Software development defect detection method based on multi-dimensional data quality index fusion

Abstract

The application discloses a software development defect detection method based on multi-dimensional data quality index fusion, comprising the following steps: step one, collecting development process data and performing time alignment processing to generate an original quality data set; step two, extracting data quality indexes and generating a multi-dimensional data quality index matrix; step three, extracting a data quality change time sequence and generating a quality propagation correlation matrix; step four, inputting the quality time sequence matrix into an improved OmniAnomaly model of an implicit state entropy evolution layer and an anti-entropy recovery disturbance layer to generate a quality state entropy sequence; step five, generating a quality state collapse sequence; step six, generating a defect source positioning result; and step seven, generating a quality state collapse propagation path and outputting a software development defect detection result. The application realizes propagation perception and defect source positioning of software development defects.

Description

Technical Field

[0001] This invention relates to the fields of software engineering and machine learning technology, and in particular to a method for detecting software development defects based on the fusion of multidimensional data quality indicators. Background Technology

[0002] With the continuous expansion of software system scale and the widespread application of continuous integration and continuous deployment development models, defect detection in the software development process is gradually showing a trend of high frequency, dynamics, and multi-source heterogeneity. There are complex temporal correlations among code submissions, API calls, log recordings, database transactions, and test execution data during the development process. For the detection of software development defects, existing technologies mainly use rule threshold analysis, log anomaly detection, or deep learning-based temporal anomaly detection methods for identification. However, the following problems are common in actual software engineering scenarios:

[0003] Traditional software defect detection methods mostly rely on a single data source or a single anomaly indicator for analysis, making it difficult to comprehensively reflect the quality propagation relationship between different development objects during the software development process. This results in insufficient ability to identify hidden defects, persistent defects, and chain-propagation defects. Existing time-series anomaly detection models typically perform anomaly judgment based on reconstruction errors or prediction errors, lacking modeling of the discrete changes in the state during the evolution of hidden states, making it difficult to identify the quality state collapse phenomenon that exists during the formation of software defects. At the same time, existing methods mostly use static feature contribution analysis for anomaly localization, lacking an anti-entropy recovery perturbation mechanism based on continuous time evolution, and failing to accurately establish the correlation between defect sources and quality state propagation paths. In addition, current defect detection models usually ignore the time lag propagation relationship between development objects and the constraint relationship of state propagation direction, making it difficult to accurately identify defect propagation chains in complex software systems.

[0004] Therefore, how to provide a software development defect detection method based on the fusion of multi-dimensional data quality indicators is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a software development defect detection method based on the fusion of multi-dimensional data quality indicators. This invention constructs a multi-dimensional data quality indicator system and a quality propagation correlation matrix, and combines the hidden state entropy evolution module, entropy feedback recursion module, and anti-entropy recovery perturbation module in the improved OmniAnomaly model. The process of quality state entropy change is embedded in the hidden state recursive update process. By utilizing the quality state collapse interval and collapse propagation path, the propagation perception, defect source location, and risk level identification of software development defects are realized, thereby improving the accuracy of defect detection and propagation identification capabilities in complex software systems.

[0006] The software development defect detection method based on multi-dimensional data quality index fusion according to an embodiment of the present invention includes the following steps:

[0007] Step 1: Collect development process data during the software development process, perform time alignment processing, and generate the original quality dataset;

[0008] Step 2: Based on the original quality dataset, extract data quality indicators and generate a multidimensional data quality indicator matrix;

[0009] Step 3: Based on the multidimensional data quality index matrix, extract the time series of data quality changes, and generate a quality propagation correlation matrix according to the time lag relationship;

[0010] Step 4: Generate a quality time series matrix based on the multidimensional data quality index matrix, and input the quality time series matrix into the improved OmniAnomaly model with hidden state entropy evolution layer and anti-entropy recovery perturbation layer to generate a quality state entropy sequence;

[0011] Step 5: Based on the mass state entropy sequence, calculate the mass state entropy change results between consecutive time windows to generate a mass state collapse sequence;

[0012] Step 6: Based on the quality state collapse sequence, perform anti-entropy recovery perturbation processing on the target data quality indicators to generate defect source location results;

[0013] Step 7: Based on the defect source location results and the quality propagation correlation matrix, generate the quality state collapse propagation path and output the software development defect detection results.

[0014] Optionally, step one specifically includes:

[0015] Collect code commit records, API call records, log records, database transaction records, test execution records, and build execution records throughout the software development lifecycle, and classify and store them according to module identifier, API identifier, service instance identifier, and database transaction identifier to generate a multi-source development process dataset;

[0016] Based on code submission time, API call time, log generation time, database transaction time, test execution time, and build execution time, extract the corresponding timestamp information and generate a continuous time segment sequence according to the timestamp interval;

[0017] Perform time interpolation on missing timestamps in a continuous time segment sequence and generate a time-compensated sequence based on the time interval between adjacent timestamps;

[0018] Based on the time compensation sequence, the time offset between different data sources is calculated, and the time axis remapping processing is performed on code commit records, interface call records, log records, database transaction records, test execution records and build execution records according to the time offset to generate a unified time base sequence;

[0019] Dynamic window partitioning is performed based on the data change density in a unified time base sequence. The code commit records, API call records, log records, database transaction records, test execution records, and build execution records are then windowed according to the dynamic window span to generate windowed development process data blocks.

[0020] Based on the time segment number, data source number, and development object identifier in the windowed development process data block, time association and data source association are established to generate the original quality dataset.

[0021] Optionally, step two specifically includes:

[0022] Perform field splitting and field type mapping processing on the code commit records, API call records, log records, database transaction records, test execution records, and build execution records in the original quality dataset to generate corresponding data attribute sequences;

[0023] The data integrity index is calculated based on the number of missing fields and the total number of fields in the data attribute sequence, and the data consistency index is calculated based on the number of times the field values ​​in the same field position are inconsistent.

[0024] The interface response fluctuation index is calculated based on the change in response time in the interface call record, and the test coverage fluctuation index is calculated based on the change in test coverage in the test execution record.

[0025] The log anomaly density index is calculated based on the number of abnormal logs in the log records and the total number of logs, and the data distribution drift index is calculated based on the change in the distribution of transaction attributes in the database transaction records.

[0026] The version commit disturbance metric is calculated based on the change in the number of commits in the code commit record, and the module call stability metric is calculated based on the change in the call path in the interface call record.

[0027] Based on the data integrity index, data consistency index, interface response fluctuation index, test coverage fluctuation index, log anomaly density index, data distribution drift index, version commit disturbance index, and module call stability index, the index fluctuation entropy corresponding to these indices is normalized to generate a standardized quality index sequence.

[0028] The standardized quality indicator sequence is matrix arranged according to the time segment number and the development object identifier to generate a multidimensional data quality indicator matrix.

[0029] Optionally, step three specifically includes:

[0030] The data quality indicators in the multidimensional data quality indicator matrix are arranged in a continuous time alignment according to the time segment number to generate the corresponding data quality change time series.

[0031] Based on adjacent time segments in the data quality change time series, calculate the difference of data quality indicators and the magnitude of indicator changes, and generate a sequence of indicator change trajectories in chronological order;

[0032] Multi-scale time offset traversal processing is performed on different data quality indicators in the indicator change trajectory sequence to generate indicator correlation change sequences under different time offset conditions;

[0033] The time lag correlation value is calculated based on the number of synchronous changes in the same direction and the number of synchronous changes in the magnitude of changes in the index correlation change sequence, and a time lag correlation sequence is generated according to the time offset.

[0034] Perform reverse time filtering on the time lag correlation values ​​in the time lag correlation sequence, and generate quality propagation correlation edges according to the time sequence;

[0035] A sparse matrix mapping process is performed based on the starting development object identifier, the target development object identifier, and the time lag association value in the quality propagation association edges to generate a quality propagation association matrix.

[0036] Optionally, step four specifically includes:

[0037] Input the quality time series matrix into the improved OmniAnomaly model;

[0038] The improved OmniAnomaly model includes a temporal recursive encoding module, a hidden state entropy evolution module, an entropy feedback recursive module, and an anti-entropy perturbation recovery module.

[0039] The temporal recursive coding module takes the quality temporal matrix as input and performs temporal recursive coding processing on the data quality indicators in continuous time segments to generate the hidden state vector sequence in the corresponding time segments.

[0040] The temporal recursive coding module includes a gated recursive unit and a hidden state distribution mapping unit;

[0041] The gated recursive unit reads the data quality indicators in the quality time series matrix in chronological order and updates the hidden state vector in the current time segment based on the hidden state vector in the previous time segment.

[0042] The hidden state distribution mapping unit performs probability distribution mapping processing on the hidden state vector to generate the hidden state distribution sequence in the corresponding time segment;

[0043] The hidden state entropy evolution module takes the hidden state distribution sequence as input, performs probability statistical processing on the hidden state distribution in each time segment, and generates a hidden state probability sequence.

[0044] Calculate the degree of dispersion of the state distribution in the corresponding time segment based on the probability change magnitude in the hidden state probability sequence;

[0045] The state distribution discreteness is continuously evolved and arranged according to time order to generate a mass state entropy sequence;

[0046] Calculate the entropy change difference between adjacent time segments in the mass state entropy sequence to generate a mass state entropy change sequence;

[0047] Perform collapse judgment processing on the continuous unidirectional entropy gradient interval in the mass state entropy change sequence to generate the mass state collapse interval.

[0048] The entropy feedback recursive module performs hidden state entropy coupling shaping processing based on the mass state entropy value in the previous time segment to generate the hidden state distribution sequence in the corresponding time segment.

[0049] The anti-entropy recovery perturbation module performs anti-entropy collapse tracing processing based on the data quality indicators corresponding to the quality state collapse interval, and generates an entropy recovery change sequence.

[0050] Optionally, the hidden state entropy coupling shaping process specifically includes:

[0051] Extract the mass state entropy value and the corresponding hidden state vector from the previous time segment;

[0052] The entropy coupling strength is calculated based on the correspondence between the mass state entropy value and the hidden state vector, and a sequence of entropy coupling coefficients is generated.

[0053] Map the entropy coupling coefficient sequence to the hidden state update path corresponding to the current time segment to generate the entropy coupling update path;

[0054] Based on the entropy coupling update path, a joint shaping process is performed on the hidden state propagation path and the hidden state retention path in the current time segment to generate a shaped hidden state sequence.

[0055] Based on the shaped hidden state sequence, the hidden state vector in the current time segment is reconstructed to generate an entropy-coupled hidden state vector.

[0056] The entropy-coupled hidden state vector is input into the hidden state distribution mapping unit to generate the hidden state distribution sequence in the corresponding time segment.

[0057] Optionally, the anti-entropy collapse tracing process specifically includes:

[0058] Extract the target data quality indicators corresponding to the quality state collapse interval;

[0059] Perform a gradual perturbation on the target data quality indicators while keeping the time series corresponding to the remaining data quality indicators unchanged;

[0060] The quality indicators of the target data after asymptotic perturbation are re-input into the time-series recursive encoding module to generate a perturbation hidden state distribution sequence.

[0061] The degree of dispersion of the state distribution in the corresponding time segment is calculated based on the perturbation hidden state distribution sequence to generate a perturbation state discrete sequence;

[0062] Perform difference calculation between the discrete sequence of the disturbed state and the discrete sequence of the original state to generate an entropy recovery change sequence;

[0063] Based on the continuous change relationship in the entropy recovery change sequence, collapse source tracing analysis is performed to generate a collapse release contribution sequence.

[0064] Optionally, step five specifically includes:

[0065] Perform continuous time alignment processing on each time segment in the mass state entropy sequence to generate a continuous mass state entropy trajectory sequence;

[0066] Based on adjacent time segments in a continuous mass state entropy trajectory sequence, calculate the mass state entropy difference and entropy change amplitude to generate a mass state entropy change sequence.

[0067] Perform entropy gradient direction detection processing on continuous time segments in the mass state entropy change sequence to generate an entropy gradient change sequence;

[0068] The cumulative change of entropy gradient is calculated based on the continuous unidirectional entropy gradient intervals in the entropy gradient change sequence, and an entropy gradient accumulation sequence is generated.

[0069] Perform state discreteness detection processing on the mass state entropy values ​​corresponding to the continuous decreasing intervals in the entropy gradient accumulation sequence to generate the mass state convergence interval;

[0070] Based on the time continuity length, entropy change amplitude, and state discreteness change in the convergence interval of the quality state, a collapse interval mapping process is performed to generate a quality state collapse sequence.

[0071] Optionally, step six specifically includes:

[0072] Perform collapse contribution correlation analysis on each collapse interval in the quality state collapse sequence to generate a set of target data quality indicators;

[0073] A hierarchical perturbation partitioning process is performed on each target data quality indicator in the target data quality indicator set to generate the corresponding asymptotic perturbation sequence;

[0074] Based on the progressive perturbation sequence, incremental perturbation update processing is performed on the index values ​​corresponding to the target data quality index in chronological order, while keeping the time series corresponding to the non-target data quality index unchanged.

[0075] The perturbation-updated target data quality index is re-input into the improved OmniAnomaly model, and the perturbation hidden state distribution sequence and perturbation quality state entropy sequence in the corresponding time segment are regenerated.

[0076] The entropy recovery change is calculated based on the difference in state discreteness between the perturbed mass state entropy sequence and the original mass state entropy sequence, and an entropy recovery trajectory sequence is generated.

[0077] Perform collapse release contribution calculation on continuous recovery intervals in the entropy recovery trajectory sequence to generate an entropy recovery contribution sequence corresponding to the target data quality index;

[0078] Based on the ranking of recovery contribution in the entropy recovery contribution sequence and the correlation of collapse propagation time corresponding to the collapse interval of quality state, the defect source location result is generated.

[0079] Optionally, step seven specifically includes:

[0080] Perform collapse propagation mapping processing on the target development object identifier, target data quality index and quality state collapse interval in the defect source location results to generate a defect propagation node sequence;

[0081] Based on the time lag correlation value and propagation direction constraint relationship in the quality propagation correlation matrix, time overlap propagation correlation matching processing is performed on adjacent nodes in the defect propagation node sequence to generate a node propagation correlation sequence.

[0082] Perform collapse propagation continuity detection processing on each propagation node in the node propagation association sequence to generate a continuous propagation interval sequence;

[0083] Based on the propagation time length, number of propagation nodes, and propagation association strength in the continuous propagation interval sequence, propagation path connection processing is performed in the order of propagation direction to generate a quality state collapse propagation path sequence.

[0084] Perform a propagation direction consistency check on each propagation path in the mass state collapse propagation path sequence to generate a propagation direction sequence;

[0085] Based on the number of propagation direction changes and the change in propagation association strength in the propagation direction sequence, a collapse propagation risk level classification process is performed to generate a path risk level sequence.

[0086] Based on the path risk level sequence, the quality state collapse propagation path sequence, and the defect source location results, corresponding defect nodes, propagation paths, and software development defect detection results corresponding to risk levels are generated.

[0087] The beneficial effects of this invention are:

[0088] This invention constructs a multi-dimensional data quality indicator system integrating code commit records, API call records, log records, database transaction records, test execution records, and build execution records. Combined with time alignment processing, dynamic window partitioning processing, and the collaborative design of a quality propagation correlation matrix, it addresses the problems of complex defect propagation relationships, hidden quality state evolution, and insufficient anomaly localization accuracy in software development. It proposes a quality propagation modeling strategy based on time lag correlation analysis, multi-scale time offset traversal, and collapse propagation mapping, significantly improving the ability to express temporal correlations between development process data and identify defect propagation. In the quality state modeling stage, an improved OmniAnomaly model is introduced, employing a time-series recursive encoding module and hidden... The joint design of the state entropy evolution module, entropy feedback recursion module, and anti-entropy recovery perturbation module embeds the quality state entropy change process into the hidden state recursive update process. It utilizes the continuous unidirectional entropy gradient interval to construct the quality state collapse interval, effectively enhancing the model's ability to perceive hidden and persistent defects in complex software systems. In the defect source localization stage, a progressive perturbation and anti-entropy recovery mechanism is constructed. The defect source localization relationship is established by combining the contribution of collapse relief and the correlation between collapse propagation time, avoiding localization bias caused by single anomaly index analysis. Finally, the software development defect detection results are generated using the quality state collapse propagation path and collapse propagation risk level, realizing the perception of the time propagation of software development defects, the identification of quality state collapse, and the intelligent localization of defect sources. Attached Figure Description

[0089] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0090] Figure 1 This is a schematic diagram of the overall process of the software development defect detection method based on the fusion of multi-dimensional data quality indicators proposed in this invention;

[0091] Figure 2 This is an overall structural diagram of the improved OmniAnomaly model in the software development defect detection method based on the fusion of multi-dimensional data quality indicators proposed in this invention. Detailed Implementation

[0092] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0093] refer to Figures 1-2 A software development defect detection method based on the fusion of multi-dimensional data quality indicators includes the following steps:

[0094] Step 1: Collect development process data during the software development process, perform time alignment processing, and generate the original quality dataset;

[0095] Step 2: Based on the original quality dataset, extract data quality indicators and generate a multidimensional data quality indicator matrix;

[0096] Step 3: Based on the multidimensional data quality index matrix, extract the time series of data quality changes, and generate a quality propagation correlation matrix according to the time lag relationship;

[0097] Step 4: Generate a quality time series matrix based on the multidimensional data quality index matrix, and input the quality time series matrix into the improved OmniAnomaly model with the hidden state entropy evolution layer and the anti-entropy recovery perturbation layer to generate a quality state entropy sequence;

[0098] Step 5: Based on the mass state entropy sequence, calculate the mass state entropy change results between consecutive time windows to generate a mass state collapse sequence;

[0099] Step 6: Based on the quality state collapse sequence, perform anti-entropy recovery perturbation processing on the target data quality indicators to generate defect source location results;

[0100] Step 7: Based on the defect source location results and the quality propagation correlation matrix, generate the quality state collapse propagation path and output the software development defect detection results.

[0101] In this embodiment, step one specifically includes:

[0102] Collect code commit records, API call records, log records, database transaction records, test execution records, and build execution records throughout the software development lifecycle, and classify and store them according to module identifier, API identifier, service instance identifier, and database transaction identifier to generate a multi-source development process dataset;

[0103] Based on code submission time, API call time, log generation time, database transaction time, test execution time, and build execution time, extract the corresponding timestamp information and generate a continuous time segment sequence according to the timestamp interval;

[0104] Perform time interpolation on missing timestamps in a continuous time segment sequence and generate a time-compensated sequence based on the time interval between adjacent timestamps;

[0105] Based on the time compensation sequence, the time offset between different data sources is calculated, and the time axis remapping processing is performed on code commit records, interface call records, log records, database transaction records, test execution records and build execution records according to the time offset to generate a unified time base sequence;

[0106] Dynamic window partitioning is performed based on the data change density in a unified time base sequence. The code commit records, API call records, log records, database transaction records, test execution records, and build execution records are then windowed according to the dynamic window span to generate windowed development process data blocks.

[0107] Based on the time segment number, data source number, and development object identifier in the windowed development process data block, time association and data source association are established to generate the original quality dataset.

[0108] In this implementation, time interpolation processing constructs a time compensation sequence using the rate of change of time intervals between adjacent time segments. When the time interval between adjacent time segments exceeds the continuous time density threshold, compensation time nodes are inserted into the data positions corresponding to the missing timestamps. Time axis remapping processing uses a unified time base to offset and align timestamps from different data sources, and rearranges the corresponding data records according to the time offset. Dynamic window partitioning processing constructs a time change density sequence using the interface call density, log generation density, and code submission density in the time segment, and dynamically adjusts the window span according to the time change density sequence, with high-density time areas corresponding to smaller window spans and low-density time areas corresponding to larger window spans. Data blocks in the windowed development process are arranged continuously in chronological order and established with a corresponding relationship with the development object identifier.

[0109] In this embodiment, step two specifically includes:

[0110] Perform field splitting and field type mapping processing on the code commit records, API call records, log records, database transaction records, test execution records, and build execution records in the original quality dataset to generate corresponding data attribute sequences;

[0111] The data integrity index is calculated based on the number of missing fields and the total number of fields in the data attribute sequence, and the data consistency index is calculated based on the number of times the field values ​​in the same field position are inconsistent.

[0112] The interface response fluctuation index is calculated based on the change in response time in the interface call record, and the test coverage fluctuation index is calculated based on the change in test coverage in the test execution record.

[0113] The log anomaly density index is calculated based on the number of abnormal logs in the log records and the total number of logs, and the data distribution drift index is calculated based on the change in the distribution of transaction attributes in the database transaction records.

[0114] The version commit disturbance metric is calculated based on the change in the number of commits in the code commit record, and the module call stability metric is calculated based on the change in the call path in the interface call record.

[0115] Based on the data integrity index, data consistency index, interface response fluctuation index, test coverage fluctuation index, log anomaly density index, data distribution drift index, version commit disturbance index, and module call stability index, the index fluctuation entropy corresponding to these indices is normalized to generate a standardized quality index sequence.

[0116] The standardized quality indicator sequence is matrix arranged according to the time segment number and the development object identifier to generate a multidimensional data quality indicator matrix.

[0117] In this implementation, the field type mapping process includes the classification mapping of character fields, numeric fields, and time-series fields. Character fields are generated into discrete attribute sequences using field value encoding, numeric fields are generated into continuous attribute sequences using interval segmentation, and time-series fields are generated into time-series attribute sequences using time-increasing sorting. The change in transaction attribute distribution is calculated using the change rate of transaction attribute frequency in continuous time segments. The change in call path is calculated by jointly using the change rate of the number of interface call nodes and the change rate of call path length in adjacent time segments. The index fluctuation entropy is calculated using the probability distribution of index change in continuous time windows, and normalized interval mapping is performed according to the change interval corresponding to the index fluctuation entropy.

[0118] In this embodiment, step three specifically includes:

[0119] The data quality indicators in the multidimensional data quality indicator matrix are arranged in a continuous time alignment according to the time segment number to generate the corresponding data quality change time series.

[0120] Based on adjacent time segments in the data quality change time series, calculate the difference of data quality indicators and the magnitude of indicator changes, and generate a sequence of indicator change trajectories in chronological order;

[0121] Multi-scale time offset traversal processing is performed on different data quality indicators in the indicator change trajectory sequence to generate indicator correlation change sequences under different time offset conditions;

[0122] The time lag correlation value is calculated based on the number of synchronous changes in the same direction and the number of synchronous changes in the magnitude of changes in the index correlation change sequence, and a time lag correlation sequence is generated according to the time offset.

[0123] Perform reverse time filtering on the time lag correlation values ​​in the time lag correlation sequence, and generate quality propagation correlation edges according to the time sequence;

[0124] A sparse matrix mapping process is performed based on the starting development object identifier, the target development object identifier, and the time lag association value in the quality propagation association edges to generate a quality propagation association matrix.

[0125] In this implementation, the multi-scale time offset traversal processing includes short-period time offset traversal and long-period time offset traversal. The short-period time offset traversal corresponds to the index change relationship between consecutive adjacent time segments, while the long-period time offset traversal corresponds to the index change relationship across time windows. The number of changes in the same direction is counted by the number of times the index change direction is consistent in consecutive time segments. The number of synchronous changes in change amplitude is counted by the range of the difference in change amplitude of different data quality indicators under the same time offset condition. The reverse time direction filtering processing deletes the reverse propagation edge by using the correspondence between the time offset direction and the time segment order. The sparse matrix mapping processing performs matrix position mapping by using the correlation strength range corresponding to the time lag correlation value, and performs zero-value filling processing on the low correlation strength position.

[0126] In this embodiment, step four specifically includes:

[0127] Input the quality time series matrix into the improved OmniAnomaly model;

[0128] The improved OmniAnomaly model includes a temporal recursive encoding module, a hidden state entropy evolution module, an entropy feedback recursive module, and an anti-entropy perturbation recovery module;

[0129] The temporal recursive coding module takes the quality temporal matrix as input, performs temporal recursive coding on the data quality indicators in continuous time segments, and generates the hidden state vector sequence in the corresponding time segment.

[0130] The temporal recursive coding module includes a gated recursive unit and a hidden state distribution mapping unit;

[0131] The gated recursive unit reads the data quality indicators in the quality time series matrix in chronological order and updates the hidden state vector in the current time segment based on the hidden state vector in the previous time segment.

[0132] The hidden state distribution mapping unit performs probability distribution mapping processing on the hidden state vector to generate the hidden state distribution sequence in the corresponding time segment;

[0133] The hidden state entropy evolution module takes the hidden state distribution sequence as input, performs probabilistic statistical processing on the hidden state distribution in each time segment, and generates a hidden state probability sequence.

[0134] Calculate the degree of dispersion of the state distribution in the corresponding time segment based on the probability change magnitude in the hidden state probability sequence;

[0135] The state distribution discreteness is continuously evolved and arranged according to time order to generate a mass state entropy sequence;

[0136] Calculate the entropy change difference between adjacent time segments in the mass state entropy sequence to generate a mass state entropy change sequence;

[0137] Perform collapse judgment processing on the continuous unidirectional entropy gradient interval in the mass state entropy change sequence to generate the mass state collapse interval.

[0138] The entropy feedback recursive module performs hidden state entropy coupling shaping processing based on the mass state entropy value in the previous time segment, generating the hidden state distribution sequence in the corresponding time segment.

[0139] The anti-entropy recovery perturbation module performs anti-entropy collapse tracing processing based on the data quality indicators corresponding to the quality state collapse interval, generating an entropy recovery change sequence.

[0140] In this embodiment, the hidden state entropy coupling shaping process specifically involves:

[0141] Extract the mass state entropy value and the corresponding hidden state vector from the previous time segment;

[0142] The entropy coupling strength is calculated based on the correspondence between the mass state entropy value and the hidden state vector, and a sequence of entropy coupling coefficients is generated.

[0143] Map the entropy coupling coefficient sequence to the hidden state update path corresponding to the current time segment to generate the entropy coupling update path;

[0144] Based on the entropy coupling update path, a joint shaping process is performed on the hidden state propagation path and the hidden state retention path in the current time segment to generate a shaped hidden state sequence.

[0145] Based on the shaped hidden state sequence, the hidden state vector in the current time segment is reconstructed to generate an entropy-coupled hidden state vector.

[0146] The entropy-coupled hidden state vector is input into the hidden state distribution mapping unit to generate the hidden state distribution sequence in the corresponding time segment.

[0147] In this embodiment, the anti-entropy collapse tracing process is specifically as follows:

[0148] Extract the target data quality indicators corresponding to the quality state collapse interval;

[0149] Perform a gradual perturbation on the target data quality indicators while keeping the time series corresponding to the remaining data quality indicators unchanged;

[0150] The quality indicators of the target data after asymptotic perturbation are re-input into the time-series recursive encoding module to generate a perturbation hidden state distribution sequence.

[0151] The degree of dispersion of the state distribution in the corresponding time segment is calculated based on the perturbation hidden state distribution sequence to generate a perturbation state discrete sequence;

[0152] Perform difference calculation between the discrete sequence of the disturbed state and the discrete sequence of the original state to generate an entropy recovery change sequence;

[0153] Based on the continuous change relationship in the entropy recovery change sequence, collapse source tracing analysis is performed to generate a collapse release contribution sequence.

[0154] In this implementation, the hidden state distribution mapping process constructs a probability distribution using the mean and variance parameters corresponding to the hidden state vectors, and calculates the degree of dispersion of the state distribution based on the probability discrete change rate in the probability distribution; the continuous unidirectional entropy gradient interval is constructed using the consistent entropy change direction relationship in multiple consecutive time segments; the quality state collapse corresponds to the process of decreasing degrees of freedom of the hidden state distribution in consecutive time segments; the asymptotic perturbation process updates the index values ​​corresponding to the target data quality index with progressively increasing perturbation intensity, and stops the incremental perturbation update when the entropy recovery change in consecutive time segments enters the stable change interval; the propagation path connection process deletes duplicate propagation nodes and generates the corresponding quality state collapse propagation path in chronological order;

[0155] Both the improved OmniAnomaly model and the OmniAnomaly model adopt a time-series coding method based on a gated recursive structure. They perform recursive hidden state update processing on multidimensional data in continuous time segments and generate a state distribution sequence in the corresponding time segment based on the hidden state distribution, thereby realizing time correlation modeling in development process data.

[0156] The improved OmniAnomaly model introduces an entropy feedback recursive module within the gated recursive structure, a hidden state entropy evolution module in the hidden state distribution generation stage, and an anti-entropy recovery perturbation module in the quality state collapse interval. The entropy feedback recursive module uses the quality state entropy value in the previous time segment to adjust the hidden state update path in the current time segment. The hidden state entropy evolution module generates a quality state entropy change sequence based on the discreteness of the state distribution in continuous time segments. The anti-entropy recovery perturbation module performs asymptotic perturbation processing on the target data quality index and regenerates the corresponding hidden state distribution sequence.

[0157] The improved OmniAnomaly model incorporates the quality state entropy change process into the hidden state recursive update process, enabling the hidden state update process to have a continuous entropy evolution constraint relationship; it performs quality state collapse judgment processing based on the continuous unidirectional entropy gradient interval, forming a quality state propagation chain; and it establishes the defect source localization relationship based on the entropy recovery change process after asymptotic perturbation, enhancing the state propagation identification capability and defect localization accuracy in the software development defect detection process.

[0158] In this embodiment, step five specifically includes:

[0159] Perform continuous time alignment processing on each time segment in the mass state entropy sequence to generate a continuous mass state entropy trajectory sequence;

[0160] Based on adjacent time segments in a continuous mass state entropy trajectory sequence, calculate the mass state entropy difference and entropy change amplitude to generate a mass state entropy change sequence.

[0161] Perform entropy gradient direction detection processing on continuous time segments in the mass state entropy change sequence to generate an entropy gradient change sequence;

[0162] The cumulative change of entropy gradient is calculated based on the continuous unidirectional entropy gradient intervals in the entropy gradient change sequence, and an entropy gradient accumulation sequence is generated.

[0163] Perform state discreteness detection processing on the mass state entropy values ​​corresponding to the continuous decreasing intervals in the entropy gradient accumulation sequence to generate the mass state convergence interval;

[0164] Based on the time continuity length, entropy change amplitude, and state discreteness change in the convergence interval of the quality state, a collapse interval mapping process is performed to generate a quality state collapse sequence.

[0165] In this implementation, the entropy gradient direction detection process uses the sign of the entropy difference between adjacent time segments to generate an entropy gradient direction sequence; the continuous unidirectional entropy gradient interval is divided by the number of times the entropy gradient direction is consistent in the continuous time segments; the cumulative change of entropy gradient is calculated by the sum of the entropy change amplitudes in the continuous unidirectional entropy gradient interval; the state discreteness detection process is calculated by the rate of change of the discrete value of the hidden state distribution in the continuous decreasing interval; the collapse interval mapping process performs joint mapping by the interval weights corresponding to the continuous time length, entropy change amplitude, and state discreteness change, and generates a quality state collapse marker in the corresponding time segment according to the mapping result.

[0166] In this embodiment, step six specifically includes:

[0167] Perform collapse contribution correlation analysis on each collapse interval in the quality state collapse sequence to generate a set of target data quality indicators;

[0168] A hierarchical perturbation partitioning process is performed on each target data quality indicator in the target data quality indicator set to generate the corresponding asymptotic perturbation sequence;

[0169] Based on the progressive perturbation sequence, incremental perturbation update processing is performed on the index values ​​corresponding to the target data quality index in chronological order, while keeping the time series corresponding to the non-target data quality index unchanged.

[0170] The perturbation-updated target data quality index is re-input into the improved OmniAnomaly model, and the perturbation hidden state distribution sequence and perturbation quality state entropy sequence in the corresponding time segment are regenerated.

[0171] The entropy recovery change is calculated based on the difference in state discreteness between the perturbed mass state entropy sequence and the original mass state entropy sequence, and an entropy recovery trajectory sequence is generated.

[0172] Perform collapse release contribution calculation on continuous recovery intervals in the entropy recovery trajectory sequence to generate an entropy recovery contribution sequence corresponding to the target data quality index;

[0173] Based on the ranking of recovery contribution in the entropy recovery contribution sequence and the correlation of collapse propagation time corresponding to the collapse interval of quality state, the defect source location result is generated.

[0174] In this implementation, the collapse contribution correlation analysis is performed by calculating the correlation strength between the frequency of changes in data quality indicators and the magnitude of entropy changes within the collapse interval of the quality state; the hierarchical perturbation division is performed by generating a multi-level perturbation intensity sequence from the indicator change intervals corresponding to the target data quality indicators; the incremental perturbation update is performed by updating the indicator values ​​corresponding to the target data quality indicators by progressively increasing the perturbation amplitude; the state dispersion difference is calculated by the difference in the discrete values ​​of the hidden state distribution before and after the perturbation; the collapse release contribution is jointly calculated by the cumulative value of the entropy recovery change in the continuous recovery interval and the interval weight corresponding to the recovery duration; and the collapse propagation time correlation is constructed by the temporal continuity relationship between the collapse intervals of the quality state and the consistency relationship of the propagation direction.

[0175] In this embodiment, step seven specifically includes:

[0176] Perform collapse propagation mapping processing on the target development object identifier, target data quality index and quality state collapse interval in the defect source location results to generate a defect propagation node sequence;

[0177] Based on the time lag correlation value and propagation direction constraint relationship in the quality propagation correlation matrix, time overlap propagation correlation matching processing is performed on adjacent nodes in the defect propagation node sequence to generate a node propagation correlation sequence.

[0178] Perform collapse propagation continuity detection processing on each propagation node in the node propagation association sequence to generate a continuous propagation interval sequence;

[0179] Based on the propagation time length, number of propagation nodes, and propagation association strength in the continuous propagation interval sequence, propagation path connection processing is performed in the order of propagation direction to generate a quality state collapse propagation path sequence.

[0180] Perform a propagation direction consistency check on each propagation path in the mass state collapse propagation path sequence to generate a propagation direction sequence;

[0181] Based on the number of propagation direction changes and the change in propagation association strength in the propagation direction sequence, a collapse propagation risk level classification process is performed to generate a path risk level sequence.

[0182] Based on the path risk level sequence, the quality state collapse propagation path sequence, and the defect source location results, corresponding defect nodes, propagation paths, and software development defect detection results corresponding to risk levels are generated.

[0183] In this implementation, the temporal overlap propagation association matching process uses the temporal overlap ratio and temporal offset interval between the quality state collapse intervals corresponding to adjacent propagation nodes for joint matching; the collapse propagation continuity detection process uses the change rate of time interval and the change rate of propagation direction between continuous propagation nodes for continuity calculation; the propagation path connection process uses the connection weight corresponding to the temporal order of propagation nodes and the propagation association strength to execute path combination; the propagation direction consistency detection process uses the number of times the propagation direction is consistent between continuous propagation nodes for statistical analysis; and the collapse propagation risk level classification process uses the interval weight corresponding to the change in propagation association strength, propagation path length, and the number of times the propagation direction changes for joint calculation, and generates the corresponding path risk level label according to the calculation results.

[0184] Example 1: To verify the feasibility of this invention in practice, it was applied to a software development defect detection scenario for a large-scale internet payment platform. The internet payment platform includes an order service module, a payment service module, a risk control service module, a database transaction module, and a message queue service module. The entire platform is deployed using a microservice architecture, with an average daily API call volume exceeding 360 million times and continuous integration build tasks exceeding 12,000. During development, there are numerous code commits, log recordings, database transactions, and API call activities. Due to the large number of system modules and the complex time dependencies between different services, when a local service experiences an abnormal state, cross-module propagation is easily formed, making it difficult for traditional defect detection methods based on log anomalies or refactoring errors to accurately identify the defect source location and defect propagation path.

[0185] In practical applications, the process begins by collecting code commit records, API call records, log records, database transaction records, test execution records, and build execution records throughout the software development lifecycle. These records are then categorized and stored according to module identifiers, API identifiers, and service instance identifiers. Next, a timeline remapping process is performed based on code commit time, log generation time, and API call time to construct a unified time base sequence. Following this, dynamic window partitioning is performed based on the API call density, log generation density, and code commit density within the unified time base sequence to generate windowed development process data blocks. Field splitting and field type mapping are then performed on the data attributes within these windowed development process data blocks. Finally, data integrity metrics, data consistency metrics, API response fluctuation metrics, test coverage fluctuation metrics, log anomaly density metrics, and module call stability metrics are calculated to construct a multi-dimensional data quality metric matrix.

[0186] In the quality propagation modeling phase, continuous time alignment is performed on each data quality indicator in the multidimensional data quality indicator matrix, and an indicator change trajectory sequence is constructed based on the indicator change amplitude in different time windows. Subsequently, multi-scale time offset traversal processing is used to analyze the time lag correlation between different development objects, and the reverse time propagation edges are deleted to generate a quality propagation correlation matrix. The quality time series matrix is ​​then input into the improved OmniAnomaly model. The model's internal time-series recursive encoding module performs recursive hidden state update processing on the data quality indicators in continuous time segments, and the hidden state entropy evolution module performs continuous entropy evolution analysis on the state dispersion in the hidden state distribution. When the quality state entropy continuously decreases in multiple consecutive time segments, the model generates the corresponding quality state collapse interval. For the target data quality indicator in the quality state collapse interval, the anti-entropy recovery perturbation module regenerates the perturbation hidden state distribution sequence according to the increasing perturbation intensity, and calculates the collapse relief contribution based on the entropy recovery change, thereby locating the corresponding defect source module.

[0187] In this embodiment, during system operation, the payment service module experienced an abnormal release of the database connection pool starting 37 minutes after the version upgrade. Since no obvious error logs were generated in the initial stage of the anomaly, traditional log detection methods could not identify the anomaly in a timely manner. As the database transaction execution time continued to increase, the quality state entropy corresponding to the payment service module continuously decreased over five consecutive time windows, and the system identified the corresponding time interval as a quality state collapse interval. Subsequently, the risk control service module and the order service module experienced interface call blocking, and the corresponding propagation correlation strength in the quality propagation correlation matrix increased synchronously. The system ultimately generated quality state collapse propagation paths from the payment service module to the risk control service module and from the payment service module to the order service module, and identified the payment service module as the defect source node.

[0188] To verify the detection effect of the present invention, the present invention was compared with traditional log rule detection methods, LSTM anomaly detection methods based on reconstruction error, and the original OmniAnomaly model. The experimental results are shown in Table 1.

[0189] Table 1 Comparison of Software Development Defect Detection Results

[0190] Log rule detection method 71.4 58.6 19.7 148 21.3 32.5 LSTM anomaly detection method 81.9 69.8 13.2 97 43.7 57.1 OmniAnomaly model 87.6 78.4 9.5 74 61.2 70.6 Method of the present invention 95.8 92.7 4.1 29 89.4 93.5

[0191] As shown in Table 1, the method of this invention exhibits high stability and accuracy across all detection metrics. Compared to log rule-based detection methods, this invention improves defect identification accuracy from 71.4% to 95.8%, demonstrating its ability to effectively identify hidden and persistent defects in complex software systems. Defect source localization accuracy increases from 58.6% to 92.7%, indicating that this invention, through its anti-entropy recovery perturbation mechanism and collapse resolution contribution analysis, can accurately locate defect propagation source nodes. The average false alarm rate decreases from 19.7% to 4.1%, demonstrating that this invention establishes defect judgment relationships based on the quality state entropy evolution process, reducing false triggering issues in traditional anomaly detection. The average detection latency is reduced to 29 seconds, indicating that this invention can quickly identify anomaly propagation trends in the early stages of quality state collapse. Furthermore, the propagation path identification rate (%) and continuous hidden defect identification rate reach 89.4% and 93.5%, respectively, demonstrating that this invention can accurately describe defect propagation relationships in complex microservice architectures and improve the detection capability for continuous hidden defects.

[0192] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A software development defect detection method based on the fusion of multi-dimensional data quality indicators, characterized in that, Includes the following steps: Step 1: Collect development process data during the software development process, perform time alignment processing, and generate the original quality dataset; Step 2: Based on the original quality dataset, extract data quality indicators and generate a multidimensional data quality indicator matrix; Step 3: Based on the multidimensional data quality index matrix, extract the time series of data quality changes, and generate a quality propagation correlation matrix according to the time lag relationship; Step 4: Generate a quality time series matrix based on the multidimensional data quality index matrix, and input the quality time series matrix into the improved OmniAnomaly model with hidden state entropy evolution layer and anti-entropy recovery perturbation layer to generate a quality state entropy sequence; Step 5: Based on the mass state entropy sequence, calculate the mass state entropy change results between consecutive time windows to generate a mass state collapse sequence; Step 6: Based on the quality state collapse sequence, perform anti-entropy recovery perturbation processing on the target data quality indicators to generate defect source location results; Step 7: Based on the defect source location results and the quality propagation correlation matrix, generate the quality state collapse propagation path and output the software development defect detection results.

2. The software development defect detection method based on multi-dimensional data quality index fusion according to claim 1, characterized in that, Step one specifically involves: Collect code commit records, API call records, log records, database transaction records, test execution records, and build execution records throughout the software development lifecycle, and classify and store them according to module identifier, API identifier, service instance identifier, and database transaction identifier to generate a multi-source development process dataset; Based on code submission time, API call time, log generation time, database transaction time, test execution time, and build execution time, extract the corresponding timestamp information and generate a continuous time segment sequence according to the timestamp interval; Perform time interpolation on missing timestamps in a continuous time segment sequence and generate a time-compensated sequence based on the time interval between adjacent timestamps; Based on the time compensation sequence, the time offset between different data sources is calculated, and the time axis remapping processing is performed on code commit records, interface call records, log records, database transaction records, test execution records and build execution records according to the time offset to generate a unified time base sequence; Dynamic window partitioning is performed based on the data change density in a unified time base sequence. The code commit records, API call records, log records, database transaction records, test execution records, and build execution records are then windowed according to the dynamic window span to generate windowed development process data blocks. Based on the time segment number, data source number, and development object identifier in the windowed development process data block, time association and data source association are established to generate the original quality dataset.

3. The software development defect detection method based on multi-dimensional data quality index fusion according to claim 1, characterized in that, Step two specifically involves: Perform field splitting and field type mapping processing on the code commit records, API call records, log records, database transaction records, test execution records, and build execution records in the original quality dataset to generate corresponding data attribute sequences; The data integrity index is calculated based on the number of missing fields and the total number of fields in the data attribute sequence, and the data consistency index is calculated based on the number of times the field values ​​in the same field position are inconsistent. The interface response fluctuation index is calculated based on the change in response time in the interface call record, and the test coverage fluctuation index is calculated based on the change in test coverage in the test execution record. The log anomaly density index is calculated based on the number of abnormal logs in the log records and the total number of logs, and the data distribution drift index is calculated based on the change in the distribution of transaction attributes in the database transaction records. The version commit disturbance metric is calculated based on the change in the number of commits in the code commit record, and the module call stability metric is calculated based on the change in the call path in the interface call record. Based on the data integrity index, data consistency index, interface response fluctuation index, test coverage fluctuation index, log anomaly density index, data distribution drift index, version commit disturbance index, and module call stability index, the index fluctuation entropy corresponding to these indices is normalized to generate a standardized quality index sequence. The standardized quality indicator sequence is matrix arranged according to the time segment number and the development object identifier to generate a multidimensional data quality indicator matrix.

4. The software development defect detection method based on multi-dimensional data quality index fusion according to claim 1, characterized in that, Step three specifically involves: The data quality indicators in the multidimensional data quality indicator matrix are arranged in a continuous time alignment according to the time segment number to generate the corresponding data quality change time series. Based on adjacent time segments in the data quality change time series, calculate the difference of data quality indicators and the magnitude of indicator changes, and generate a sequence of indicator change trajectories in chronological order; Multi-scale time offset traversal processing is performed on different data quality indicators in the indicator change trajectory sequence to generate indicator correlation change sequences under different time offset conditions; The time lag correlation value is calculated based on the number of synchronous changes in the same direction and the number of synchronous changes in the magnitude of changes in the index correlation change sequence, and a time lag correlation sequence is generated according to the time offset. Perform reverse time filtering on the time lag correlation values ​​in the time lag correlation sequence, and generate quality propagation correlation edges according to the time sequence; A sparse matrix mapping process is performed based on the starting development object identifier, the target development object identifier, and the time lag association value in the quality propagation association edges to generate a quality propagation association matrix.

5. The software development defect detection method based on multi-dimensional data quality index fusion according to claim 1, characterized in that, Step four specifically involves: Input the quality time series matrix into the improved OmniAnomaly model; The improved OmniAnomaly model includes a temporal recursive encoding module, a hidden state entropy evolution module, an entropy feedback recursive module, and an anti-entropy perturbation recovery module. The temporal recursive coding module takes the quality temporal matrix as input and performs temporal recursive coding processing on the data quality indicators in continuous time segments to generate the hidden state vector sequence in the corresponding time segments. The temporal recursive coding module includes a gated recursive unit and a hidden state distribution mapping unit; The gated recursive unit reads the data quality indicators in the quality time series matrix in chronological order and updates the hidden state vector in the current time segment based on the hidden state vector in the previous time segment. The hidden state distribution mapping unit performs probability distribution mapping processing on the hidden state vector to generate the hidden state distribution sequence in the corresponding time segment; The hidden state entropy evolution module takes the hidden state distribution sequence as input, performs probability statistical processing on the hidden state distribution in each time segment, and generates a hidden state probability sequence. Calculate the degree of dispersion of the state distribution in the corresponding time segment based on the probability change magnitude in the hidden state probability sequence; The state distribution discreteness is continuously evolved and arranged according to time order to generate a mass state entropy sequence; Calculate the entropy change difference between adjacent time segments in the mass state entropy sequence to generate a mass state entropy change sequence; Perform collapse judgment processing on the continuous unidirectional entropy gradient interval in the mass state entropy change sequence to generate the mass state collapse interval. The entropy feedback recursive module performs hidden state entropy coupling shaping processing based on the mass state entropy value in the previous time segment to generate the hidden state distribution sequence in the corresponding time segment. The anti-entropy recovery perturbation module performs anti-entropy collapse tracing processing based on the data quality indicators corresponding to the quality state collapse interval, and generates an entropy recovery change sequence.

6. The software development defect detection method based on multi-dimensional data quality index fusion according to claim 5, characterized in that, The hidden state entropy coupling shaping process specifically involves: Extract the mass state entropy value and the corresponding hidden state vector from the previous time segment; The entropy coupling strength is calculated based on the correspondence between the mass state entropy value and the hidden state vector, and a sequence of entropy coupling coefficients is generated. Map the entropy coupling coefficient sequence to the hidden state update path corresponding to the current time segment to generate the entropy coupling update path; Based on the entropy coupling update path, a joint shaping process is performed on the hidden state propagation path and the hidden state retention path in the current time segment to generate a shaped hidden state sequence. Based on the shaped hidden state sequence, the hidden state vector in the current time segment is reconstructed to generate an entropy-coupled hidden state vector. The entropy-coupled hidden state vector is input into the hidden state distribution mapping unit to generate the hidden state distribution sequence in the corresponding time segment.

7. The software development defect detection method based on multi-dimensional data quality index fusion according to claim 5, characterized in that, The anti-entropy collapse tracing process specifically involves: Extract the target data quality indicators corresponding to the quality state collapse interval; Perform a gradual perturbation on the target data quality indicators while keeping the time series corresponding to the remaining data quality indicators unchanged; The quality indicators of the target data after asymptotic perturbation are re-input into the time-series recursive encoding module to generate a perturbation hidden state distribution sequence. The degree of dispersion of the state distribution in the corresponding time segment is calculated based on the perturbation hidden state distribution sequence to generate a perturbation state discrete sequence; Perform difference calculation between the discrete sequence of the disturbed state and the discrete sequence of the original state to generate an entropy recovery change sequence; Based on the continuous change relationship in the entropy recovery change sequence, collapse source tracing analysis is performed to generate a collapse release contribution sequence.

8. The software development defect detection method based on multi-dimensional data quality index fusion according to claim 1, characterized in that, Step five specifically involves: Perform continuous time alignment processing on each time segment in the mass state entropy sequence to generate a continuous mass state entropy trajectory sequence; Based on adjacent time segments in a continuous mass state entropy trajectory sequence, calculate the mass state entropy difference and entropy change amplitude to generate a mass state entropy change sequence. Perform entropy gradient direction detection processing on continuous time segments in the mass state entropy change sequence to generate an entropy gradient change sequence; The cumulative change of entropy gradient is calculated based on the continuous unidirectional entropy gradient intervals in the entropy gradient change sequence, and an entropy gradient accumulation sequence is generated. Perform state discreteness detection processing on the mass state entropy values ​​corresponding to the continuous decreasing intervals in the entropy gradient accumulation sequence to generate the mass state convergence interval; Based on the time continuity length, entropy change amplitude, and state discreteness change in the convergence interval of the quality state, a collapse interval mapping process is performed to generate a quality state collapse sequence.

9. The software development defect detection method based on multi-dimensional data quality index fusion according to claim 1, characterized in that, Step six specifically involves: Perform collapse contribution correlation analysis on each collapse interval in the quality state collapse sequence to generate a set of target data quality indicators; A hierarchical perturbation partitioning process is performed on each target data quality indicator in the target data quality indicator set to generate the corresponding asymptotic perturbation sequence; Based on the progressive perturbation sequence, incremental perturbation update processing is performed on the index values ​​corresponding to the target data quality index in chronological order, while keeping the time series corresponding to the non-target data quality index unchanged. The perturbation-updated target data quality index is re-input into the improved OmniAnomaly model, and the perturbation hidden state distribution sequence and perturbation quality state entropy sequence in the corresponding time segment are regenerated. The entropy recovery change is calculated based on the difference in state discreteness between the perturbed mass state entropy sequence and the original mass state entropy sequence, and an entropy recovery trajectory sequence is generated. Perform collapse release contribution calculation on continuous recovery intervals in the entropy recovery trajectory sequence to generate an entropy recovery contribution sequence corresponding to the target data quality index; Based on the ranking of recovery contribution in the entropy recovery contribution sequence and the correlation of collapse propagation time corresponding to the collapse interval of quality state, the defect source location result is generated.

10. The software development defect detection method based on multi-dimensional data quality index fusion according to claim 1, characterized in that, Step seven specifically involves: Perform collapse propagation mapping processing on the target development object identifier, target data quality index and quality state collapse interval in the defect source location results to generate a defect propagation node sequence; Based on the time lag correlation value and propagation direction constraint relationship in the quality propagation correlation matrix, time overlap propagation correlation matching processing is performed on adjacent nodes in the defect propagation node sequence to generate a node propagation correlation sequence. Perform collapse propagation continuity detection processing on each propagation node in the node propagation association sequence to generate a continuous propagation interval sequence; Based on the propagation time length, number of propagation nodes, and propagation association strength in the continuous propagation interval sequence, propagation path connection processing is performed in the order of propagation direction to generate a quality state collapse propagation path sequence. Perform a propagation direction consistency check on each propagation path in the mass state collapse propagation path sequence to generate a propagation direction sequence; Based on the number of propagation direction changes and the change in propagation association strength in the propagation direction sequence, a collapse propagation risk level classification process is performed to generate a path risk level sequence. Based on the path risk level sequence, the quality state collapse propagation path sequence, and the defect source location results, corresponding defect nodes, propagation paths, and software development defect detection results corresponding to risk levels are generated.

Images