate the test parameter difference between adjacent states, solve the parameter adjustment magnitude based on the test parameter difference, perform statistical analysis to obtain the adjustment distribution characteristics, construct a state transition reward matrix based on the adjustment distribution characteristics and the test performance evaluation value, perform value propagation calculation to obtain the state value distribution, and solve the mapping weight from different states to parameter adjustment actions in the test state space based on the state value distribution to obtain the state action mapping table.

[0043] The parameter optimization execution unit is used to obtain the current test status of the integrated circuit and query the candidate adjustment action set based on the status action mapping table, calculate the expected test performance corresponding to each candidate adjustment action in the candidate adjustment action set and filter to obtain the target adjustment action, perform constraint matching based on the target adjustment action and the resource occupancy status of the current test environment and correct the target adjustment action, obtain the optimal parameter adjustment instruction and send it to the test equipment for execution.

[0044] A third aspect of the present invention provides an electronic device, comprising:

[0045] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0046] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0047] In this invention, a test state space integrating time-frequency domain features and performance evaluation values ​​is constructed to comprehensively characterize the dynamic behavior of the test system, providing a precise and high-dimensional state representation foundation for intelligent decision-making. Based on the statistical analysis of state transition sequences and parameter differences, effective adjustment patterns in the historical optimization process can be accurately captured, thereby extracting statistically significant adjustment distribution characteristics. Combined with the reward matrix and value propagation mechanism constructed from performance evaluation values, the long-term optimization potential under different states can be quantified, generating precise mapping weights to guide parameter adjustments. Resource occupancy states are introduced as constraints to match and correct target actions, ensuring that optimization instructions are applied in real-world testing. The feasibility and execution efficiency in the test environment have achieved closed-loop optimization from state awareness and strategy generation to action execution. This enables test parameters to be autonomously adjusted based on system state and historical experience, effectively overcoming the limitations of fixed parameter strategies. Through continuous interaction and learning in the reinforcement learning framework, optimization experience can be continuously accumulated and the state-action mapping relationship can be updated, thereby gradually improving the accuracy and adaptability of parameter adjustment strategies. This reduces reliance on external expert experience, lowers the cost and subjective error risk of manual parameter tuning, and significantly improves the test pass rate and overall test efficiency, providing a reliable automated optimization solution for large-scale and highly complex testing of integrated circuits. Attached Figure Description

[0048] Figure 1 This is a flowchart illustrating the adaptive optimization method for integrated circuit test parameters based on reinforcement learning, as described in an embodiment of the present invention.

[0049] Figure 2 This is a flowchart illustrating the test state analysis and parameter adjustment process of the integrated circuit test parameter adaptive optimization method based on reinforcement learning, as described in an embodiment of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0052] Figure 1 This is a flowchart illustrating the adaptive optimization method for integrated circuit test parameters based on reinforcement learning, as described in an embodiment of the present invention. Figure 1As shown, the method includes:

[0053] The test parameter sequence and test result sequence of the integrated circuit test equipment are collected. The time domain features and frequency domain features of the test parameter sequence are extracted and concatenated to obtain the test feature vector. The test pass rate and test coverage are calculated based on the test result sequence, and the test performance evaluation value is solved. The test feature vector and the test performance evaluation value are correlated and mapped to establish the test state space.

[0054] The test state space is divided into states to construct a state transition sequence and the test parameter difference between adjacent states is calculated. Based on the test parameter difference, the parameter adjustment magnitude is solved and statistical analysis is performed to obtain the adjustment distribution characteristics. Based on the adjustment distribution characteristics and the test performance evaluation value, a state transition reward matrix is ​​constructed and value propagation calculation is performed to obtain the state value distribution. Based on the state value distribution, the mapping weight from different states to parameter adjustment actions in the test state space is solved to obtain the state action mapping table.

[0055] The current test status of the integrated circuit is obtained, and a set of candidate adjustment actions is obtained by querying the state action mapping table. The expected test performance corresponding to each candidate adjustment action in the set of candidate adjustment actions is calculated and the target adjustment action is selected. The target adjustment action is matched with the resource occupancy status of the current test environment and the target adjustment action is corrected to obtain the optimal parameter adjustment instruction and send it to the test equipment for execution.

[0056] In one alternative implementation,

[0057] The process involves collecting test parameter sequences and test result sequences from integrated circuit testing equipment, extracting time-domain and frequency-domain features from the test parameter sequences and concatenating them to obtain a test feature vector, calculating the test pass rate and test coverage based on the test result sequences, and obtaining a test performance evaluation value. Finally, the test feature vector and the test performance evaluation value are correlated and mapped to establish a test state space, including:

[0058] The test parameter sequence and test result sequence are collected from the integrated circuit test equipment for multiple test cycles. The test parameter sequence is segmented according to time order and the statistics in each segment are calculated to obtain time domain features. The test parameter sequence is subjected to spectrum transformation and the dominant frequency component and energy distribution are extracted to obtain frequency domain features. The time domain features and the frequency domain features are concatenated according to a preset dimension order to obtain a test feature vector.

[0059] The test pass rate is obtained by statistically analyzing the number of samples that passed the test in the test result sequence and the total number of samples, and calculating the ratio. The test coverage rate is obtained by identifying the number of test item types in the test result sequence and combining it with the preset total number of test item types. The test pass rate and the test coverage rate are assigned weights and weighted summed to obtain a comprehensive evaluation value. A timeliness penalty is calculated based on the comprehensive evaluation value and preset timeliness constraints. The test performance evaluation value is obtained by solving based on the timeliness penalty and the comprehensive evaluation value.

[0060] The test feature vectors are clustered to obtain multiple feature clusters and assigned state labels. The correlation between the test feature vectors in each feature cluster and the test performance evaluation value is calculated and a state feature matrix is ​​constructed. Based on the state feature matrix and the test performance evaluation value, state boundary conditions are determined. The state labels and the state boundary conditions are combined to obtain the test state space.

[0061] The test parameter sequence is acquired from the data acquisition interface of the integrated circuit test equipment for multiple consecutive test cycles. Each test cycle corresponds to a complete test process of the integrated circuit sample. The test parameter sequence includes, but is not limited to, key parameters such as test voltage, test frequency, signal rise time, hold time, setup time, and current threshold. These key parameters are arranged in timestamp order to form time series data. Simultaneously, the corresponding test result sequence is acquired, recording the pass or failure status of each test item, represented in binary form, where pass is marked as 1 and failure as 0. The data acquisition period needs to cover at least 72 consecutive hours of the test process to ensure that various state changes and anomalies during the test can be captured.

[0062] The collected test parameter sequences are segmented according to preset time windows. The length of the time window is determined based on the update frequency of the test parameters, typically set to 10 to 30 consecutive test samples. Within each time window, statistics such as the mean, standard deviation, maximum value, minimum value, rate of change, and trend of the test parameters are calculated. The mean reflects the central level of the parameter within that time period, the standard deviation reflects the degree of parameter fluctuation, the rate of change is calculated as the ratio of the parameter difference between adjacent test samples to the time interval, and the trend is obtained by using a linear fitting method to obtain the overall direction of parameter change within that segment. The statistics corresponding to all time windows are arranged in chronological order to form a time-domain feature vector. Time-domain features can effectively capture the dynamic changes of test parameters over time, reflecting the stability of the testing process and the continuity of parameter adjustments.

[0063] Frequency domain analysis was performed on the test parameter sequence, using Fast Fourier Transform (FFT) to convert the time-domain signal to the frequency domain. FFT decomposes the time-domain signal into a superposition of sinusoidal components of different frequencies, revealing the periodic variation patterns in the test parameter sequence. The resulting spectrum data is shown, with the horizontal axis representing frequency and the vertical axis representing the amplitude of the corresponding frequency component. The top 5 to 10 frequency components with the largest amplitudes were extracted from the spectrum data as dominant frequency components, reflecting the main periodic characteristics of the test parameter changes. The energy distribution of each frequency band was calculated, dividing the entire spectrum into low-frequency, mid-frequency, and high-frequency bands. The total energy of all frequency components within each band was calculated, and the energy distribution characterizes the frequency response of the test parameter changes. A higher proportion of energy in the low-frequency band usually indicates slower and more stable test parameter changes, while a higher proportion in the high-frequency band indicates rapid fluctuations or abrupt changes. The extracted dominant frequency components, their corresponding amplitudes, and the energy distribution proportions of each frequency band were combined to form a frequency domain feature vector.

[0064] The time-domain feature vector and the frequency-domain feature vector are concatenated according to a preset dimensional order. The concatenation order follows the principle of time-domain first, then frequency-domain, arranging the statistics in the time-domain features first, followed by the dominant frequency components and energy distribution parameters in the frequency-domain features. During the concatenation process, features of different dimensions need to be normalized to ensure that the features of each dimension are comparable in numerical scale. The normalization method uses max-min normalization, mapping the value of each feature dimension to the interval between 0 and 1. After concatenation and normalization, a complete test feature vector is obtained. The dimension of the test feature vector is typically between 50 and 200, depending on the number of time-domain statistics and the number of dominant frequency components in the frequency domain.

[0065] The test result sequence is analyzed by counting the number of samples that passed the test. This is done by traversing the sequence and accumulating the number of samples marked as 1. The total length of the test result sequence is then calculated to obtain the total number of samples. The pass rate is calculated as the ratio of the number of passed samples to the total number of samples. The pass rate ranges from 0 to 1, with a value closer to 1 indicating a higher proportion of passed samples. The test result sequence is then identified by identifying the types of test items, including functional tests, timing tests, voltage margin tests, and temperature stress tests. Metadata information from the test result sequence is parsed to extract the types of test items that have been executed and deduplicated to obtain the actual number of test item types. Based on the integrated circuit product's test specification document, the total number of preset test item types is obtained. This number represents the total number of test items required for complete test coverage. The ratio of the actual number of test item types to the total number of test item types is calculated to obtain the test coverage rate, which reflects the comprehensiveness and completeness of the test.

[0066] Test pass rate and test coverage are each assigned a weighting coefficient, with the weighting coefficient set based on the priority of the testing objectives. In scenarios prioritizing high yield, the weighting coefficient for test pass rate is set to 0.6 to 0.7, and the weighting coefficient for test coverage is set to 0.3 to 0.4. In scenarios prioritizing comprehensive verification, the weighting coefficient for test coverage can be increased to 0.5 to 0.6. A weighted summation is performed to obtain a comprehensive evaluation value, which is calculated as the product of test pass rate and its corresponding weighting coefficient, plus the product of test coverage and its corresponding weighting coefficient. The comprehensive evaluation value comprehensively reflects the quality and completeness of the testing.

[0067] The timeliness penalty is calculated based on preset timeliness constraints. These constraints include test cycle duration limits and test response time limits. The test cycle duration limit specifies the maximum allowed duration of a single test cycle, while the test response time limit specifies the maximum allowed delay from test initiation to result output. The actual test cycle duration and actual test response time are statistically analyzed and compared with the limits. When the actual value exceeds the limit, the excess percentage is calculated as the actual value minus the limit, divided by the limit. The timeliness penalty uses an exponential decay function, with the excess percentage as the input parameter; the penalty intensity increases exponentially as the excess percentage increases. The value of the timeliness penalty ranges from 0 to 1, with smaller values ​​indicating more severe penalties.

[0068] The overall evaluation value is multiplied by the timeliness penalty to obtain the test performance evaluation value. The test performance evaluation value comprehensively considers three dimensions: test quality, test completeness, and test efficiency, providing a holistic assessment of the current testing status. The test performance evaluation value ranges from 0 to 1, with higher values ​​indicating better test performance.

[0069] The test feature vectors are clustered using either K-means clustering or hierarchical clustering. Before clustering, the appropriate number of clusters needs to be determined. The elbow rule or silhouette coefficient method is used to evaluate the clustering effect under different numbers of clusters, selecting a number that maximizes intra-cluster similarity and inter-cluster dissimilarity. The number of clusters is typically set between 8 and 20. After executing the clustering algorithm, the test feature vector space is divided into multiple feature clusters, with each cluster containing test feature vectors exhibiting a similar feature distribution. A unique state identifier is assigned to each feature cluster, using integer numbers starting from 1 and incrementing sequentially.

[0070] Calculate the correlation between test feature vectors and corresponding test performance evaluation values ​​within each feature cluster. The correlation is measured using the Pearson correlation coefficient or Spearman rank correlation coefficient. For example, for a given feature cluster, extract all test feature vectors and their corresponding test performance evaluation values ​​within that cluster, and calculate the correlation coefficient between each dimension of the test feature vector and the test performance evaluation value. A larger absolute value of the correlation coefficient indicates a stronger correlation between that dimension and the test performance. Organize the correlation calculation results for each feature cluster into a matrix, where rows represent different feature clusters, columns represent the dimensions of the test feature vectors, and matrix elements are the corresponding correlation coefficients, forming a state feature matrix.

[0071] State boundary conditions are determined based on the state feature matrix and test performance evaluation values. These conditions define the boundary criteria between different states. For each feature cluster, the distribution range of test performance evaluation values ​​within that cluster is statistically analyzed, and the minimum, maximum, and mean values ​​are calculated. The mean of the test performance evaluation values ​​is used as the center value of that state, and the minimum and maximum values ​​are used to determine the effective range of that state. The state boundary conditions also include a maximum distance threshold between the test feature vector and the center point of the feature cluster. This maximum distance threshold is obtained by calculating the distances from all vectors within the feature cluster to the cluster center and taking the 95th percentile.

[0072] By combining state identifiers with state boundary conditions, a mapping relationship is established between state identifiers and state boundary conditions. Each state identifier corresponds to a set of boundary conditions, which include the range of test performance evaluation values, the coordinates of the center point of the test feature vector, and the maximum distance threshold. The set of all state identifiers and their corresponding boundary conditions forms a test state space, which is stored in the form of a data structure that supports quick querying of the state based on the test feature vector and the test performance evaluation value.

[0073] In one alternative implementation,

[0074] Calculating the correlation between the test feature vectors within each feature cluster and the test performance evaluation value, and constructing a state feature matrix, then determining the state boundary conditions based on the state feature matrix and the test performance evaluation value includes:

[0075] Extract test feature vector sets from different feature clusters and calculate the corresponding center vectors. Calculate the deviation product between the test feature vectors in each feature cluster and the test performance evaluation value and solve for the linear correlation degree. Sort the test feature vectors in each feature cluster and the test performance evaluation value respectively and calculate the cumulative value of the difference in sorting position to obtain the nonlinear correlation degree. Perform a weighted sum of the linear correlation degree and the nonlinear correlation degree to obtain the comprehensive correlation degree.

[0076] The state identifier of each feature cluster and the center vector are arranged in order to form an initial state feature matrix. The comprehensive correlation degree is embedded as a weight parameter into the corresponding state identifier position in the initial state feature matrix, and the elements in the initial state feature matrix are scaled and adjusted to obtain the state feature matrix.

[0077] Extract the test feature vector distribution corresponding to each state identifier from the state feature matrix and calculate the feature distribution range. Statistically analyze the numerical distribution of the test performance evaluation value under different state identifiers and calculate the value range of the test performance evaluation value. Perform cross-matching between the feature distribution range and the value range of the test performance evaluation value and identify overlapping and interval regions. Based on the overlapping and interval regions, determine the critical conditions for state switching to obtain the state boundary conditions.

[0078] After constructing the test state space, it needs to be finely divided. Specifically, a clustering algorithm is used to divide the test feature vectors into multiple feature clusters, each representing a set of states with similar test features. For each feature cluster, the correlation between internal test features and test performance needs to be quantified and expressed in a structured manner using a state feature matrix. Simultaneously, the boundary conditions between different states must be clearly defined to avoid ambiguity in state identification.

[0079] For calculating the correlation between test feature vectors and test performance evaluation values ​​within a feature cluster, a dual linear and nonlinear measurement mechanism is employed. (The last part, "for the 1st...", appears to be incomplete and requires further context.) For each feature cluster, extract all test feature vectors contained within it. ,in This represents the number of samples within the feature cluster. Calculate the center vector of the feature cluster. The center vector is obtained by taking the arithmetic mean of all test feature vectors within the cluster, specifically calculated as a vector composed of the mean values ​​of each dimension of the feature within the cluster. The center vector can represent the typical test feature state of this feature cluster and serves as a reference benchmark for subsequent state recognition.

[0080] The linear correlation degree is calculated based on the co-variation relationship between the eigenvectors and the test performance evaluation values. For each test eigenvector within a feature cluster... Extract its corresponding test performance evaluation value. The calculation involves multiplying the deviations of each dimension of the test feature vector from the test performance evaluation value. Specifically, this is done by subtracting the mean of that dimension within the cluster from the value of each dimension of the test feature vector, and simultaneously subtracting the mean of the test performance evaluation values ​​within the cluster from the test performance evaluation value. The products of the deviations for each dimension are then summed. (Linear correlation degree) It reflects the strength of the direct linear correlation between test features and test performance. The larger the value, the stronger the linear correlation between changes in test parameters within that feature cluster and improvements in test performance.

[0081] The calculation of nonlinear correlation is achieved through ranking consistency analysis. Test feature vectors within a feature cluster are arranged according to a specific ranking rule, which can be based on the feature vector's magnitude, the value of a specific dimension, or a comprehensive score. Simultaneously, the corresponding test performance evaluation values ​​are also ranked from smallest to largest. For each sample, its position in the ranking of test feature vectors is recorded. and its position in the ranking of test performance evaluation values. Calculate the difference in sorting positions. The total ranking difference is obtained by summing the ranking position differences of all samples within the cluster, and then converted into a non-linear correlation degree through normalization. Nonlinear correlation can capture the monotonic relationship and nonlinear dependency pattern between test features and test performance. Even if the two do not satisfy a strict linear relationship, a high nonlinear correlation can be obtained as long as they maintain a consistent trend of change.

[0082] The overall correlation degree is obtained by weighted summation of linear and nonlinear correlation degrees. The weighting parameters are determined based on the statistical characteristics of historical test data. When the test system exhibits strong linear characteristics, higher weight is given to linear correlation; when complex nonlinear dependencies exist in the test process, the weight of nonlinear correlation is increased. Overall Correlation Degree ,in and The weighting coefficients for the linear and nonlinear components are respectively, satisfying... The overall correlation score provides a quantified importance score for each feature cluster, reflecting the degree of influence of that state category on test performance.

[0083] The state feature matrix is ​​constructed based on the organization and correlation information of feature clusters. Each feature cluster is assigned a unique state identifier, and they are arranged according to their sequence number or priority. Each row of the initial state feature matrix corresponds to a feature cluster, and the elements in the row include the state identifier and the components of each dimension of the cluster's central vector. The matrix's column structure includes a state number column and multiple feature dimension columns, with the number of feature dimension columns matching the dimensions of the test feature vectors. The initial state feature matrix provides the basic topological structure of the state space, but it does not yet incorporate the correlation information between states and test performance.

[0084] To embed the comprehensive correlation degree into the state feature matrix, the matrix undergoes weight enhancement processing. Specifically, a column is added to the initial state feature matrix to specifically store the comprehensive correlation degree values; this column is adjacent to the state label column. For the... For each status identifier, fill in the corresponding row in the Comprehensive Relevance column. Furthermore, the comprehensive correlation degree is used to scale and adjust the components of each dimension of the central vector. The scaling method is to multiply each component of the original central vector by the corresponding comprehensive correlation degree to obtain a weighted state feature representation. The final state feature matrix obtained after weight embedding and scaling adjustment not only contains the feature description of the state, but also incorporates the importance information of the state, forming a structured state knowledge representation.

[0085] Determining state boundary conditions requires clearly defining the separating boundaries between different state identifiers. The distribution of test feature vectors corresponding to each state identifier is extracted from the state feature matrix. This distribution is obtained by statistically analyzing the value ranges of all original test feature vectors contained within that state identifier. For the ... Each state, in each feature dimension Above, calculate the minimum value of all test feature vector components in this dimension. and maximum value This determines the range of characteristic distribution. The feature distribution range describes the coverage area of ​​the state in each feature dimension, constituting the occupied domain of the state in the feature space.

[0086] The statistical distribution of test performance evaluation values ​​under different state markers is analyzed. For each state marker, the test performance evaluation values ​​corresponding to all test samples under that state are collected, and the minimum test performance evaluation value is calculated. and maximum value The range of values ​​for the test performance evaluation is obtained. The range of test performance evaluation values ​​reflects the range of test performance levels that can be achieved under a given state, and is a quantitative description of the state's performance.

[0087] State boundary identification is achieved by cross-matching the feature distribution range with the range of test performance evaluation values. For two adjacent state identifiers... and In the feature space, the feature distribution ranges of two states are compared to identify overlapping and gap regions. Overlapping regions indicate that two states intersect in the feature dimension; the test feature vectors in these regions may be classified into different states, requiring the introduction of additional discrimination conditions. The discrimination conditions are based on the range of test performance evaluation values. When the test performance evaluation value corresponding to the test feature vector within the overlapping region falls within the range of state... When the test performance evaluation value falls within the specified range, it is classified as a state. Otherwise, it is classified as a state. The interval region represents a clear separation between two states in the feature space, and the boundary of this region directly serves as the critical condition for state transition.

[0088] The critical conditions for state switching are determined through a comprehensive analysis of overlapping and gap regions. For overlapping regions, the center position is calculated and used as the initial boundary for state division, then fine-tuned based on differences in test performance evaluation values. The fine-tuning rule is that if the mean test performance evaluation value of samples within the overlapping region is closer to the mean test performance evaluation value of a certain state, the boundary is moved towards the other state to reduce the possibility of misclassification. For gap regions, the boundary points of the gap are directly extracted as state boundaries. The final state boundary conditions are expressed in the form of a multidimensional hyperplane or piecewise function, clarifying how the test feature vector maps to specific state identifiers, providing an accurate basis for state identification. These boundary conditions are used in subsequent real-time testing to quickly determine the state identifier to which the current test state belongs, supporting the querying of the state-action mapping table and the generation of parameter adjustment decisions.

[0089] In one alternative implementation,

[0090] The test state space is divided into states to construct a state transition sequence, and the test parameter differences between adjacent states are calculated. Based on the test parameter differences, the parameter adjustment magnitude is solved, and statistical analysis is performed to obtain the adjustment distribution characteristics, including:

[0091] Based on the pre-acquired state boundary conditions, the test feature vector in the test state space is divided into multiple state regions and assigned a state region identifier. The jump records between the state region identifiers in the pre-stored historical test data are extracted and arranged to obtain a state transition trajectory. The repeated state region identifier pairs in the state transition trajectory are counted and the transition frequency is calculated. Based on the transition frequency, a high-frequency transition path is determined and the state region identifiers are connected according to the transition direction to obtain a state transition sequence.

[0092] The test feature vectors corresponding to adjacent state region identifiers are extracted from the state transition sequence, and the numerical differences of each dimension are calculated to obtain the test parameter differences. The test performance evaluation values ​​corresponding to adjacent state region identifiers are extracted, and the numerical differences are calculated to obtain the change in test performance evaluation values. The ratio between the difference of each dimension in the test parameter differences and the change in test performance evaluation values ​​is calculated, and the influence weight of each dimension is obtained by scaling. The absolute value of the difference of each dimension in the test parameter differences is calculated, and the influence weights are combined to perform a weighted summation to obtain the parameter adjustment range.

[0093] The parameter adjustment ranges are summarized, sorted, and divided into intervals. The number of samples in each interval is counted to obtain the frequency distribution. The central tendency and dispersion measures of the frequency distribution are calculated, and the adjustment distribution characteristics are constructed.

[0094] After establishing the test state space, a deep, structured analysis is required. State partitioning rules are retrieved from a pre-configured state boundary condition library. These rules include segmentation thresholds for each dimension of the test feature vector and region definition criteria. For each test feature vector in the test state space, its dimensions are read sequentially. The first dimension's value is compared with the set boundary threshold for that dimension to determine its corresponding value interval. The values ​​for the second, third, and last dimensions are then read, and interval assignments are determined for each. Once the interval assignments for all dimensions are determined, the test feature vector is assigned to a specific state region according to the multi-dimensional space partitioning rules, and a unique state region identifier is assigned. This identifier is represented in an encoded form, such as using a combination of letters and numbers, like S001, S002, etc. All test feature vectors in the test state space are traversed to complete the comprehensive partitioning of state regions, forming multiple state regions with clearly defined boundaries.

[0095] Access the historical test data storage module, which records a large amount of test data generated during past integrated circuit testing. Read each historical test record sequentially; each record contains the test time, test feature vector, and corresponding state region identifier. Extract the state region identifiers generated by the same test batch or the same test equipment in chronological order, arranging them into a time series to obtain the state transition trajectory. In the state transition trajectory, two consecutive state region identifiers form a state region identifier pair; for example, a jump from S001 to S003 forms the identifier pair (S001, S003). Scan all state transition trajectories, record all occurrences of state region identifier pairs, and establish a counter to accumulate the occurrence count of each identifier pair to obtain the transition frequency. Sort the transition frequencies from high to low, and select identifier pairs whose transition frequencies exceed a preset frequency threshold as high-frequency transition paths. Based on the sequential relationship of identifier pairs in the high-frequency transition paths, connect the state region identifiers according to the transition direction to form a directional state transition sequence. This sequence reflects common state evolution paths in the test state space.

[0096] Select two adjacent state regions from the constructed state transition sequence, denoted as state A and state B respectively. Search the test state space to find all test feature vectors belonging to state A, and calculate the mean of these vectors as the representative feature vector of state A. Obtain the representative feature vector of state B using the same method. Subtract the representative feature vector of state A from the representative feature vector of state A dimension by dimension to obtain the numerical difference in each dimension. Assume the test feature vector has... The dimension, then the The numerical difference of the dimension is ,in Indicates that state B is in the 1st... The numerical value of the dimension. Indicates that state A is in the th... The numerical values ​​of all dimensions. The difference vector of test parameters is obtained by combining the numerical differences of all dimensions. .

[0097] Extract the test performance evaluation values ​​corresponding to state A and state B from the stored test performance evaluation value data, and denot them as follows: and Calculate the change in the test performance evaluation value. For each dimension of the test parameter difference vector, the ratio between the difference in that dimension and the change in the test performance evaluation value is calculated. To avoid an excessively large ratio due to a small change in the test performance evaluation value, regularization is introduced. The original ratio is calculated in the form of , where For small positive numbers used for numerical stability. The absolute value of the original ratios across all dimensions is then calculated. Then, normalization is performed using... The method of calculating the first The influence weight of a dimension. The influence weight reflects the degree to which changes in the parameter of that dimension contribute to changes in test performance. The larger the value, the more significant the impact of the parameter of that dimension on test performance.

[0098] The absolute value of the difference in each dimension of the test parameter difference vector is calculated to obtain... The weighted difference is obtained by multiplying the absolute difference of each dimension by its corresponding influence weight. Sum the weighted differences across all dimensions to obtain the parameter adjustment magnitude. The parameter adjustment range comprehensively considers the overall degree of change of parameters in multiple dimensions and the differences in the impact of each dimension on test performance, and can quantitatively characterize the comprehensive intensity of parameter adjustment when transitioning from one state to another.

[0099] Repeat the above calculation process for all adjacent state region identifiers in the state transition sequence to obtain multiple parameter adjustment magnitude values. Arrange these values ​​in ascending order to determine the minimum and maximum values ​​of the value range. Based on the data distribution characteristics, divide the value range into several intervals using equal width or equal frequency methods. Assume the intervals are divided as follows: Given a set of intervals, each interval has definite upper and lower bounds. Iterate through all parameter adjustment values, determine which interval each value falls into, and increment the counter for the corresponding interval to obtain the sample size within each interval, forming a frequency distribution. The frequency distribution can be represented as... ,in Indicates the first Number of samples within each interval.

[0100] Statistical features are extracted from the frequency distribution. The arithmetic mean of all parameter adjustment values ​​is calculated as the mean in the measure of central tendency. The median is selected from the middle value in the sorted parameter adjustment value sequence; this indicator is robust to outliers. The variance of the parameter adjustment values ​​is calculated, reflecting the degree of dispersion of the data relative to the mean. The variance is calculated as follows: Where n is the total number of samples for parameter adjustment magnitude. For the first One sample, The mean is the average. The standard deviation is obtained by taking the square root of the variance. The mean, median, standard deviation, skewness coefficient, and kurtosis coefficient are calculated as measures of dispersion. Simultaneously, the skewness coefficient describes the symmetry of the distribution, and the kurtosis coefficient describes the sharpness of the distribution. These statistics—frequency distribution, mean, median, standard deviation, skewness coefficient, and kurtosis coefficient—are organized into a structured data format to form the adjusted distribution features. These adjusted distribution features comprehensively characterize the distribution patterns of parameter adjustment amplitudes, providing statistical basis for subsequent reward design and strategy optimization. This allows reinforcement learning algorithms to make decisions based on the actual distribution of historical parameter adjustments, avoiding parameter adjustments that exceed normal ranges and ensuring the stability and feasibility of parameter optimization.

[0101] In one alternative implementation,

[0102] Based on the adjustment distribution characteristics and the test performance evaluation value, a state transition reward matrix is ​​constructed and value propagation calculation is performed to obtain the state value distribution. Based on the state value distribution, the mapping weights from different states to parameter adjustment actions in the test state space are solved to obtain a state action mapping table, including:

[0103] The central tendency measure in the adjustment distribution characteristics is used as the expected adjustment range. The expected return value is calculated based on the expected adjustment range and the test performance evaluation value. The adjustment risk value is calculated based on the pre-acquired dispersion measure and combined with the expected return value to obtain the state transition reward value. All state region identifier pairs in the test state space are traversed and the corresponding state transition reward values ​​are filled to construct the state transition reward matrix.

[0104] Initialize the value estimate of each state region identifier, extract the state transition reward value corresponding to the target state region identifier from the state transition reward matrix and combine it with the set of subsequent state region identifiers, calculate the value estimate of different subsequent state region identifiers in the set of subsequent state region identifiers and solve for the cumulative value by combining the corresponding state transition reward value, update the value estimate of the target state region identifier based on the cumulative value, and repeat the update until the value estimate converges to obtain the state value distribution.

[0105] The estimated value of each state region identifier is extracted from the state value distribution, and the correlation between the value and the preset parameter adjustment action is calculated to obtain the mapping weight. A state action mapping table is constructed based on the mapping weight.

[0106] After obtaining the adjustment distribution characteristics, a value assessment mechanism for state transitions is established. A measure of central tendency is extracted from the adjustment distribution characteristics; this measure characterizes the typical magnitude of change of the test parameter during historical adjustments. The measure of central tendency can be obtained by calculating the mean or median of the adjustment magnitude; the choice of which statistic depends on the distribution pattern of the adjusted data. When the adjusted data exhibits a symmetrical distribution, the mean effectively reflects central tendency; when extreme adjustment values ​​exist, the median is more robust. The extracted measure of central tendency is defined as the expected adjustment magnitude, which represents the reasonable range of parameter adjustment during a specific state transition.

[0107] The expected return is calculated based on the expected adjustment range and the test performance evaluation value. The test performance evaluation value reflects the test quality level under the current state, while the expected adjustment range characterizes the typicality of parameter changes. The increment of the test performance evaluation value is correlated with the expected adjustment range to obtain the performance improvement per unit adjustment range. The calculation of the expected return needs to consider the difference in test performance evaluation values ​​before and after the state transition, and this difference is normalized with the expected adjustment range to eliminate the influence of different dimensions. When the performance improvement is significant and the adjustment range is moderate, the expected return is higher; when the performance improvement is limited or the adjustment range is too large, the expected return is correspondingly lower.

[0108] The adjustment risk value is calculated based on a pre-acquired dispersion metric. The dispersion metric reflects the stability of test parameter adjustments and can be quantified using indicators such as standard deviation, coefficient of variation, or interquartile range. Larger dispersion indicates increased uncertainty in the adjustment behavior, potentially leading to fluctuations in test performance. The adjustment risk value is positively correlated with the dispersion metric, and the influence factor of the adjustment magnitude needs to be introduced. When the adjustment magnitude is close to the expected adjustment magnitude and the dispersion is small, the adjustment risk value remains within a reasonable range; when the adjustment magnitude deviates from the expected value or the dispersion increases significantly, the adjustment risk value increases accordingly. To comprehensively evaluate the value of state transitions, the expected return value and the adjustment risk value are combined to solve for the state transition reward value. Specifically, the expected return value is subtracted from the product of the adjustment risk value and the risk coefficient. The risk coefficient is used to control the sensitivity to risk and can be set according to the fault tolerance requirements of the actual test scenario.

[0109] Traverse all possible state region identifier pairs in the test state space and construct a state transition reward matrix. A state region identifier pair represents a transition relationship from one state region to another. For each state region identifier pair, extract the corresponding adjustment distribution characteristics and test performance evaluation value changes, and calculate the state transition reward value using the aforementioned method. Fill the calculated reward value into the corresponding position in the matrix, with the row index corresponding to the starting state region identifier and the column index corresponding to the target state region identifier. For state pairs without a direct transition relationship, set the reward value at the corresponding position to negative infinity or a minimum value, indicating that the transition path is unreachable. After completing the traversal, a complete state transition reward matrix is ​​obtained, which describes the immediate benefits of all transition behaviors in the state space.

[0110] The value propagation process calculates the state value distribution and initializes the value estimates for each state region identifier, typically by setting them as a zero vector or assigning initial values ​​based on prior knowledge. Value propagation employs an iterative update strategy, processing one or more state region identifiers in each iteration. For a selected target state region identifier, the corresponding row vector is extracted from the state transition reward matrix. This row vector contains the reward values ​​for transitioning from the target state to each subsequent state. Simultaneously, a set of subsequent state region identifiers for the target state is obtained, containing all states that can be directly transitioned from the target state.

[0111] For each state in the set of successor state region identifiers, extract its current value estimate. Multiply the successor state's value estimate by a discount factor, which ranges from zero to one and measures the importance of future returns. A discount factor closer to one indicates a greater emphasis on long-term returns; a discount factor closer to zero indicates a greater focus on immediate returns. Add the discounted value estimate to the corresponding state transition reward value to obtain the cumulative value of that transition path. Iterate through all successor states, calculate the cumulative value of each transition path, and select the maximum value as the basis for updating the target state's value. Update the target state region identifier's value estimate using the aforementioned maximum cumulative value. The update process can employ either a complete replacement or a weighted average strategy. When using a weighted average, a learning rate parameter needs to be set to control the update magnitude.

[0112] The value propagation and update process is repeated, sequentially processing the identifiers of each state region in the state space. After each iteration, the change in the estimated value of all states is calculated, typically using the maximum or average change as a convergence criterion. When the change in value is less than a preset convergence threshold, the value estimate is considered to have converged, and at this point, the estimated values ​​of each state constitute the state value distribution. The convergence threshold needs to be set in a balance between computational accuracy and efficiency; a threshold that is too small will lead to an increased number of iterations, while a threshold that is too large may result in inaccurate value estimation. In practical applications, a maximum number of iterations can be set as a termination condition to avoid wasting computational resources due to failure to converge in certain situations.

[0113] The state-action mapping table is constructed based on the correlation analysis between state value distribution and parameter adjustment actions. The estimated value corresponding to each state region identifier is extracted from the state value distribution; this value represents the expected long-term return when in that state. The preset set of parameter adjustment actions includes various executable parameter modification operations, such as increasing voltage amplitude, decreasing test frequency, and adjusting temperature range. For each state region identifier, historical state transition data is analyzed to statistically analyze the distribution of subsequent states reached after performing different parameter adjustment actions starting from that state.

[0114] The correlation between state region identifiers and preset parameter adjustment actions is calculated. This correlation calculation requires comprehensive consideration of the state transition probability after action execution and the estimated value of the target state. Specifically, for a given state and action, the frequency of transitions to each subsequent state after executing the action is statistically analyzed, and the transition probability distribution is calculated. The estimated value of each subsequent state is multiplied by its corresponding transition probability and summed to obtain the expected value of executing the action. This calculation is repeated for all preset parameter adjustment actions to obtain a sequence of expected values ​​for different actions. The expected value sequence is then normalized so that the sum of the correlations of all actions equals one; the normalized value is the mapping weight. The mapping weight reflects the priority of executing each action in a specific state; a larger weight indicates a higher long-term benefit from the action.

[0115] A state-action mapping table is constructed based on the calculated mapping weights. The table is stored in a key-value pair structure, where the key is the state region identifier and the value is an action weight vector. The dimension of the action weight vector equals the number of actions adjusted by preset parameters, and each element in the vector corresponds to the mapping weight of an action. To improve query efficiency, the mapping table can be optimized with an index, such as using a hash table structure or creating a fast retrieval index for state identifiers. The completed state-action mapping table can quickly respond to state query requests during real-time testing, returning recommended weights for actions adjusted based on the current state region identifier, providing a decision-making basis for subsequent action selection.

[0116] Figure 2This is a flowchart illustrating the test state analysis and parameter adjustment process of the integrated circuit test parameter adaptive optimization method based on reinforcement learning, as described in an embodiment of the present invention.

[0117] In one alternative implementation,

[0118] The current test status of the integrated circuit is obtained, and a set of candidate adjustment actions is obtained based on the state action mapping table. The expected test performance corresponding to each candidate adjustment action in the set of candidate adjustment actions is calculated, and the target adjustment action is selected, including:

[0119] The current test state of the integrated circuit is obtained and time-domain and frequency-domain features are extracted to construct the current state feature vector. The cosine similarity between the current state feature vector and the test feature vector corresponding to each state region identifier in the test state space is calculated. The state region identifier with the highest cosine similarity is selected as the matching state identifier. The mapping weight corresponding to each parameter adjustment action is queried in the state action mapping table. The parameter adjustment actions with the mapping weight greater than the preset weight threshold are extracted to construct a candidate adjustment action set.

[0120] The value estimate corresponding to the matching state identifier is extracted from the state value distribution as the benchmark value. The parameter adjustment amount corresponding to each candidate adjustment action in the candidate adjustment action set is determined and the simulated successor feature vector is calculated. The simulated successor feature vector is matched with the test feature vector corresponding to each state region identifier in the test state space to obtain the predicted successor state region identifier. The value estimate corresponding to each predicted successor state region identifier is extracted from the state value distribution and combined with the benchmark value to calculate the value increment. Based on the value increment and the state transition reward value corresponding to the predicted successor state region identifier, the expected test performance corresponding to each candidate adjustment action is solved, and the candidate adjustment action with the highest expected test performance is taken as the target adjustment action.

[0121] During integrated circuit testing, after the test equipment completes one or more test cycles, it needs to retrieve executable parameter adjustment schemes from the established state-action mapping table based on the current test status. Obtaining the current test status relies on the acquisition of real-time test data, specifically including the current values ​​of test parameters such as test frequency, voltage threshold, timing margin, and power consumption, as well as test result data such as yield data, test duration records, and coverage metrics from the most recent test batch.

[0122] When extracting features from the current test state, the test parameter sequence needs to be arranged in chronological order. The mean, variance, maximum, minimum, and trend coefficients of the parameter changes within the sliding time window are calculated; these statistics constitute the time-domain features. Simultaneously, a Fast Fourier Transform is performed on the test parameter sequence to extract the amplitude, phase information, and energy distribution characteristics of the main frequency components in the spectrum. The frequency-domain information reflects the periodic patterns and oscillatory characteristics of the parameter changes. The time-domain feature vector and the frequency-domain feature vector are then concatenated according to their dimensions to form a complete current-state feature vector, typically between 80 and 150 dimensions.

[0123] The state matching process uses cosine similarity to calculate the similarity between the current state feature vector and the representative test feature vector corresponding to each state region identifier in the test state space. The formula for calculating cosine similarity is as follows: ,in Represents the feature vector of the current state. Let represent the test feature vector corresponding to the state region identifier, · denotes the vector dot product operation, and ||·| denotes the Euclidean norm of the vector. Traverse all state region identifiers in the state space, calculate the similarity sequence, and select the state region identifier with the highest similarity as the matching state identifier. This identifier uniquely corresponds to the position of the current test state in the known state space.

[0124] The query operation is performed in the state-action mapping table based on the matching state identifier. The state-action mapping table records the association weight between each state region identifier and multiple parameter adjustment actions. The mapping weight value reflects the historical effect and priority of performing a parameter adjustment action in a specific state. The higher the weight, the more significant the test performance improvement produced by the action in historical data. A preset weight threshold is set, which is usually determined based on the statistical distribution of the mapping weights. It can be the mean of all mapping weights plus 0.5 times the standard deviation, or a fixed value between 0.6 and 0.8 based on practical engineering experience. All parameter adjustment actions with mapping weights greater than the preset weight threshold are filtered out. These actions constitute the candidate adjustment action set, which typically contains 5 to 15 different adjustment schemes.

[0125] To evaluate the actual effect of each action in the candidate adjustment action set, it is necessary to quantify the expected test performance. This involves extracting the value estimate corresponding to the matching state identifier from the state value distribution. The value estimate reflects the baseline level of test performance in the current state and is denoted as the baseline value. For any candidate adjustment action in the candidate adjustment action set, determine its specific parameter adjustment amount, including the increment or decrement value of each test parameter, such as adjusting the voltage parameter to +0.05V, adjusting the frequency parameter to +50MHz, etc.

[0126] The current state feature vector is corrected based on the parameter adjustment amount to generate a simulated successor feature vector. Specifically, the parameter adjustment amount is updated numerically according to the mapping relationship of the corresponding dimensions in the feature vector. For the mean dimension in the time domain features, the adjustment amount is directly added; for the variance dimension, it is scaled accordingly based on the absolute value of the adjustment amount; and for the frequency domain features, the amplitude of the frequency components is adjusted based on the impact of parameter changes on the spectral distribution. The generated simulated successor feature vector approximately represents the expected state features after performing this candidate adjustment action.

[0127] The simulated successor feature vector is matched with the test feature vector corresponding to each state region identifier in the test state space. Using the same cosine similarity calculation method as described above, the state region identifier with the highest similarity is selected as the predicted successor state region identifier. The predicted successor state region identifier represents the target state region that the test state is most likely to transition to after executing the current candidate adjustment action.

[0128] Extract the value estimate corresponding to the predicted successor state region identifier from the state value distribution, denoted as [value]. Calculate the value increment A positive value increment indicates that the expected performance improvement after implementing the adjustment action, while a negative value indicates a potential decrease in performance. Simultaneously, the state transition reward value corresponding to the region from the matched state identifier to the predicted subsequent state identifier is extracted from the state transition reward matrix and denoted as... The state transition reward value is calculated based on historical state transition data and reflects the immediate benefits of the state transition path.

[0129] The expected test performance of the candidate adjustment action is calculated using the combined value increment and the state transition reward value, as shown in the following formula: ,in This is a discount factor, ranging from 0.85 to 0.95, used to balance the weights of immediate rewards and long-term value. Iterate through all candidate adjustment actions in the candidate adjustment action set and calculate the expected test performance for each. The value is used to identify the candidate adjustment action with the highest expected test performance as the target adjustment action.

[0130] The selection of target adjustment actions depends not only on the magnitude of the value increment and state transition reward, but also implicitly on the stability of state transitions and historical transition success rates. When the expected test performance values ​​of multiple candidate adjustment actions are similar, actions with a higher frequency of occurrence in historical data and a higher state transition success rate are prioritized to reduce decision-making risk. The entire process of candidate action selection and target action determination is completed within a millisecond timescale, ensuring that the real-time requirements of test parameter adjustment are met, while quantitative calculations guarantee the scientific nature and traceability of the decision-making.

[0131] In one alternative implementation,

[0132] Constraint matching is performed based on the target adjustment action and the current resource occupancy status of the test environment, and the target adjustment action is corrected to obtain the optimal parameter adjustment command, which is then sent to the test equipment for execution, including:

[0133] Obtain the current resource usage status of the test environment and the resource load index of the test equipment, count the number of test tasks currently being executed and determine the concurrent task load, calculate the resource availability based on the resource load index and combine it with the concurrent task load to solve for the resource constraint strength;

[0134] Extract parameter adjustment amounts from the target adjustment action and calculate the resource consumption required to execute the target adjustment action. Calculate the resource gap and identify the resource bottleneck type based on the resource consumption and resource availability. Construct a constraint matching matrix based on the resource bottleneck type and the resource constraint strength and determine the constraint satisfaction degree corresponding to the target adjustment action.

[0135] Determine whether the constraint satisfaction is greater than a preset constraint threshold. If it is not greater, reduce the parameter adjustment amount in the target adjustment action based on the resource bottleneck type to obtain a corrected adjustment amount. Calculate the expected test performance corresponding to the corrected adjustment amount and balance it with the resource constraint strength to obtain a corrected target adjustment action. Iterate and verify the corrected target adjustment action and the constraint matching matrix until the constraint satisfaction is greater than the constraint threshold.

[0136] The verified target adjustment action is converted into a device-recognizable instruction format and encapsulated to obtain the optimal parameter adjustment instruction. The optimal parameter adjustment instruction is then sent to the test device for execution via a communication connection.

[0137] After obtaining the target adjustment action, it is necessary to consider the resource constraints in the actual test environment to ensure that the parameter adjustment operation does not exceed the carrying capacity of the test equipment. A real-time data link is established with the test equipment to collect resource load indicators, including CPU utilization, memory usage, channel load rate, and power consumption. These resource load indicators are reported by the equipment's internal monitoring module at a fixed sampling period. Simultaneously, a list of test tasks currently executing in the test environment is obtained, and the number of concurrently executing tasks is counted through the task management module. For the collected resource load indicators, the ratio of the current usage to the maximum capacity of each type of resource is calculated, and the resource availability is obtained based on a weighted average of these ratios. The calculation of resource availability needs to consider the weighting coefficients of different resource types, typically assigning higher weights to critical resources such as test channels. When determining the concurrent task load, not only the number of tasks is counted, but the resource occupancy of each task also needs to be evaluated. The overall concurrent load value is obtained by summing the products of task priority and resource occupancy. The resource availability and concurrent task load are normalized, and the resource constraint strength is calculated using a defined mapping function. This strength value reflects the degree of restriction on parameter adjustment operations in the current environment.

[0138] The specific adjustment amounts of each parameter are extracted from the target adjustment actions, including changes in test frequency, signal amplitude, and the increase or decrease in the number of sampling points. For each parameter adjustment, the amount of resources required to perform the adjustment is calculated based on a pre-established resource consumption model. The resource consumption model is obtained by fitting historical data and records the correlation between different parameter adjustment operations and resource consumption. The calculated resource consumption amounts are compared with the current available resources. When the demand for a certain type of resource exceeds the available resources, the difference between the two is calculated as a resource gap. By analyzing the distribution of resource gaps, the main resource types that restrict the execution of parameter adjustments are identified and marked as resource bottleneck types. Resource bottlenecks may include insufficient computing power, limited storage space, or saturated test channels. Based on the identified resource bottleneck types and the previously calculated resource constraint strength, a constraint matching matrix is ​​constructed. The rows of the constraint matching matrix represent different resource types, the columns represent the parameter adjustments involved in the target adjustment actions, and the matrix element values ​​represent the matching relationship between the resource consumption degree of the parameter adjustment and the current resource constraint strength. By comprehensively evaluating the constraint matching matrix, the overall constraint satisfaction of the target adjustment action is calculated using a weighted summation method. This value is between 0 and 1, with a larger value indicating a higher degree of constraint satisfaction.

[0139] The calculated constraint satisfaction is compared with a preset constraint threshold, which is determined based on the safety operation boundary and service quality requirements of the test equipment. When the constraint satisfaction is greater than the constraint threshold, it indicates that the target adjustment action can be directly executed under the current resource conditions. If the constraint satisfaction is not greater than the constraint threshold, the target adjustment action needs to be corrected. The correction process adopts targeted strategies based on different resource bottleneck types. When the bottleneck is computing resources, the adjustment magnitude of computing-intensive parameters is reduced first; when the bottleneck is storage resources, the adjustment amount of data cache-related parameters is reduced. Specifically, the adjustment amount of each parameter in the target adjustment action is multiplied by a reduction coefficient. The reduction coefficient is dynamically determined based on the size of the resource gap; the larger the resource gap, the smaller the reduction coefficient. The corrected parameter adjustment amount is obtained through reduction processing. These corrected adjustment amounts are then substituted back into the previously established resource consumption model to calculate the resource consumption corresponding to the corrected operation, and the expected test performance after correction is estimated based on the state value distribution.

[0140] After obtaining the expected test performance corresponding to the adjusted parameters, a trade-off needs to be struck between improving test performance and satisfying resource constraints. A trade-off objective function is constructed, with expected test performance as the benefit term and resource constraint strength as the penalty term. A weighted combination is used to find the adjustment scheme that maximizes the objective function. During the trade-off process, the weight coefficients of test performance and resource constraint strength are dynamically adjusted according to the urgency of the current test task. Urgent tasks tend to prioritize performance and moderately relax resource constraints, while regular tasks adhere more strictly to resource limits. The corrected target adjustment action is obtained through optimization calculation. This action can improve test performance to a certain extent without severely violating resource constraints. The corrected target adjustment action is then substituted back into the constraint matching matrix for verification, and the constraint satisfaction is recalculated. If the corrected constraint satisfaction is still not greater than the constraint threshold, an iterative correction process is initiated. In each iteration, the parameter adjustment magnitude is further reduced while monitoring the degree of decline in expected test performance and setting a tolerance lower limit for performance loss. The iterative process employs a binary search strategy, continuously adjusting the reduction coefficient to quickly converge to a parameter adjustment scheme that satisfies the constraints. When the constraint satisfaction rate first exceeds the constraint threshold, the iteration is terminated and the current correction target adjustment action is confirmed as the final executable solution.

[0141] After confirming the executable target adjustment action, it is converted into an instruction format that the test equipment can recognize and execute. According to the test equipment's communication protocol specifications, each adjustment value in the parameter adjustment action is mapped to the corresponding instruction code and parameter field. The instruction format includes an instruction header, parameter type identifier, parameter value, checksum, etc. Parameter values ​​need to be converted according to the numerical precision and units required by the equipment. For cases where multiple parameters need to be adjusted simultaneously, these parameters are packaged into a single instruction or arranged according to the instruction sequence supported by the equipment, ensuring that the execution order of each parameter adjustment conforms to the equipment's operating logic. During instruction encapsulation, a timestamp and sequence number are added for instruction tracking, and a priority identifier is added for scheduling processing on the equipment side. After instruction encapsulation is completed, format verification is performed to confirm that the instruction structure conforms to the protocol requirements and that the parameter values ​​are within the equipment's allowed range. Through the established communication connection with the test equipment, the optimal parameter adjustment instruction is sent to the test equipment using TCP or a dedicated industrial bus protocol. An acknowledgment mechanism is used when the instruction is sent, waiting for the equipment to return a confirmation signal. If no confirmation is received within the set timeout period, retransmission is performed. After receiving the instruction, the equipment's internal controller parses the instruction content, adjusts the test parameter configuration according to the instruction requirements, and initiates the test process under the new parameters. By monitoring the equipment's execution status, it is confirmed that the parameter adjustment has taken effect, providing a basis for subsequent test performance evaluation.

[0142] A second aspect of this invention provides an adaptive optimization system for integrated circuit test parameters based on reinforcement learning, comprising:

[0143] The test state modeling unit is used to collect the test parameter sequence and test result sequence of the integrated circuit test equipment, extract the time domain features and frequency domain features of the test parameter sequence and concatenate them to obtain the test feature vector, calculate the test pass rate and test coverage based on the test result sequence and solve the test performance evaluation value, and establish a test state space by associating and mapping the test feature vector and the test performance evaluation value.

[0144] The state value mapping unit is used to divide the test state space into states, construct a state transition sequence, calculate the test parameter difference between adjacent states, solve the parameter adjustment magnitude based on the test parameter difference, perform statistical analysis to obtain the adjustment distribution characteristics, construct a state transition reward matrix based on the adjustment distribution characteristics and the test performance evaluation value, perform value propagation calculation to obtain the state value distribution, and solve the mapping weight from different states to parameter adjustment actions in the test state space based on the state value distribution to obtain the state action mapping table.

[0145] The parameter optimization execution unit is used to obtain the current test status of the integrated circuit and query the candidate adjustment action set based on the status action mapping table, calculate the expected test performance corresponding to each candidate adjustment action in the candidate adjustment action set and filter to obtain the target adjustment action, perform constraint matching based on the target adjustment action and the resource occupancy status of the current test environment and correct the target adjustment action, obtain the optimal parameter adjustment instruction and send it to the test equipment for execution.

[0146] A third aspect of the present invention provides an electronic device, comprising:

[0147] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0148] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0149] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0150] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions 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 adaptive optimization method for integrated circuit test parameters based on reinforcement learning, characterized in that, include: The test parameter sequence and test result sequence of the integrated circuit test equipment are collected. The time domain features and frequency domain features of the test parameter sequence are extracted and concatenated to obtain the test feature vector. The test pass rate and test coverage are calculated based on the test result sequence, and the test performance evaluation value is solved. The test feature vector and the test performance evaluation value are correlated and mapped to establish the test state space. The test state space is divided into states to construct a state transition sequence and the test parameter difference between adjacent states is calculated. Based on the test parameter difference, the parameter adjustment magnitude is solved and statistical analysis is performed to obtain the adjustment distribution characteristics. Based on the adjustment distribution characteristics and the test performance evaluation value, a state transition reward matrix is ​​constructed and value propagation calculation is performed to obtain the state value distribution. Based on the state value distribution, the mapping weight from different states to parameter adjustment actions in the test state space is solved to obtain the state action mapping table. The current test status of the integrated circuit is obtained, and a set of candidate adjustment actions is obtained by querying the state action mapping table. The expected test performance corresponding to each candidate adjustment action in the set of candidate adjustment actions is calculated and the target adjustment action is selected. The target adjustment action is matched with the resource occupancy status of the current test environment and the target adjustment action is corrected to obtain the optimal parameter adjustment instruction and send it to the test equipment for execution.

2. The method according to claim 1, characterized in that, The process involves collecting test parameter sequences and test result sequences from integrated circuit testing equipment, extracting time-domain and frequency-domain features from the test parameter sequences and concatenating them to obtain a test feature vector, calculating the test pass rate and test coverage based on the test result sequences, and obtaining a test performance evaluation value. Finally, the test feature vector and the test performance evaluation value are correlated and mapped to establish a test state space, including: The test parameter sequence and test result sequence are collected from the integrated circuit test equipment for multiple test cycles. The test parameter sequence is segmented according to time order and the statistics in each segment are calculated to obtain time domain features. The test parameter sequence is subjected to spectrum transformation and the dominant frequency component and energy distribution are extracted to obtain frequency domain features. The time domain features and the frequency domain features are concatenated according to a preset dimension order to obtain a test feature vector. The test pass rate is obtained by statistically analyzing the number of samples that passed the test in the test result sequence and the total number of samples, and calculating the ratio. The test coverage rate is obtained by identifying the number of test item types in the test result sequence and combining it with the preset total number of test item types. The test pass rate and the test coverage rate are assigned weights and weighted summed to obtain a comprehensive evaluation value. A timeliness penalty is calculated based on the comprehensive evaluation value and preset timeliness constraints. The test performance evaluation value is obtained by solving based on the timeliness penalty and the comprehensive evaluation value. The test feature vectors are clustered to obtain multiple feature clusters and assigned state labels. The correlation between the test feature vectors in each feature cluster and the test performance evaluation value is calculated and a state feature matrix is ​​constructed. Based on the state feature matrix and the test performance evaluation value, state boundary conditions are determined. The state labels and the state boundary conditions are combined to obtain the test state space.

3. The method according to claim 2, characterized in that, Calculating the correlation between the test feature vectors within each feature cluster and the test performance evaluation value, and constructing a state feature matrix, then determining the state boundary conditions based on the state feature matrix and the test performance evaluation value includes: Extract test feature vector sets from different feature clusters and calculate the corresponding center vectors. Calculate the deviation product between the test feature vectors in each feature cluster and the test performance evaluation value and solve for the linear correlation degree. Sort the test feature vectors in each feature cluster and the test performance evaluation value respectively and calculate the cumulative value of the difference in sorting position to obtain the nonlinear correlation degree. Perform a weighted sum of the linear correlation degree and the nonlinear correlation degree to obtain the comprehensive correlation degree. The state identifier of each feature cluster and the center vector are arranged in order to form an initial state feature matrix. The comprehensive correlation degree is embedded as a weight parameter into the corresponding state identifier position in the initial state feature matrix, and the elements in the initial state feature matrix are scaled and adjusted to obtain the state feature matrix. Extract the test feature vector distribution corresponding to each state identifier from the state feature matrix and calculate the feature distribution range. Statistically analyze the numerical distribution of the test performance evaluation value under different state identifiers and calculate the value range of the test performance evaluation value. Perform cross-matching between the feature distribution range and the value range of the test performance evaluation value and identify overlapping and interval regions. Based on the overlapping and interval regions, determine the critical conditions for state switching to obtain the state boundary conditions.

4. The method according to claim 1, characterized in that, The test state space is divided into states to construct a state transition sequence, and the test parameter differences between adjacent states are calculated. Based on the test parameter differences, the parameter adjustment magnitude is solved, and statistical analysis is performed to obtain the adjustment distribution characteristics, including: Based on the pre-acquired state boundary conditions, the test feature vector in the test state space is divided into multiple state regions and assigned a state region identifier. The jump records between the state region identifiers in the pre-stored historical test data are extracted and arranged to obtain a state transition trajectory. The repeated state region identifier pairs in the state transition trajectory are counted and the transition frequency is calculated. Based on the transition frequency, a high-frequency transition path is determined and the state region identifiers are connected according to the transition direction to obtain a state transition sequence. The test feature vectors corresponding to adjacent state region identifiers are extracted from the state transition sequence, and the numerical differences of each dimension are calculated to obtain the test parameter differences. The test performance evaluation values ​​corresponding to adjacent state region identifiers are extracted, and the numerical differences are calculated to obtain the change in test performance evaluation values. The ratio between the difference of each dimension in the test parameter differences and the change in test performance evaluation values ​​is calculated, and the influence weight of each dimension is obtained by scaling. The absolute value of the difference of each dimension in the test parameter differences is calculated, and the influence weights are combined to perform a weighted summation to obtain the parameter adjustment range. The parameter adjustment ranges are summarized, sorted, and divided into intervals. The number of samples in each interval is counted to obtain the frequency distribution. The central tendency and dispersion measures of the frequency distribution are calculated, and the adjustment distribution characteristics are constructed.

5. The method according to claim 1, characterized in that, Based on the adjustment distribution characteristics and the test performance evaluation value, a state transition reward matrix is ​​constructed and value propagation calculation is performed to obtain the state value distribution. Based on the state value distribution, the mapping weights from different states to parameter adjustment actions in the test state space are solved to obtain a state action mapping table, including: The central tendency measure in the adjustment distribution characteristics is used as the expected adjustment range. The expected return value is calculated based on the expected adjustment range and the test performance evaluation value. The adjustment risk value is calculated based on the pre-acquired dispersion measure and combined with the expected return value to obtain the state transition reward value. All state region identifier pairs in the test state space are traversed and the corresponding state transition reward values ​​are filled to construct the state transition reward matrix. Initialize the value estimate of each state region identifier, extract the state transition reward value corresponding to the target state region identifier from the state transition reward matrix and combine it with the set of subsequent state region identifiers, calculate the value estimate of different subsequent state region identifiers in the set of subsequent state region identifiers and solve for the cumulative value by combining the corresponding state transition reward value, update the value estimate of the target state region identifier based on the cumulative value, and repeat the update until the value estimate converges to obtain the state value distribution. The estimated value of each state region identifier is extracted from the state value distribution, and the correlation between the value and the preset parameter adjustment action is calculated to obtain the mapping weight. A state action mapping table is constructed based on the mapping weight.

6. The method according to claim 1, characterized in that, The current test status of the integrated circuit is obtained, and a set of candidate adjustment actions is obtained based on the state action mapping table. The expected test performance corresponding to each candidate adjustment action in the set of candidate adjustment actions is calculated, and the target adjustment action is selected, including: The current test state of the integrated circuit is obtained and time-domain and frequency-domain features are extracted to construct the current state feature vector. The cosine similarity between the current state feature vector and the test feature vector corresponding to each state region identifier in the test state space is calculated. The state region identifier with the highest cosine similarity is selected as the matching state identifier. The mapping weight corresponding to each parameter adjustment action is queried in the state action mapping table. The parameter adjustment actions with the mapping weight greater than the preset weight threshold are extracted to construct a candidate adjustment action set. The value estimate corresponding to the matching state identifier is extracted from the state value distribution as the benchmark value. The parameter adjustment amount corresponding to each candidate adjustment action in the candidate adjustment action set is determined and the simulated successor feature vector is calculated. The simulated successor feature vector is matched with the test feature vector corresponding to each state region identifier in the test state space to obtain the predicted successor state region identifier. The value estimate corresponding to each predicted successor state region identifier is extracted from the state value distribution and combined with the benchmark value to calculate the value increment. Based on the value increment and the state transition reward value corresponding to the predicted successor state region identifier, the expected test performance corresponding to each candidate adjustment action is solved, and the candidate adjustment action with the highest expected test performance is taken as the target adjustment action.

7. The method according to claim 1, characterized in that, Constraint matching is performed based on the target adjustment action and the current resource occupancy status of the test environment, and the target adjustment action is corrected to obtain the optimal parameter adjustment command, which is then sent to the test equipment for execution, including: Obtain the current resource usage status of the test environment and the resource load index of the test equipment, count the number of test tasks currently being executed and determine the concurrent task load, calculate the resource availability based on the resource load index and combine it with the concurrent task load to solve for the resource constraint strength; Extract parameter adjustment amounts from the target adjustment action and calculate the resource consumption required to execute the target adjustment action. Calculate the resource gap and identify the resource bottleneck type based on the resource consumption and resource availability. Construct a constraint matching matrix based on the resource bottleneck type and the resource constraint strength and determine the constraint satisfaction degree corresponding to the target adjustment action. Determine whether the constraint satisfaction is greater than a preset constraint threshold. If it is not greater, reduce the parameter adjustment amount in the target adjustment action based on the resource bottleneck type to obtain a corrected adjustment amount. Calculate the expected test performance corresponding to the corrected adjustment amount and balance it with the resource constraint strength to obtain a corrected target adjustment action. Iterate and verify the corrected target adjustment action and the constraint matching matrix until the constraint satisfaction is greater than the constraint threshold. The verified target adjustment action is converted into a device-recognizable instruction format and encapsulated to obtain the optimal parameter adjustment instruction. The optimal parameter adjustment instruction is then sent to the test device for execution via a communication connection.

8. An adaptive optimization system for integrated circuit test parameters based on reinforcement learning, used to implement the method of any one of claims 1-7, characterized in that, include: The test state modeling unit is used to collect the test parameter sequence and test result sequence of the integrated circuit test equipment, extract the time domain features and frequency domain features of the test parameter sequence and concatenate them to obtain the test feature vector, calculate the test pass rate and test coverage based on the test result sequence and solve the test performance evaluation value, and establish a test state space by associating and mapping the test feature vector and the test performance evaluation value. The state value mapping unit is used to divide the test state space into states, construct a state transition sequence, calculate the test parameter difference between adjacent states, solve the parameter adjustment magnitude based on the test parameter difference, perform statistical analysis to obtain the adjustment distribution characteristics, construct a state transition reward matrix based on the adjustment distribution characteristics and the test performance evaluation value, perform value propagation calculation to obtain the state value distribution, and solve the mapping weight from different states to parameter adjustment actions in the test state space based on the state value distribution to obtain the state action mapping table. The parameter optimization execution unit is used to obtain the current test status of the integrated circuit and query the candidate adjustment action set based on the status action mapping table, calculate the expected test performance corresponding to each candidate adjustment action in the candidate adjustment action set and filter to obtain the target adjustment action, perform constraint matching based on the target adjustment action and the resource occupancy status of the current test environment and correct the target adjustment action, obtain the optimal parameter adjustment instruction and send it to the test equipment for execution.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to 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 the processor, they implement the method described in any one of claims 1 to 7.

Images