Intelligent test case screening method and system, medium and equipment

Abstract

The invention provides an intelligent screening method and system for test cases, a medium and equipment, and the method comprises the steps: building a dual-channel attention mechanism through an LSTM-Transform hybrid architecture, and extracting a transverse coupling feature and a cross-platform longitudinal migration feature of hardware-software-service parameters; establishing a three-level attenuation network, respectively quantizing exponential attenuation of hardware iteration, nonlinear attenuation of a software version and step attenuation of API change, and fusing attenuation factors and spatio-temporal features to construct an association degree prediction model; calculating a dynamic confidence threshold based on the wavelet feature energy entropy and the platform migration compensation factor; according to a correlation degree result and a threshold value, screening necessary and rejectable use cases, and generating a visual map; and through incremental learning of a closed loop, data updating model parameters are executed after screening, and a continuously optimized test decision cycle is formed. According to the method, multi-dimensional features can be accurately captured, platform differences are dynamically adapted, the rate of missing report and false report is remarkably reduced, and the test efficiency and the resource utilization rate are improved.

Description

Technical Field

[0001] This application relates to the field of computer software testing technology, and in particular to a method, system, medium, and device for intelligent screening of test cases. Background Technology

[0002] In the iterative process of computer hardware and software technology, servers, as the core computing power carriers, experience frequent updates to their hardware architecture, supporting software, and business parameters. Effective test case selection is crucial for ensuring the stability and compatibility of servers after iteration. Currently, the mainstream test case selection techniques in the industry fall into two main categories: one is a static rule-based selection method, which filters test cases by pre-setting matching rules for hardware configuration and software versions, and judges whether test cases meet testing requirements based on fixed parameter thresholds; the other is a single-dimensional machine learning-based selection method, which learns test case characteristics based on historical test data, and achieves selection by predicting test case priorities. In some scenarios, it also relies on human experience to define selection criteria, using testers' subjective judgment to determine reusable or discardable test cases.

[0003] However, these existing technologies have significant shortcomings at the principle level, making them difficult to adapt to the complex needs of server testing: First, existing technologies generally treat hardware configuration, software version, and business parameters as independent variables, ignoring the implicit coupling relationship between them. They fail to capture the impact of multi-dimensional parameter interactions on test case relevance, especially in cross-platform testing scenarios. The failure to extract migration features such as platform architecture and instruction set differences leads to prominent issues of test case failure and false negatives. Second, existing technologies lack the ability to model the decay of test case relevance caused by technology iteration. This applies to everything from the precipitous drop in relevance caused by major API changes, the gradual change in failure probability caused by software version iterations, to the gradual weakening of relevance due to hardware updates. First, none of the above were included in the quantitative analysis, resulting in the retention of a large number of invalid test cases that were out of touch with the new platform, leading to a waste of testing resources. Second, the existing technology uses fixed thresholds for screening decisions, without considering the differences in characteristics between different platforms. It cannot dynamically adjust the thresholds according to the different sensitivity and adaptation requirements of the platform to parameters, and it also does not combine changes in feature energy to optimize the judgment criteria, resulting in the screening accuracy being greatly affected by platform characteristics. Third, the existing technology lacks incremental learning mechanisms and visual decision support. After the model parameters are trained, it is difficult to update and optimize them with new test data. As technology iterates, the model accuracy gradually decreases. Moreover, it only outputs the screening results without providing an intuitive display of parameter correlation and platform migration path, resulting in low efficiency of operation and maintenance decisions.

[0004] Furthermore, the server testing industry is a niche field with insufficient industry collaboration. Due to intellectual property barriers, it is difficult for manufacturers to share core data on technology iteration and test case failures, which further limits the large-scale optimization of existing technologies and exacerbates the impact of the aforementioned technological deficiencies. Summary of the Invention

[0005] In view of this, this application provides a method, system, medium and device for intelligent screening of test cases, with the main purpose of improving testing efficiency and resource utilization.

[0006] According to one aspect of this application, a test case intelligent screening method is provided, comprising: a spatiotemporal feature extraction step: acquiring historical test case data and technology update data, wherein the historical test case data includes hardware configuration, software version, and business parameter information; constructing a dual-channel attention mechanism using an LSTM-Transformer hybrid architecture, quantifying the coupling effect of hardware-software-business parameters through a horizontal attention channel to generate a parameter sensitivity matrix, and extracting cross-platform migration features through a vertical convolution channel to construct a platform migration feature map, thereby achieving cross-spatiotemporal feature fusion and feature differentiation; and a correlation decay modeling and fusion step: establishing a three-layer decay network, respectively quantifying the correlation decay caused by hardware iteration through an exponential decay model, and quantifying the correlation decay caused by software iteration through a nonlinear decay model. The attenuation caused by version iteration and the attenuation driven by API changes are quantified through a step attenuation model, and attenuation factors at each level are output. The parameter sensitivity matrix, platform migration feature map, and attenuation factors at each level are fused to construct a test case correlation prediction model. Dynamic threshold calculation step: feature energy is extracted through wavelet transform, combined with platform migration compensation factor, and a dynamic confidence threshold is calculated through an adaptive algorithm. Test case screening step: necessary test cases and removable test cases are screened based on the comparison between the correlation results output by the test case correlation prediction model and the dynamic confidence threshold, combined with preset rejection conditions. Visualization step: parameter sensitivity heatmap and platform migration path map are generated based on cross-platform migration features, and the screening process and results are visualized.

[0007] In one implementation, in the spatiotemporal feature extraction step, the lateral attention channel maps hardware-software-business parameters into high-dimensional feature vectors through tensor decomposition, and then quantifies the parameter coupling effect through matrix projection and softmax activation function; the longitudinal convolution channel encodes the differences in hardware architecture and instruction set of different platforms through a convolutional spatiotemporal encoder, and extracts the failure mode features of cross-generational platforms to construct a platform migration feature map.

[0008] In one implementation, the correlation decay modeling and fusion step involves: the exponential decay model expression for the hardware-level decay factor is γ_hw = e^(-λ·Δt), where Δt is the hardware iteration cycle and λ is the hardware decay coefficient; the nonlinear decay model expression for the software-level decay factor is γ_sw = 1 / (1+β·ΔVer), where ΔVer is the software version difference and β is the software decay coefficient; when the API change rate exceeds a preset threshold, the business-level decay factor experiences a step decrease; the fusion process maps the parameter sensitivity matrix and each level decay factor to the same feature space through matrix multiplication, and combines the weight allocation of the platform migration feature map to output the test case correlation prediction result.

[0009] In one implementation, the process of extracting feature energy through wavelet transform in the dynamic threshold calculation step includes: performing wavelet decomposition on the fused features obtained in the spatiotemporal feature extraction step to obtain multi-scale wavelet coefficients; calculating the total feature energy of the wavelet coefficients at each scale, selecting wavelet coefficients that meet the preset validity conditions and calculating their effective feature energy; calculating the ratio of effective feature energy to total feature energy as wavelet feature energy entropy, and the adaptive algorithm uses the wavelet feature energy entropy as the core input and combines it with the platform migration compensation factor to calculate the dynamic confidence threshold.

[0010] In one implementation, the parameter sensitivity matrix contains quantified failure probability values ​​of various business parameters under different hardware configurations and software versions; the platform migration feature map contains the mapping relationship between cross-platform architecture differences, instruction set differences and test case failure modes.

[0011] In one implementation, during the visualization step, cross-platform instruction set difference annotations are overlaid on the generated parameter sensitivity heatmap. The annotations include the influence coefficients of different platform instruction set types and instruction set compatibility on the relevance of test cases.

[0012] In one implementation, after the visualization step, the method further includes: an incremental learning closed-loop step: collecting the filtered test case execution data as incremental data, updating the model parameters based on the incremental data, completing the model update, and returning to the spatiotemporal feature extraction step to form a test decision closed loop; wherein, the model parameter update includes: calculating the error loss of the parameter sensitivity matrix based on the incremental data, optimizing the matrix projection parameters of the lateral attention channel through backpropagation; adjusting the feature encoding weights of the vertical convolution channel according to the cross-platform test results; and updating the decay coefficients (λ, β) of the three-layer decay network based on the correspondence between API changes and test case failures.

[0013] According to one aspect of this application, a test case intelligent screening system is provided, comprising: a data acquisition module for collecting historical test case data and technology update data, wherein the historical test case data includes hardware configuration, software version, and business parameter information, and the technology update data includes hardware iteration, software version update, and API change information; a spatiotemporal feature fusion module for constructing a dual-channel attention unit using an LSTM-Transformer hybrid architecture, including a horizontal attention unit and a vertical convolutional unit; wherein the horizontal attention unit is used to quantify the coupling effect of hardware-software-business parameters to generate a parameter sensitivity matrix, and the vertical convolutional unit is used to extract cross-platform migration features to construct a platform migration feature map; and a correlation decay fusion module. The system comprises a three-level attenuation unit and a feature fusion unit. The three-level attenuation unit quantifies the attenuation factors corresponding to hardware, software, and API changes through exponential attenuation, nonlinear attenuation, and step attenuation models, respectively. The feature fusion unit is used to fuse the parameter sensitivity matrix, platform migration feature map, and attenuation factors to construct a test case correlation prediction model. The dynamic threshold decision module is used to extract feature energy through wavelet transform and calculate the dynamic confidence threshold by combining it with the platform migration compensation factor. The test case screening module is used to output necessary test cases and removable test cases based on the comparison between the test case correlation prediction results and the dynamic confidence threshold. The visualization module is used to generate parameter sensitivity heatmaps and platform migration path maps based on cross-platform migration features.

[0014] According to one aspect of this application, a storage medium is provided that stores a computer program, wherein the computer program is configured to execute the above-described method at runtime.

[0015] According to one aspect of this application, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the methods described above.

[0016] By employing the above technical solutions, this application provides a method, system, medium, and device for intelligent test case screening. It utilizes a spatiotemporal feature extraction method based on an LSTM-Transformer hybrid architecture, which can both break the coupling effect of hardware-software-business parameters to accurately extract information features and capture cross-platform migration features, significantly reducing the platform adaptation false negative rate. Dynamic decay modeling quantifies the step decay caused by API changes, and combined with version iteration descriptions, it analyzes key change points and performs reasonable threshold screening, effectively avoiding the drawback of undetected historical test case failures, significantly enhancing prediction accuracy and significantly reducing the false positive rate of test case failures. Adaptive confidence thresholding dynamically adjusts screening criteria based on the characteristics of different platforms such as AMD, ARM, and X86, coupling platform features and integrating their requirements for the evaluation model, improving the credibility of confidence intervals and significantly increasing the accuracy of necessary test case screening. Simultaneously, it can generate visualization support such as parameter sensitivity heatmaps and platform migration path maps, doubling the efficiency of operation and maintenance decisions.

[0017] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This paper illustrates a flowchart of an intelligent test case selection method provided in an embodiment of this application. Figure 2 This illustration shows a logical diagram of a test case intelligent screening method provided in an embodiment of this application; Figure 3 This paper illustrates a schematic diagram of the structure of an intelligent test case screening system provided in an embodiment of this application. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, and not all of them. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present application. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present application can be combined with each other.

[0020] This application addresses the shortcomings of existing test case screening methods. It tackles the issues of neglecting hardware-software-business parameter coupling and missing cross-platform features in current technologies by proposing a dual-channel attention mechanism with an LSTM-Transformer hybrid architecture. This mechanism quantifies multi-parameter coupling relationships through horizontal attention and extracts cross-platform migration features through vertical convolution. To address the lack of modeling for technology iteration attenuation, a three-layer attenuation network (hardware exponential attenuation, software nonlinear attenuation, and API step attenuation) is designed, and attenuation factors are fused with spatiotemporal features to construct a correlation prediction model. For the poor adaptability of fixed thresholds, a dynamic confidence threshold algorithm based on wavelet feature energy entropy is proposed, combined with a platform migration compensation factor to achieve adaptive threshold adjustment. Finally, to address the lack of incremental learning and visualization support, an incremental learning closed loop (updating model parameters with filtered data) and a cross-platform migration feature visualization system are constructed (generating heatmaps and path maps and annotating instruction set differences). Ultimately, this forms a complete technical solution covering feature extraction, attenuation modeling, threshold calculation, and iterative optimization, adapting to the needs of server testing scenarios.

[0021] The terminology is explained below.

[0022] Step decay: A quantitative model in which the relevance of test cases drops sharply when the API undergoes a major change (e.g., a sudden change from γ_biz=0 to 1).

[0023] Multidimensional correlation analysis: Analyze the implicit coupling relationships between hardware configuration, software version, and business parameters (such as the correlation between Thread_per_block and GPU architecture).

[0024] Dynamic confidence threshold: The test case selection threshold is adjusted in real time based on the historical effective feature energy to adapt to platform differences.

[0025] Nonlinear decay: Software version iteration causes the probability of use case failure to decrease in an S-shaped curve (e.g., γ_sw=1 / (1+β·ΔVer)).

[0026] LSTM-Transformer hybrid architecture: lateral attention captures parameter sensitivity, vertical convolution extracts platform transfer features, and achieves cross-temporal and spatial dimension feature fusion.

[0027] See Figure 1 The flowchart of a test case intelligent screening method provided in this application embodiment includes the following steps S101-S106.

[0028] S101, Spatiotemporal Feature Extraction: Acquire historical test case data and technology update data. The historical test case data includes hardware configuration, software version, and business parameter information. A dual-channel attention mechanism is built using an LSTM-Transformer hybrid architecture. The coupling effect of hardware-software-business parameters is quantified through the horizontal attention channel to generate a parameter sensitivity matrix. Cross-platform migration features are extracted through the vertical convolution channel to construct a platform migration feature map, thereby achieving cross-spatiotemporal feature fusion and feature differentiation.

[0029] In this embodiment, to overcome the limitations of traditional technologies that treat hardware configuration, software version, and business parameters as independent variables, a dual-channel attention mechanism built with an LSTM-Transformer hybrid architecture is used to achieve collaborative capture of multi-parameter coupled features and cross-platform transfer features across spatiotemporal dimensions. Specifically, the lateral attention channel relies on the Transformer's self-attention mechanism to first transform hardware configuration, software version, and business parameters into high-dimensional feature tensors. Then, through tensor decomposition operations and the softmax activation function, the implicit coupling effects among the three are quantified. For example, the limiting relationship between different GPU architectures and the range of business parameter values ​​under a specific software version is shown, ultimately forming a parameter sensitivity matrix. Each element in the matrix corresponds to the correlation strength of a set of parameter combinations, intuitively reflecting the potential effectiveness of test cases under that combination. The vertical convolutional channel combines the learning ability of LSTM to learn temporal patterns with the ability of convolutional neural networks to capture spatial features. First, the LSTM layer sorts out the temporal context of hardware or software iteration, such as the update sequence of different generations of GPU architecture. Then, the convolutional layer extracts the architectural difference features between platforms, such as the expansion or adjustment of instruction set functions of different platforms. In this way, a platform migration feature map is constructed. The map clearly presents the failure modes that test cases may face when migrating between different platforms, thereby achieving accurate differentiation and extraction of multi-dimensional features and laying the foundation for subsequent correlation analysis.

[0030] In one implementation, spatiotemporal feature extraction may specifically include: the horizontal attention channel maps hardware-software-business parameters into high-dimensional feature vectors through tensor decomposition, and then quantifies the parameter coupling effect through matrix projection and softmax activation function; the vertical convolution channel encodes the differences in hardware architecture and instruction set of different platforms through convolutional spatiotemporal encoder, and extracts the failure mode features of cross-generation platforms to construct a platform migration feature map.

[0031] In this embodiment, the parameter sensitivity matrix may contain the quantified failure probability values ​​of each business parameter under different hardware configurations and software versions; the platform migration feature map contains the mapping relationship between cross-platform architecture differences, instruction set differences and test case failure modes.

[0032] S102, Correlation Attenuation Modeling and Fusion: A three-level attenuation network is established, which quantifies the correlation attenuation caused by hardware iteration through an exponential attenuation model, the attenuation caused by software version iteration through a nonlinear attenuation model, and the attenuation driven by API changes through a step attenuation model, and outputs the attenuation factor of each level; the parameter sensitivity matrix, platform migration feature map and the attenuation factor of each level are fused to construct a test case correlation prediction model.

[0033] The correlation decay modeling and fusion aims to transform the declining trend of existing test case correlation into a computable mathematical model and deeply integrate it with previously extracted spatiotemporal features to construct a test case correlation prediction model. First, a three-level decay network is designed for different types of technology iterations. Correlation decay caused by hardware iteration is calculated using an exponential decay model, as hardware iterations typically have a fixed cycle, and correlation decreases slowly exponentially with the extension of the iteration cycle. The decay rate is determined by the characteristics of the hardware itself, and the decay coefficient varies for different types of hardware. Decrease caused by software version iteration is quantified using a nonlinear decay model. The greater the software version difference, the more the increase in test case failure probability conforms to an S-shaped curve. For example, the failure probability changes slowly with small version updates, but rises rapidly with multiple version updates. Decrease driven by API changes is described using a step decay model. When the API change rate exceeds a certain threshold, the correlation between the test case and the new environment drops sharply; below the threshold, the correlation remains relatively stable. After calculating the attenuation factors at each level, the horizontal coupling characteristics reflected by the parameter sensitivity matrix and the vertical migration characteristics reflected by the platform migration feature map are mapped to the same feature space with the three-level attenuation factors through matrix multiplication. After processing by the fully connected layer, the correlation prediction result of the test cases is finally output. This result can comprehensively reflect the effectiveness of the test cases in the current technical environment.

[0034] In one implementation, the specific method for modeling and fusing correlation decay is as follows: the exponential decay model expression for the hardware-level decay factor is γ_hw = e^(-λ·Δt), where Δt is the hardware iteration period and λ is the hardware decay coefficient; the nonlinear decay model expression for the software-level decay factor is γ_sw = 1 / (1+β·ΔVer), where ΔVer is the software version difference and β is the software decay coefficient; when the API change rate exceeds a preset threshold, the business-level decay factor undergoes a step decrease; the fusing process maps the parameter sensitivity matrix and the decay factors of each level to the same feature space through matrix multiplication, and combines the weight allocation of the platform migration feature map to output the correlation prediction results of test cases.

[0035] S103, Dynamic Threshold Calculation: Feature energy is extracted through wavelet transform, and combined with the platform migration compensation factor, the dynamic confidence threshold is calculated through an adaptive algorithm.

[0036] In this embodiment, to address the issue that traditional fixed thresholds cannot adapt to different platform characteristics, an adaptive confidence threshold is generated by combining feature energy analysis and platform characteristic compensation. First, wavelet transform is used to process the previously extracted parameter sensitivity matrix and platform migration feature map, decomposing the feature signal into high-frequency and low-frequency components. The high-frequency components are mostly irrelevant noise, while the low-frequency components contain key information strongly correlated with the validity of the test cases. The ratio of the total feature energy of all wavelet components to the effective feature energy corresponding to the low-frequency components is calculated; this ratio is the wavelet feature energy entropy. A higher entropy value indicates a higher proportion of effective information in the features, and a stronger reliability in judging the correlation of test cases. Subsequently, a platform migration compensation factor is introduced. This factor is determined based on the characteristics of different platforms. For example, some platforms are more sensitive to instruction set compatibility, and the compensation factor value will be adjusted accordingly. Combined with the basic threshold and adjustment coefficient, the threshold is smoothly and dynamically adjusted through the Sigmoid function. That is, when the wavelet feature energy entropy is high, the threshold will be appropriately increased to ensure that the selected test cases have higher effectiveness. When the platform characteristics have high requirements for test sensitivity, the compensation factor will make the threshold adapt to the platform requirements, thereby avoiding the misjudgment problem caused by a fixed threshold and ensuring that the threshold can accurately match the current test scenario.

[0037] In one implementation, the specific process of extracting feature energy through wavelet transform may include: performing wavelet decomposition on the fused features obtained from the spatiotemporal feature extraction step to obtain multi-scale wavelet coefficients; calculating the total feature energy of the wavelet coefficients at each scale, selecting wavelet coefficients that meet the preset validity conditions and calculating their effective feature energy; calculating the ratio of effective feature energy to total feature energy as wavelet feature energy entropy, and using the wavelet feature energy entropy as the core input of the adaptive algorithm, combined with the platform migration compensation factor to calculate the dynamic confidence threshold.

[0038] S104, Test case screening: Based on the comparison between the correlation result output by the test case correlation prediction model and the dynamic confidence threshold, and combined with preset elimination conditions, necessary test cases and test cases that can be eliminated are screened.

[0039] In this embodiment, test cases are accurately classified based on the correlation prediction results and dynamic confidence thresholds, combined with preset rejection conditions. Specifically, the correlation prediction results of test cases are first compared with the dynamic confidence threshold. If the correlation result is higher than the dynamic threshold, it indicates that the test case has a strong correlation with the current hardware, software, and business parameters, and has been verified by both attenuation effect and feature energy, thus contributing significantly to defect detection. Therefore, it is determined to be a necessary test case, which should be prioritized for execution in subsequent tests. If the correlation result is lower than the preset rejection threshold, it indicates that the test case has become severely out of touch with the current testing environment due to hardware iteration, software updates, or API changes. Even if executed, it is difficult to detect defects, and it can be directly determined as a rejectable test case, thereby reducing the consumption of ineffective testing resources. For test cases with correlation results between the dynamic threshold and the preset rejection threshold, they are marked as test cases to be verified. Subsequent manual sampling verification can be used to further determine whether they need to be executed, ultimately forming a clear test case classification result. This ensures the detection rate of critical defects, significantly improves testing efficiency, and optimizes the allocation of testing resources.

[0040] S105, Visualization: Based on cross-platform migration characteristics, a parameter sensitivity heatmap and a platform migration path map are generated to visualize the screening process and results.

[0041] In this embodiment, abstract feature data, attenuation model calculation processes, and screening results are transformed into an intuitive graphical interface through visualization, helping testers quickly understand the screening logic and locate the root cause of problems. The parameter sensitivity heatmap uses hardware configuration, business parameters, and software version as different dimensions, and visually presents the correlation strength in the parameter sensitivity matrix through color depth differences. Darker colors represent stronger correlations between parameter combinations and higher test case effectiveness, while darker colors indicate weaker correlations and lower effectiveness. Testers can quickly identify high-risk or high-effectiveness parameter combinations through the heatmap. The platform migration path map adopts a directed graph structure, with nodes representing different platform types and edges representing migration paths between platforms. The edge colors and labels correspond to the test case failure rate and key failure reasons during migration, such as failures due to instruction set incompatibility or API changes, allowing testers to clearly trace the core issues causing cross-platform test case failures. At the same time, it will generate an interactive filtering result report containing test case ID, correlation value, classification result, decay factor details and threshold data, and provide comparative information on the number of test cases, time consumption and defect detection rate before and after filtering. This allows testers to quickly grasp the filtering effect and key information without analyzing complex raw data, improve the efficiency of operation and maintenance decision-making, and provide clear data support for cross-team collaboration and knowledge transfer.

[0042] In one implementation, during the visualization step, cross-platform instruction set difference annotations are overlaid on the generated parameter sensitivity heatmap. The annotations include the influence coefficients of different platform instruction set types and instruction set compatibility on the relevance of test cases.

[0043] Following step S105, this embodiment can further perform an incremental learning closed-loop step: collecting the filtered test case execution data as incremental data, updating the model parameters based on the incremental data, completing the model update, and returning to the spatiotemporal feature extraction step to form a test decision closed loop; wherein, the model parameter update includes: calculating the error loss of the parameter sensitivity matrix based on the incremental data, optimizing the matrix projection parameters of the lateral attention channel through backpropagation; adjusting the feature encoding weights of the vertical convolution channel according to the cross-platform test results; and updating the decay coefficients (λ, β) of the three-layer decay network based on the correspondence between API changes and test case failures.

[0044] The incremental learning loop aims to break through the limitations of traditional test case selection models where parameters are fixed after training. By continuously collecting actual execution data of selected test cases, it transforms this data into incremental data to feed back into model optimization, ultimately forming a test decision-making loop of "data collection - model update - application selection - data feedback," ensuring the model can continuously adapt to technological iterations and changes in test scenarios. First, after test case selection and execution, the system collects various execution data. This data includes not only the results of whether the test cases pass or fail, but also the actual operating status of hardware configurations, software version compatibility, actual and expected deviations of business parameters, and abnormal feedback during API calls. This multi-dimensional actual execution data constitutes incremental data for model optimization, providing a realistic and current test scenario-appropriate basis for subsequent parameter updates. In the model parameter update phase, optimizations and adjustments are made to the core modules built in the early stages. For the matrix projection parameters of the lateral attention channel, the error loss between the current parameter sensitivity matrix and the actual execution results is first calculated based on incremental data. For example, by comparing the correlation strength predicted by the parameter sensitivity matrix with the failure of test cases under the same parameter combination in actual testing, the degree of deviation in matrix prediction is quantified. Then, the matrix projection parameters are adjusted along the error gradient direction using the backpropagation algorithm, so that the updated lateral attention channel can more accurately quantify the coupling effect of hardware-software-business parameters and reduce test case misjudgments caused by parameter prediction deviations. For the feature encoding weights of the vertical convolution channel, adjustments are made based on the cross-platform test results in the incremental data. For example, if it is found in actual testing that instruction set details not fully considered in the original feature encoding cause test case failures during a certain platform migration, the proportion of the instruction set features in the encoding weights will be increased accordingly, making the cross-platform migration features extracted by the vertical convolution channel more consistent with the actual migration scenario and improving the accuracy of the platform migration feature map. The decay coefficients of the three-level decay network are updated based on the correspondence between API changes and test case failures in the incremental data. For example, if it is found that the actual test case correlation decay rate is faster after a certain type of hardware iteration than the original λ coefficient is calculated, the λ value will be appropriately increased. If the change in test case failure probability does not fit well with the original β coefficient when a certain type of software version is updated, the β coefficient will be adjusted so that the exponential decay model and the nonlinear decay model can more accurately describe the actual decay law. After all model parameters are updated, the updated model is reapplied to the spatiotemporal feature extraction step, so that subsequent spatiotemporal feature extraction, correlation decay modeling, dynamic threshold calculation, and other steps are all based on the optimized model, ensuring that each screening can use the latest test data to improve accuracy.This continuous iterative closed-loop mechanism effectively prevents the model from gradually becoming ineffective due to technological iterations (such as the emergence of new hardware architectures and frequent software version updates) or changes in scenarios (such as the shift of testing requirements from general-purpose servers to dedicated computing servers). It maintains the model's adaptability to the testing scenario. At the same time, as incremental data accumulates, the model's characterization of parameter coupling effects and decay patterns becomes more accurate, thereby continuously improving the efficiency and quality of test case selection and forming a virtuous cycle.

[0045] In summary, the intelligent test case selection method provided in this application addresses the problems of existing technologies by providing a targeted solution through a layered design. First, existing technologies treat hardware configuration, software version, and business parameters as independent variables, failing to capture multi-dimensional coupling effects and neglecting to extract cross-platform migration features, leading to cross-platform test failures and missed detections. This application solves this problem through an "LSTM-Transformer hybrid architecture dual-channel attention mechanism": the horizontal attention channel quantifies the coupling relationship between hardware, software, and business parameters, generating a parameter sensitivity matrix to clarify the impact of multi-dimensional parameter interactions on test case relevance; the vertical convolutional channel extracts cross-platform migration features (such as architectural differences and instruction set changes), constructing a platform migration feature map, achieving accurate capture of test case relevance features in cross-platform scenarios, and fundamentally eliminating the failure and missed detection risks caused by the "independent variable assumption." Second, existing technologies do not quantify the attenuation of test case relevance caused by API changes, software iterations, and hardware updates, retaining a large number of invalid test cases. In this embodiment, a targeted modeling approach using a "three-level decay network" is employed: at the hardware level, an exponential decay model quantifies the gradual weakening of correlation caused by hardware updates; at the software level, a nonlinear decay model characterizes the gradual change in failure probability caused by version iterations; and at the business level, a step decay model captures the precipitous drop in correlation caused by major API changes. Simultaneously, the decay factor is fused with spatiotemporal features to construct a correlation prediction model, which can accurately identify invalid use cases that are out of sync with the new platform, reducing test resource waste at its source. Third, existing technologies use fixed thresholds for screening, which are not adapted to different platform characteristics and do not adjust the judgment criteria based on feature energy, resulting in screening accuracy being greatly affected by the platform. In this embodiment, this problem is solved by a dynamic confidence threshold algorithm based on wavelet feature energy entropy: first, feature energy is extracted through wavelet transform, and the ratio of effective feature energy to total feature energy (wavelet feature energy entropy) is calculated. Then, a platform migration compensation factor (adapting to the sensitivity differences of parameters on different platforms) is used to dynamically generate a threshold, ensuring that the screening criteria match platform characteristics and are optimized in real time with changes in feature energy, avoiding the adaptability defects of fixed thresholds. Fourth, existing technologies lack an incremental learning mechanism, preventing models from being updated with new data and providing no visualization, resulting in low efficiency in operational decision-making. This application addresses this issue through a dual solution: an incremental learning closed-loop architecture and a cross-platform migration feature visualization system. The incremental learning closed loop uses the filtered test case execution data as incremental data to update feature extraction parameters and decay model coefficients, ensuring continuous model optimization with technological iterations. The visualization system generates parameter sensitivity heatmaps and platform migration path maps, overlaid with instruction set difference annotations, intuitively presenting parameter correlations and failure causes, significantly improving operational decision-making efficiency and overcoming the shortcomings of existing technologies in model iteration and result interpretation.As can be seen, the embodiments of this application provide a complete technical solution from four dimensions: feature extraction, attenuation modeling, threshold calculation, and iterative optimization, which can effectively adapt to the complex needs of server testing scenarios.

[0046] See Figure 2 This is a logical diagram of a test case intelligent screening method provided in an embodiment of this application.

[0047] First, the system retrieves data from the historical test database and the technology update database. The historical test database stores information such as hardware configuration, software version, and business parameters of past test cases, while the technology update database contains data on hardware iterations, software version updates, API changes, and other technological iterations. This data is then input into the spatiotemporal feature joint mining module. This module processes the data using a dual-channel attention neural network. The horizontal attention channel quantifies the coupling effect of hardware-software-business parameters to generate a parameter sensitivity matrix, while the vertical convolutional channel extracts cross-platform migration features to construct a platform migration feature map, achieving feature fusion and differentiation across spatiotemporal dimensions.

[0048] Next, the processed feature data is fed into a dynamic confidence threshold decision-maker. This decision-maker extracts feature energy through wavelet transform, combines it with a platform migration compensation factor, and uses an adaptive algorithm to calculate the dynamic confidence threshold. Afterward, the data enters a correlation decay modeling engine. This engine establishes a three-level decay network to quantify the decay factors corresponding to hardware iteration, software version updates, and API changes, and fuses them with spatiotemporal features to construct a test case correlation prediction model.

[0049] Finally, the data processed by the correlation decay modeling engine is input into the visual intelligent test case screening system. Based on the comparison between the correlation prediction results and the dynamic confidence threshold, the system screens out necessary test cases and test cases that can be eliminated, and generates visualization results such as parameter sensitivity heatmaps and platform migration path maps, which intuitively present the screening process and results and assist testers in making decisions.

[0050] The following is a specific example illustrating an embodiment of this application.

[0051] Step 1: Spatiotemporal feature extraction based on LSTM-Transformer hybrid architecture.

[0052] The input consists of 10,000+ historical use cases and technology update data (including GPU architecture / driver version / API call chain) and technology update data. First, an LSTM-Transformer hybrid architecture is adopted. The horizontal attention channel generates a parameter sensitivity matrix, the idea of ​​which is shown in the pseudocode, used for spatiotemporal coupled analysis of the input data. The vertical attention channel constructs a platform migration feature map, extracting cross-generational platform failure modes (such as instruction set difference mapping from Intel Skylake to Ice Lake) through a convolutional spatiotemporal encoder. The execution approach is as follows: lateral attention: generating parameter sensitivity matrices (e.g., the failure probability of thread_per_block>1024 in the Ampere architecture is as high as 82.7% based on existing empirical data); vertical convolution: constructing platform migration feature maps (e.g., AVX-512 instruction set difference mapping from Intel Skylake to Ice Lake).

[0053] # Pseudocode Implementation def horizontal_attention(test_parameters): Historical test parameters sensitivity_matrix = einsum('btd,dh->bth', parameters, W_att) # Tensor decomposition return softmax(sensitivity_matrix @ W_proj) # Coupling quantization Step 2: Model based on dynamic three-level attenuation network.

[0054] A three-layer decay network is established, consisting of a correlation network layer caused by hardware iteration, a platform nonlinear decay network layer, and a step decay driven by API changes (this example only uses the GPU; for other components, this API value needs to be assigned based on the manufacturer and test results). The three-layer decay network modeling function is shown below, using the GPU example parameters: Hardware layer: γ_{hw} = e^{-λ·Δt} (Hardware iteration leads to exponential decay of correlation, Δt = hardware iteration period, λ = 0.32 @NVIDIA); Software layer: γ_{sw} = \frac{1}{1+β·ΔVer} (Version differences cause non-linear decay, ΔVer = version difference degree, β = 1.57@CUDA); Business layer: γ_{biz} = Φ(ΔAPI) (Step decay driven by API change. When ΔAPI>50%, the step decays to 0.1).

[0055] Step 3: Establishment and definition of the adaptive decision-making mechanism.

[0056] Calculate the dynamic threshold:

[0057] in =Feature energy extracted by wavelet transform =Total feature energy extracted by wavelet transform, θ =Platform migration compensation factor (AMD platform θ=0.32, ARM platform θ=0.41).

[0058] Step 4: Criteria for real-time screening of test cases.

[0059] Required use case: Confidence = γ_{hw} \times γ_{sw} \times γ_{biz}>τ_{dynamic}; Excluded use cases: Confidence < 0.18.

[0060] Step 5: Visualize the decision.

[0061] Visualized decision-making can utilize correlation analysis algorithms for visualization. Based on confidence intervals, characteristic data change trends, change intervals, etc., feature vectors are formed and correlation analysis is performed with temporary features based on moving windows. The correlation value is output. After all the analysis is completed, feature normalization is performed to form parameter sensitivity heatmaps and platform migration path maps, thus achieving visualization effects.

[0062] GPU driver test case application scenarios: Input: 10,000+ test cases (including a technical update database covering Turing / Ampere / Hopper architectures), with key change data being assigned based on version output data: API change rate from CUDA 11.4 to CUDA 12.0 ΔAPI = 41%.

[0063] The system automatically identifies: (1) Attenuation path: γ_{sw}=e^{-1.57\times0.41}=0.32; (2) Highly sensitive parameter: thread_per_block>1024 failure probability in Ampere architecture.

[0064] Output: 237 necessary test cases, judged based on Confidence > 0.91; 7683 test cases can be removed, judged based on attenuation factor γ < 0.18; Results: Resource savings; test time reduced from 18.5 hours to 2.1 hours. As can be seen from the above examples, the embodiments of this application adopt a spatiotemporal feature extraction method based on the LSTM-Transformer hybrid architecture. This method can not only break the coupling effect of hardware-software-business parameters to accurately extract information features, but also capture cross-platform migration features, reducing the platform adaptation false alarm rate by more than 50%. Through dynamic decay modeling, the step decay caused by API changes can be quantified. Combined with version iteration descriptions, key change points are analyzed and reasonable thresholds are selected, effectively avoiding the drawback of undetected historical use case failures, significantly improving prediction accuracy and reducing the false alarm rate of case failures by 60%. The adaptive confidence threshold can dynamically adjust the selection criteria according to the characteristics of different platforms such as AMD, ARM, and X86, coupling the characteristics of each platform and integrating their requirements for the evaluation model, improving the credibility of the confidence interval, and increasing the accuracy of necessary use case selection by 30%. At the same time, it can generate parameter sensitivity heatmaps and platform migration path maps for visualized decision support, doubling the efficiency of operation and maintenance decisions.

[0065] In summary, the embodiments of this application have at least the following technical features and advantages: First, a spatiotemporal feature fusion extraction method based on a dual-channel attention mechanism of an LSTM-Transformer hybrid architecture is adopted, which can extract the horizontal parameter sensitivity matrix of the coupling effect between hardware-software-business parameters of test cases, and capture the vertical platform migration map to achieve effective feature differentiation; Second, this spatiotemporal feature extraction is integrated with the correlation decay factor tool to construct a predictive model that can clearly quantify features such as hardware exponential decay, software nonlinear decay, and business step decay. The time-effect decay is quantified through a three-layer model, and this application scenario is the first of its kind in a test case platform; Third, a dynamic confidence threshold algorithm based on wavelet feature energy entropy is proposed; Fourth, a test decision closed-loop architecture that supports incremental learning is constructed to form a deployment and application closed loop of "historical data → model update → real-time filtering → incremental data → parameter weight update → model update"; Fifth, a cross-platform migration feature visualization map generation system is created, which supports heatmap overlay and instruction set difference annotation. These protection points, through combined innovation, focus on new application areas and system scalability, and have significant industrial-grade value. For example, they can accurately eliminate invalid use cases in NVIDIA CUDA upgrade scenarios.

[0066] Corresponding to the above methods, embodiments of this application also provide a test case intelligent screening system. (See attached document) Figure 3 The system includes: Data acquisition module: used to collect historical test case data and technology update data. The historical test case data includes hardware configuration, software version, and business parameter information. The technology update data includes hardware iteration, software version update, and API change information. Spatiotemporal feature fusion module: A dual-channel attention unit is built using an LSTM-Transformer hybrid architecture, including a horizontal attention unit and a vertical convolution unit; the horizontal attention unit is used to quantify the coupling effect of hardware-software-business parameters to generate a parameter sensitivity matrix, and the vertical convolution unit is used to extract cross-platform migration features to construct a platform migration feature map; The correlation attenuation fusion module includes a three-level attenuation unit and a feature fusion unit. The three-level attenuation unit quantifies the attenuation factors corresponding to hardware, software, and API changes through exponential attenuation, nonlinear attenuation, and step attenuation models, respectively. The feature fusion unit is used to fuse the parameter sensitivity matrix, platform migration feature map, and attenuation factor to construct a test case correlation prediction model. Dynamic threshold decision module: used to extract feature energy through wavelet transform and calculate dynamic confidence threshold by combining platform migration compensation factor; Test Case Filtering Module: This module compares the predicted test case relevance with the dynamic confidence threshold to output necessary test cases and test cases that can be removed. Visualization module: used to generate parameter sensitivity heatmaps and platform migration path maps based on cross-platform migration features.

[0067] For details on the implementation principles and specific implementation methods of the system, please refer to the aforementioned method implementation examples, which will not be repeated here.

[0068] Embodiments of this application also provide a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0069] Optionally, in this embodiment, the storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0070] Embodiments of this application also provide an electronic device, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0071] Optionally, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor and the input / output device is connected to the processor.

[0072] Optionally, specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0073] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0074] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0075] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0076] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0077] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0078] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0079] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for intelligently filtering test cases, characterized in that, include: Spatiotemporal feature extraction steps: Obtain historical test case data and technology update data. The historical test case data includes hardware configuration, software version, and business parameter information. A dual-channel attention mechanism is built using an LSTM-Transformer hybrid architecture. The coupling effect of hardware-software-business parameters is quantified through the horizontal attention channel to generate a parameter sensitivity matrix. Cross-platform migration features are extracted through the vertical convolution channel to construct a platform migration feature map, thereby achieving cross-spatiotemporal feature fusion and feature differentiation. The correlation decay modeling and fusion steps are as follows: A three-level decay network is established, and the correlation decay caused by hardware iteration is quantified by the exponential decay model, the decay caused by software version iteration is quantified by the nonlinear decay model, and the decay driven by API change is quantified by the step decay model. The decay factors of each level are then output. The parameter sensitivity matrix, platform migration feature map and attenuation factors at each level are fused to construct a test case correlation prediction model. The dynamic threshold calculation steps are as follows: feature energy is extracted through wavelet transform, combined with the platform migration compensation factor, and the dynamic confidence threshold is calculated through an adaptive algorithm. Test case screening steps: Based on the comparison between the correlation results output by the test case correlation prediction model and the dynamic confidence threshold, and combined with preset elimination conditions, necessary test cases and test cases that can be eliminated are screened. Visualization steps: Generate parameter sensitivity heatmaps and platform migration path maps based on cross-platform migration characteristics to visualize the screening process and results.

2. The intelligent test case selection method according to claim 1, characterized in that, In the spatiotemporal feature extraction step, the horizontal attention channel maps hardware-software-business parameters into high-dimensional feature vectors through tensor decomposition, and then quantifies the parameter coupling effect through matrix projection and softmax activation function; the vertical convolution channel encodes the differences in hardware architecture and instruction set of different platforms through convolutional spatiotemporal encoder, and extracts the failure mode features of cross-generation platforms to construct a platform migration feature map.

3. The intelligent test case selection method according to claim 1, characterized in that, In the correlation decay modeling and fusion step: The exponential decay model expression for the hardware-level decay factor is γ_hw = e^(-λ·Δt), where Δt is the hardware iteration period and λ is the hardware decay coefficient. The nonlinear decay model expression for the software hierarchy decay factor is γ_sw = 1 / (1+β·ΔVer), where ΔVer is the software version difference degree and β is the software decay coefficient; When the API change rate exceeds a preset threshold, the business-level decay factor drops sharply. The fusion process maps the parameter sensitivity matrix and the attenuation factors of each level to the same feature space through matrix multiplication, and outputs the test case correlation prediction results by combining the weight allocation of the platform migration feature map.

4. The intelligent test case selection method according to claim 1, characterized in that, The dynamic threshold calculation step includes the process of extracting feature energy through wavelet transform, which includes: The fused features obtained from the spatiotemporal feature extraction step are decomposed into wavelet coefficients to obtain multi-scale wavelet coefficients. The total feature energy of the wavelet coefficients at each scale is calculated, and wavelet coefficients that meet the preset validity conditions are selected and their effective feature energy is calculated. The ratio of effective feature energy to total feature energy is calculated as the wavelet feature energy entropy. The adaptive algorithm uses the wavelet feature energy entropy as the core input and combines it with the platform migration compensation factor to calculate the dynamic confidence threshold.

5. The intelligent test case selection method according to claim 1, characterized in that, The parameter sensitivity matrix contains quantified failure probability values ​​for each business parameter under different hardware configurations and software versions; the platform migration feature map contains the mapping relationship between cross-platform architecture differences, instruction set differences and test case failure modes.

6. The intelligent test case selection method according to claim 1, characterized in that, In the visualization step, cross-platform instruction set difference annotations are superimposed on the generated parameter sensitivity heatmap. The annotations include the influence coefficients of different platform instruction set types and instruction set compatibility on the relevance of test cases.

7. The intelligent test case selection method according to any one of claims 1-6, characterized in that, Following the visualization step, the following is also included: Incremental learning closed-loop steps: Collect the filtered test case execution data as incremental data, update the model parameters based on the incremental data, complete the model update and return to the spatiotemporal feature extraction step to form a test decision closed loop; wherein, the model parameter update includes: calculating the error loss of the parameter sensitivity matrix based on the incremental data, optimizing the matrix projection parameters of the lateral attention channel through backpropagation; adjusting the feature encoding weights of the vertical convolution channel according to the cross-platform test results; and updating the decay coefficients (λ, β) of the three-layer decay network based on the correspondence between API changes and test case failures.

8. A test case intelligent screening system, characterized in that, include: Data acquisition module: used to collect historical test case data and technology update data. The historical test case data includes hardware configuration, software version, and business parameter information. The technology update data includes hardware iteration, software version update, and API change information. Spatiotemporal feature fusion module: A dual-channel attention unit is built using an LSTM-Transformer hybrid architecture, including a horizontal attention unit and a vertical convolution unit; the horizontal attention unit is used to quantify the coupling effect of hardware-software-business parameters to generate a parameter sensitivity matrix, and the vertical convolution unit is used to extract cross-platform migration features to construct a platform migration feature map; The correlation attenuation fusion module includes a three-level attenuation unit and a feature fusion unit. The three-level attenuation unit quantifies the attenuation factors corresponding to hardware, software, and API changes through exponential attenuation, nonlinear attenuation, and step attenuation models, respectively; the feature fusion unit is used to fuse the parameter sensitivity matrix, platform migration feature map, and attenuation factor to construct a test case correlation prediction model. Dynamic threshold decision module: used to extract feature energy through wavelet transform and calculate dynamic confidence threshold by combining platform migration compensation factor; Test Case Filtering Module: This module compares the predicted test case relevance with the dynamic confidence threshold to output necessary test cases and test cases that can be removed. Visualization module: used to generate parameter sensitivity heatmaps and platform migration path maps based on cross-platform migration features.

9. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method described in any one of claims 1 to 7 when it is run.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method as described in any one of claims 1 to 7.

Images