A financial service system gray release and intelligent verification method

Abstract

The application discloses a financial service system gray release and intelligent verification method, which forms a gray execution structure by constructing the running state of a target version and a benchmark version in parallel in a production environment; continuously observes the business request processing process, collects system running behavior and business result data, and constructs a phased system behavior observation sequence in different release stages; generates modeling input of behavior distribution change based on the data difference of adjacent release stages, forms statistical evidence reflecting the change degree of the system, and performs time series analysis on the statistical evidence to locate the abnormal change stage triggered by version change; and maps the detection result to a release control instruction to realize automatic decision and execution of release promotion, release freezing or version rollback. The application realizes continuous verification and risk control of the gray release process of the financial service system by associating system behavior change with the release control process.

Description

Technical Field

[0001] This invention relates to the fields of financial information technology and software engineering technology, and in particular to a method for gray-scale release and intelligent verification of financial business systems. Background Technology

[0002] With the large-scale evolution of financial business systems and the widespread adoption of continuous delivery models, canary releases have become an important implementation method for version upgrades and feature iterations of core financial systems. Currently, financial institutions typically reduce the impact of version releases on the continuity and stability of core businesses by introducing new and old versions to run in parallel in the production environment and gradually adjusting the proportion of business traffic. During the release process, they also monitor and verify the system's operational status and business results.

[0003] Existing technologies for canary releases and verification in financial systems still have significant shortcomings. On the one hand, traditional canary releases mainly rely on static rules or manual experience to set the release pace, and system behavior monitoring is mostly based on single indicators or simple threshold judgments. This makes it difficult to characterize the changes in the overall system behavior distribution between different release phases, and it is easy to miss hidden anomalies caused by version switching. On the other hand, existing verification methods usually separate anomaly detection from the release control process, lacking causal relationship analysis between changes in system behavior and release phases. This makes it impossible to accurately locate the specific stage at which anomalies are triggered, and it is also difficult to provide a verifiable basis for decision-making regarding continued release, freezing, or rollback. This results in delayed risk identification and insufficiently refined control response, making it difficult to meet the requirements of financial business systems for high stability and controllable releases.

[0004] Therefore, how to provide a method for canary release and intelligent verification of financial business systems 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 gray-scale release and intelligent verification method for financial business systems. By constructing a gray-scale execution structure in the production environment where new and old versions run in parallel, the system's operational behavior and business results are continuously collected, modeled, and analyzed at different release stages. By utilizing a system behavior distribution change detection and mutation location mechanism, the method can automatically identify version change risks and directly map the detection results to release control decisions, thus completing closed-loop control for release promotion, release freeze, or version rollback.

[0006] A method for canary release and intelligent verification of a financial business system according to an embodiment of the present invention includes the following steps: In the production environment, construct parallel-running grayscale execution states based on the target version and the baseline version, and generate grayscale execution structures; Within each ramp-up phase defined by the gray-scale execution structure, the processing of business requests in the gray-scale execution state is continuously observed, service operation behavior characteristics and corresponding business result data are collected, forming a multi-dimensional system behavior observation sequence that corresponds one-to-one with the ramp-up phase; Perform rank transformation and staged time slicing on the observation sequence of multidimensional system behavior to construct the input sequence for distribution modeling; Using the distribution modeling input sequence as the sole input, a RuLSIF operation based on relative density ratio is performed to generate a distribution shift score, which is then solidified into a statistical evidence unit bound to the volume expansion phase. Using time series composed of statistical evidence units as the detection object, perform the Pettitt single mutation point test to generate mutation triggering identifiers for the corresponding volume expansion stage; The mutation trigger flag is used as a necessary trigger condition for the transition of the gray-scale execution state. Based on the mutation trigger flag, release control instructions are output to continue scaling, freeze scaling, or rollback the version, and drive the gray-scale execution structure to undergo corresponding state changes.

[0007] Optionally, generating a grayscale execution structure includes: Create separate sets of isolated running instances for the target version and the baseline version in the same production environment, and bind a unique version identifier to each set of running instances; An instance mapping table is generated based on the set of running instances, and the instance mapping table is maintained as the instance mapping part in the gray-scale execution structure. Generate a set of volume increase stages based on the preset volume increase stage sequence, generate a corresponding stage boundary record for each volume increase stage, and write the stage boundary record into the grayscale execution structure. When starting in grayscale execution mode, the traffic weight field in the instance mapping table is initialized to the weight distribution state dominated by the baseline version, and the initial state is used as the starting point of the stage boundary of the first ramp-up phase. During the rollout phase, the traffic weight field in the instance mapping table is updated in stages based on the stage boundary records. The target version is in the rollout phase. The traffic share meets the following requirements: ; in, Indicates the target version is in stage The corresponding traffic weight value, Indicates the baseline version at stage The corresponding traffic weight value; After each traffic weight update, the instance mapping relationship table and the stage boundary record are consistent and solidified to generate a snapshot of the gray-scale execution structure for the corresponding volume ramp-up stage.

[0008] Optionally, creating a collection of mutually isolated running instances includes: In the same production environment, perform independent deployment operations for the target version and the baseline version. Start a preset number of service running instances for each version and write version identifiers and instance identifiers to the instances. Register the target version instances and the baseline version instances as two separate sets of instance lists. Configure the traffic scheduling component to bind requests based only on the instance list and version identifier.

[0009] Optionally, forming a multidimensional system behavior observation sequence includes: Read the scaling phase boundary records and instance mapping relationship table from the grayscale execution structure snapshot, determine the time interval of the current scaling phase based on the scaling phase boundary records, and determine the target version instance set and the base version instance set corresponding to the current scaling phase based on the instance mapping relationship table; During the current ramp-up phase, a request identifier is generated for each business request processed by the target version instance set and the baseline version instance set, and service operation behavior characteristics associated with the request identifier are collected during the business request processing. Within the current timeframe of the ramp-up phase, based on the request identifier or the business flow identifier associated with the request identifier, extract the business result data corresponding to the business request from the business processing results, and associate the business result data with the service operation behavior characteristics according to the request identifier; The data associated with the completed request identifier is time-aligned according to a preset time granularity. Data records falling within the same time granularity range are merged into the same time index, and multi-dimensional observation vectors are formed by combining them according to the feature dimensions under the time index. Arrange the multidimensional observation vectors according to the time index to construct the stage observation matrix corresponding to the current volume increase stage. ,in This refers to the sequence number of the volume increase phase; The stage observation matrices corresponding to each volume surge stage are arranged into a multi-dimensional system behavior observation sequence according to the volume surge stage number. .

[0010] Optionally, the input sequences for generating distribution modeling include: Based on the volume increase stage number, the stage observation matrix corresponding to the current volume increase stage and the stage observation matrix corresponding to the next volume increase stage are selected sequentially from the multi-dimensional system behavior observation sequence, and the two stage observation matrices are determined as a set of adjacent volume increase stage comparison data pairs. For each pair of adjacent volume increase phase comparison data, the observation values ​​in the observation matrices of the two phases are subjected to rank transformation according to the feature dimension, and the observation values ​​are mapped to rank values ​​that reflect the relative order relationship within the phase, thus forming rank-transformed observation data. In the stage observation data after rank transformation, the time segmentation within the stage is performed on adjacent volume expansion stages according to the preset time granularity, generating a set of staged time slices corresponding to each volume expansion stage. Based on the order of time slices in each volume expansion stage, a slice pairing relationship is established for the staged time slice sets of adjacent volume expansion stages. Time slices with the same slice number position form a set of stage comparison slice pairs. For each pair of stage comparison slices, multidimensional observation data within the slices are extracted, and a slice sample representation for distribution comparison is constructed while maintaining the consistency of the feature dimension order. At the same time, the slice sample representation is associated with the corresponding volume expansion stage number, slice number and time interval identifier. The slice sample representations generated in each adjacent volume expansion stage are aggregated in the order of volume expansion stage number and slice number to form the input sequence for distribution modeling.

[0011] Optionally, the rank transformation of the observations includes: in the stage observation matrix corresponding to each volume expansion stage, extracting all observations under each feature dimension, sorting the observations according to their numerical values, and replacing the corresponding observations with their sorted position numbers to form rank values. When there are observations with the same numerical value, assigning the same rank value to the group of observations according to a preset parallel rank rule or determining a unified rank value based on their sorting position, thereby converting the stage observation matrix into a rank-transformed observation matrix composed of rank values ​​and using it as input for subsequent staged time slices and distribution modeling.

[0012] Optionally, generating statistical evidence units includes: According to the volume increase stage number and slice number, the stage comparison slice pairs are read sequentially from the distribution modeling input sequence; Align the phase comparison slices with the volume surge phase. The sliced ​​sample representation is constructed as a reference sample matrix; Align the phase comparison slices with the volume surge phase. The sliced ​​samples are represented as the target sample matrix, and the dimensions of the reference sample matrix and the target sample matrix are verified. Handling identical values, consistent column order, and missing values; Configure a quantile index set using the reference sample matrix as the sole source. The set of quantile indexes remains fixed within the same publication control link; For each feature dimension of the reference sample matrix Calculate the quantile values ​​corresponding to the quantile index set, and construct the anchor kernel center vector one by one according to the quantile index. Anchoring kernel center vector The Dimensional reference sample matrix Dimension in quantile index The quantile values ​​below form the anchored core center set. ; The anchored core center set is used to generate center records according to fixed serialization rules, and the center record check value is calculated and written into the statistical evidence unit. Using the set of anchor kernel centers as the kernel basis function center, Gaussian kernel basis function is used to calculate the kernel similarity from the sample to the anchor kernel center. Based on the set of anchor kernel centers, the set of Euclidean distances between each pair of anchor kernel centers is calculated, and the kernel width is determined by the median of the set of Euclidean distances. If the median is 0, the kernel width is determined according to the preset lower limit rule, and the kernel width value rule identifier and kernel width are written into the statistical evidence unit. Based on the kernel width, construct reference design matrices for the reference sample matrix and the target sample matrix respectively. With the target design matrix Each row corresponds to a sample, and each column corresponds to an anchor kernel center; Weighting coefficients The regularization coefficient is fixed to the preset configuration value. Fixed by the number of samples and The numerical values ​​are determined by the explicit function, and a linear term vector is constructed. With quadratic term matrix The first-order term vector is obtained by averaging the target design matrix over the samples, and the second-order term matrix is ​​obtained by weighting the second-order term estimates of the target design matrix and the reference design matrix. Obtained by weighted synthesis; Define the relative density ratio function as: ;; in, This represents the probability density corresponding to the reference sample matrix. This represents the probability density corresponding to the target sample matrix. This indicates the preset weighting coefficient; In regularization coefficient Solving the least squares closed-form solution under constraints to obtain the parameter vector The expression for solving its parameters is: ; in, It is a quadratic term matrix. For a linear term vector, The identity matrix is ​​used, and the parameter solution status identifier is written into the statistical evidence unit; For each sample in the target sample matrix, extract its corresponding kernel similarity vector. And calculate the relative density ratio estimate. If the relative density ratio estimate is less than 0, it is cut to 0 according to the fixed non-negative cutting rule and the cutting ratio is calculated. The distribution offset score is generated using the squared mean of the deviation from the baseline value by 1. And generate a summary statistical field of the relative density ratio estimate and write it into the statistical evidence unit; A list of feature groups is generated based on the feature dimension identifiers using deterministic grouping rules. And the set of dimension indices corresponding to each feature group, and write the grouping rule identifier and the set of dimension indices into the statistical evidence unit; for each feature group Calculate the median rank for each dimension of the feature group in the reference sample matrix and concatenate them to form the feature group masking vector. ; Without reconstructing the set of anchor kernel centers, redetermining the kernel width, or resolving the parameter vectors. Under the premise of each feature group Construct the target sample matrix after masking The target sample matrix that belongs to the feature group Replace the dimension value with the corresponding feature group mask vector. The remaining dimensions remain unchanged, and the same set of anchored kernel centers, kernel width, and parameter vectors are reused to calculate the shielding offset score. ; The difference between the distribution offset score and the shielding offset score is used as the feature group contribution value. The feature group contribution value and the corresponding feature group identifier are written into the contribution record; Include the scaling phase number, slice number, time interval identifier, slice pair identifier, number of reference samples, number of target samples, feature dimension verification value, and weight coefficient. Kernel width value rules identifier and kernel width, regularization coefficient Value selection rules and values, anchor kernel center set verification values, parameter vectors Checksum, parameter solution status indicator, distribution offset score The relative density ratio estimate summary statistics, pruning ratio and contribution record are written into the statistical evidence unit record to form a statistical evidence unit bound to the volume expansion stage; The records of each statistical evidence unit are aggregated according to the sequence number of the volume expansion stage and the slice number to generate a statistical evidence unit sequence and output it.

[0013] Optionally, generating mutation trigger identifiers includes: The statistical evidence unit sequence is sorted according to the chronological order of its time interval identifiers, resulting in a sequence of length [length missing]. An ordered sequence of evidence; The distribution offset scores are extracted one by one from the ordered evidence sequence to form an offset score sequence, and a location index table is generated simultaneously. Construct an evidence integrity vector for an ordered sequence of evidence, and for each statistical evidence unit, determine whether the following fields exist in its fixed fields: The anchored core center record check value, parameter vector check value, contribution record, pruning ratio, and relative density ratio estimate are summarized and statistically analyzed, and the number of existing fields is used as the integrity count of the statistical evidence unit. When there are scores with the same value in the offset score sequence, the scores are first grouped according to their size, and then sorted a second time within each group according to the integrity count from largest to smallest. If the integrity counts are still the same, the scores are sorted a third time according to the lexicographical order of the release stage number and the slice number in the position index table to generate a stable order of rank distribution. Rank assignment is performed on the offset score sequence based on the stable order to obtain the evidence order sequence. Each of them For position The corresponding score ranks under a stable order are used to determine the evidence order column as the unique input sequence for the Pettitt single mutation point test. Perform Pettitt rank statistic calculation on the evidence order sequence for each candidate time position. Divide the evidence sequence into a front set. With the latter part of the set And calculate the rank statistic of the candidate time positions. Its calculation expression is: ; in, The sign function is used; the rank statistic is calculated sequentially for all candidate time positions to form a rank statistic sequence; Calculate the corresponding time position for each candidate time position in the rank statistic sequence. , determine The candidate time position with the maximum value is taken as the initial candidate mutation position. Its defining expression is ; And the initial mutation candidate positions, corresponding Write candidate records in association with the location index table; Based on candidate records, perform quantitative structural consistency checks on mutation candidate points and read the initial mutation candidate positions using the position index table. The volume ramp-up stage number and slice number are obtained, and the volume ramp-up stage number and slice number corresponding to at least one preceding position and at least one following position adjacent to the initial mutation candidate position are read. It is then determined whether the initial mutation candidate position satisfies the following conditions: The initial mutation candidate position corresponds to the volume release stage number and at least one of the volume release stage numbers in the adjacent positions have a stage boundary change relationship, and the initial mutation candidate position and at least one of the slice numbers in the adjacent positions satisfy the same slice pair relationship. When the consistency check passes, the initial mutation candidate position is determined as the single mutation point position, and the corresponding scale-up stage number, slice number, time interval identifier and statistical evidence unit identifier are obtained by reverse lookup through the position index table, and the mutation location result is generated; when the consistency check fails, a no-mutation result is generated and the current Pettitt single mutation point test process is terminated. Based on the mutation location results, a mutation trigger identifier is generated and written into the mutation trigger table; the mutation trigger table is output as the trigger input for the grayscale execution state transition.

[0014] The beneficial effects of this invention are: RuLSIF computation models the relative density ratio of system behavior data in adjacent release phases, directly quantifying the intensity of the difference in behavior distribution between the two phases, forming statistical evidence bound to the release phase. This makes gray-scale release verification no longer dependent on single indicator thresholds or empirical rules, and can identify latent shifts in the joint space of multi-dimensional behavioral features and business result data, improving the sensitivity and stability of anomaly identification.

[0015] The Pettitt single mutation point test uses the time series of statistical evidence units as the detection object. It can automatically locate the time position of the distribution shift triggered by the version change in a continuous volume expansion phase, and reverse look up the corresponding volume expansion phase and time interval identifier. This solves the problem of "detecting anomalies but finding it difficult to accurately locate the triggering phase" in the existing technology, and improves the stage accuracy and operability of anomaly location. Attached Figure Description

[0016] 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: Figure 1 This is a flowchart of a method for gray-scale release and intelligent verification of a financial business system proposed in this invention. Figure 2 This is a schematic diagram of the RuLSIF operation process proposed in this invention. Detailed Implementation

[0017] 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.

[0018] refer to Figure 1 - Figure 2A method for canary release and intelligent verification of a financial business system includes the following steps: In the production environment, a parallel canary execution state is built based on the target version and the baseline version, and a canary execution structure containing the scale-up phase boundary and instance mapping relationship is generated. Within each ramp-up phase defined by the gray-scale execution structure, the processing of business requests in the gray-scale execution state is continuously observed, service operation behavior characteristics and corresponding business result data are collected, and a multi-dimensional system behavior observation sequence corresponding to each ramp-up phase is formed. We perform rank transformation and staged time slicing on the multidimensional system behavior observation sequence to construct a distribution modeling input sequence that characterizes the differences in the distribution of system behavior between adjacent volume expansion stages; Using the distribution modeling input sequence as the sole input, a RuLSIF operation based on relative density ratio is performed to generate a distribution shift score that reflects the intensity of changes in the distribution of system behavior. The distribution shift score is then solidified into a statistical evidence unit that is bound to the volume expansion phase. Using time series composed of statistical evidence units as the detection object, perform Pettitt single mutation point test to determine the time position of the system behavior distribution triggered by version change and generate mutation trigger identifiers for the corresponding volume expansion stage; The mutation trigger flag is used as a necessary trigger condition for the transition of the gray-scale execution state. Based on the mutation trigger flag, release control instructions are output to continue scaling, freeze scaling, or rollback the version, and drive the gray-scale execution structure to undergo corresponding state changes.

[0019] In this embodiment, generating the grayscale execution structure includes: In the same production environment, create separate sets of isolated running instances for the target version and the baseline version, and bind a unique version identifier to each set of running instances, so that the target version and the baseline version can maintain an independent execution state for processing business requests under the condition of shared infrastructure; An instance mapping table is generated based on the set of running instances. The instance mapping table contains at least the instance identifier, version identifier and current traffic weight fields, and is maintained as the instance mapping part in the gray-scale execution structure. A set of volume expansion stages is generated based on a preset volume expansion stage sequence. A corresponding stage boundary record is generated for each volume expansion stage. The stage boundary record contains at least the stage number, the stage start identifier, and the stage end judgment condition. The stage boundary record is then written into the grayscale execution structure. When starting in grayscale execution mode, the traffic weight field in the instance mapping table is initialized to the weight distribution state dominated by the baseline version, and this initial state is used as the starting point of the stage boundary of the first ramp-up phase. During the rollout phase, the traffic weight field in the instance mapping table is updated in stages based on the stage boundary records, ensuring the target version is optimized during the rollout phase. The traffic share meets the following requirements: ; in, Indicates the target version is in stage The corresponding traffic weight value, Indicates the baseline version at stage The corresponding traffic weight value; After each traffic weight update, the instance mapping relationship table and the stage boundary record are consistent and solidified to generate a snapshot of the gray-scale execution structure for the corresponding volume release stage. The snapshot of the gray-scale execution structure is then used as the structural input for subsequent system behavior observation, distribution modeling, and release control decisions.

[0020] In this implementation, creating a set of mutually isolated running instances includes: In the same production environment, the target version and the baseline version are deployed independently. A preset number of service instances are started for each version, and version identifiers and instance identifiers are written to the instances. The target version instances and the baseline version instances are registered as two separate instances. The traffic scheduling component is configured to bind requests based only on the instance list and version identifier, ensuring that any business request enters only the target version instance group or only the baseline version instance group in a single processing link, and that different version instances do not reuse the same running process.

[0021] In this embodiment, generating a multidimensional system behavior observation sequence includes: Read the scaling phase boundary records and instance mapping relationship table from the grayscale execution structure snapshot, determine the time interval of the current scaling phase based on the scaling phase boundary records, and determine the target version instance set and the base version instance set corresponding to the current scaling phase based on the instance mapping relationship table; During the current ramp-up phase, a request identifier is generated for each business request processed by the target version instance set and the baseline version instance set. Service operation behavior characteristics associated with the request identifier are collected during the business request processing. The service operation behavior characteristics include at least the request processing latency and the processing result status code. Within the current timeframe of the ramp-up phase, based on the request identifier or the business flow identifier associated with the request identifier, extract the business result data corresponding to the business request from the business processing results. The business result data includes at least the business processing result identifier and the core business value field, and associate the business result data with the service operation behavior characteristics according to the request identifier. The data associated with the completed request identifier is time-aligned according to a preset time granularity. Data records falling within the same time granularity range are merged into the same time index, and multi-dimensional observation vectors are formed by combining them according to feature dimensions under this time index. Arrange the multidimensional observation vectors according to the time index to construct the stage observation matrix corresponding to the current volume increase stage. ,in This refers to the sequence number of the volume increase phase; The stage observation matrices corresponding to each volume surge stage are arranged into a multi-dimensional system behavior observation sequence according to the volume surge stage number. .

[0022] In this embodiment, constructing the input sequence for distribution modeling includes: Based on the volume increase stage number, select the stage observation matrix corresponding to the current volume increase stage and the stage observation matrix corresponding to the next volume increase stage from the multi-dimensional system behavior observation sequence, and determine the two stage observation matrices as a set of adjacent volume increase stage comparison data pairs. For each pair of adjacent volume increase phase comparison data, the observation values ​​in the observation matrices of the two phases are subjected to rank transformation according to the feature dimension, and the observation values ​​are mapped to rank values ​​that reflect the relative order relationship within the phase, forming rank-transformed observation data for distribution comparison between phases. In the stage observation data after rank transformation, the time segmentation within the stage is performed on adjacent volume expansion stages according to the preset time granularity, generating a set of staged time slices corresponding to each volume expansion stage. Based on the order of time slices in each volume expansion stage, a slice pairing relationship is established for the staged time slice sets of adjacent volume expansion stages, so that time slices with the same slice number position form a set of stage comparison slice pairs. For each pair of stage comparison slices, multidimensional observation data within the slices are extracted, and a slice sample representation for distribution comparison is constructed while maintaining the consistency of the feature dimension order. At the same time, the slice sample representation is associated with the corresponding volume increase stage number, slice number and time interval identifier. The slice sample representations generated in each adjacent volume expansion stage are aggregated in the order of volume expansion stage number and slice number to form a distribution modeling input sequence for characterizing the differences in system behavior distribution between adjacent volume expansion stages.

[0023] In this embodiment, the statistical evidence generation unit includes: Read the phase comparison slice pairs sequentially from the distribution modeling input sequence according to the volume increase phase number and slice number, and then select the slice pairs from the volume increase phase. The sliced ​​sample representation is constructed as a reference sample matrix; Align the phase comparison slices with the volume surge phase. The sliced ​​sample representation is constructed as a target sample matrix, where each sample for Dimensional vectors, and verify the dimension of the reference sample matrix and the target sample matrix. Same, consistent column order, and missing values; Configure a quantile index set using the reference sample matrix as the sole source. The set of quantile indexes remains fixed within the same publication control link; For each feature dimension of the reference sample matrix Calculate the quantile values ​​corresponding to the quantile index set, and construct the anchor kernel center vector one by one according to the quantile index. , so that the anchoring core center vector The Dimensional reference sample matrix Dimension in quantile index The quantile values ​​below form the anchored core center set. ; The anchored core center set generates a center record according to a fixed serialization rule and calculates the center record check value, which is then written into the statistical evidence unit so that different slices within the same release control link can share a verifiable core center definition. Using the set of anchor kernel centers as the kernel basis function center, Gaussian kernel basis function is used to calculate the kernel similarity from the sample to the anchor kernel center. Based on the set of anchor kernel centers, the set of Euclidean distances between each pair of anchor kernel centers is calculated, and the kernel width is determined by the median of the set of Euclidean distances. If the median is 0, the kernel width is determined according to the preset lower limit rule, and the kernel width value rule identifier and kernel width are written into the statistical evidence unit. Based on the kernel width, construct reference design matrices for the reference sample matrix and the target sample matrix respectively. With the target design matrix Each row corresponds to a sample, and each column corresponds to an anchor kernel center; Weighting coefficients The regularization coefficient is fixed to the preset configuration value. Fixed by the number of samples and The numerical values ​​are determined by the explicit function, and a linear term vector is constructed. With quadratic term matrix The first-order term vector is obtained by averaging the target design matrix over the samples, and the second-order term matrix is ​​obtained by weighting the second-order term estimates of the target design matrix and the reference design matrix. Obtained by weighted synthesis; Define the relative density ratio function as: ; in, This represents the probability density corresponding to the reference sample matrix. This represents the probability density corresponding to the target sample matrix. This indicates the preset weighting coefficient; In regularization coefficient Solving the least squares closed-form solution under constraints to obtain the parameter vector Its parameter solution expression is as follows ; in, It is a quadratic term matrix. For a linear term vector, The identity matrix is ​​used, and the parameter solution status identifier is written into the statistical evidence unit; For each sample in the target sample matrix, extract its corresponding kernel similarity vector. And calculate the relative density ratio estimate. If the estimated relative density ratio is less than 0, it is cut to 0 according to the fixed non-negative cut rule and the cut ratio is calculated; the distribution offset score is generated using the square mean of the deviation from the baseline value of 1. And generate a summary statistical field of the relative density ratio estimate and write it into the statistical evidence unit; A list of feature groups is generated based on the feature dimension identifiers using deterministic grouping rules. And the set of dimension indices corresponding to each feature group, and write the grouping rule identifier and the set of dimension indices into the statistical evidence unit; for each feature group Calculate the median rank of each dimension of the feature group in the reference sample matrix and concatenate them to form the feature group masking vector. ; Without reconstructing the set of anchor kernel centers, redetermining the kernel width, or resolving the parameter vectors. Under the premise of each feature group Construct the target sample matrix after masking The target sample matrix that belongs to the feature group Replace the dimension value with the corresponding feature group mask vector. The remaining dimensions remain unchanged, and the same set of anchored kernel centers, kernel width, and parameter vectors are reused to calculate the shielding offset score. The difference between the distribution offset score and the shielding offset score is used as the feature group contribution value. The feature group contribution value and the corresponding feature group identifier are written into the contribution record; Include the scaling phase number, slice number, time interval identifier, slice pair identifier, number of reference samples, number of target samples, feature dimension verification value, and weight coefficient. Kernel width value rules identifier and kernel width, regularization coefficient Value selection rules and values, anchor kernel center set verification values, parameter vectors Checksum, parameter solution status indicator, distribution offset score The relative density ratio estimate summary statistics, pruning ratio and contribution record are written into the statistical evidence unit record to form a statistical evidence unit bound to the volume expansion stage; The records of each statistical evidence unit are aggregated according to the sequence number of the volume expansion stage and the slice number to generate a statistical evidence unit sequence and output it.

[0024] In this embodiment, generating the distribution offset score includes: After estimating the relative density ratio, for each sample in the target sample matrix, the estimated relative density ratio value is read, and the difference between the estimated relative density ratio value and the preset benchmark value 1 is calculated to obtain the deviation value. The deviation value is then squared to obtain the squared deviation. The squared deviation values ​​corresponding to all samples in the target sample matrix are summed and divided by the number of target samples to obtain the mean of the squared deviation values. The mean of the squared deviation values ​​is determined as the distribution offset score corresponding to the stage comparison slice. .

[0025] In this embodiment, constructing the anchoring kernel center vector includes: In the reference sample matrix, each quantile index of the quantile index set is processed one by one. First, all sample values ​​corresponding to each feature dimension are extracted and sorted by numerical value. The sorting position is determined according to the quantile index, and the sample value at the sorting position is read as the quantile value of the feature dimension. Then, according to the feature dimension order of the reference sample matrix, the quantile values ​​obtained from each feature dimension are sequentially filled into the corresponding dimension positions of the same vector, forming a one-to-one correspondence with the quantile index. Anchor kernel center vectors are defined, and the anchor kernel center vectors corresponding to all point indices are collected to form an anchor kernel center set.

[0026] In this embodiment, generating the mutation trigger flag includes: The statistical evidence unit sequence is sorted according to the chronological order of its time interval identifiers, resulting in a sequence of length [length missing]. The ordered evidence sequence; the distribution offset score is extracted from each of the ordered evidence sequences to form an offset score sequence, and a position index table is generated simultaneously. The position index table contains at least a one-to-one correspondence between the sequence position, the volume release stage number, the slice number, the time interval identifier, and the statistical evidence unit identifier. Construct an evidence integrity vector for an ordered sequence of evidence, and for each statistical evidence unit, determine whether the following fields exist in its fixed fields: The anchored core center record check value, parameter vector check value, contribution record, pruning ratio, and relative density ratio estimate are summarized and statistically analyzed, and the number of existing fields is used as the integrity count of the statistical evidence unit. When there are scores with the same value in the offset score sequence, the scores are first grouped according to their size, and then sorted a second time within each group according to the integrity count from largest to smallest. If the integrity counts are still the same, the scores are sorted a third time according to the lexicographical order of the release stage number and the slice number in the position index table to generate a stable order for rank allocation. Rank assignment is performed on the offset score sequence based on the stable order to obtain the evidence order sequence. Each of them For position The corresponding score ranks under a stable order are used to determine the evidence order column as the unique input sequence for the Pettitt single mutation point test. Perform Pettitt rank statistic calculation on the evidence order sequence for each candidate time position. Divide the evidence sequence into a front set. With the latter part of the set And calculate the rank statistic of the candidate time position. Its calculation expression is: ; in, The sign function is used; the rank statistic is calculated sequentially for all candidate time positions to form a rank statistic sequence; Calculate the corresponding time position for each candidate time position in the rank statistic sequence. , determine The candidate time position with the maximum value is taken as the initial candidate mutation position. Its defining expression is ; And the initial mutation candidate positions, corresponding Write candidate records in association with the location index table; Based on candidate records, perform quantitative structural consistency checks on mutation candidate points and read the initial mutation candidate positions using the position index table. The volume expansion stage number and slice number are obtained, and the volume expansion stage number and slice number corresponding to at least one preceding position and at least one following position adjacent to the initial mutation candidate position are read. It is determined whether the initial mutation candidate position meets the following conditions: the volume expansion stage number corresponding to the initial mutation candidate position has a stage boundary change relationship with at least one of the volume expansion stage numbers in the adjacent positions, and the initial mutation candidate position and at least one of the slice numbers in the adjacent positions satisfy the same slice pair relationship. When the consistency check passes, the initial mutation candidate position is determined as the single mutation point position, and the corresponding scale-up stage number, slice number, time interval identifier and statistical evidence unit identifier are obtained by reverse lookup through the position index table, and the mutation location result is generated; when the consistency check fails, a no-mutation result is generated and the current Pettitt single mutation point test process is terminated. Based on the mutation location results, a mutation trigger identifier is generated. The mutation trigger identifier includes at least the volume expansion stage number, slice number, time interval identifier, single mutation point location identifier, statistical evidence unit identifier, and candidate record identifier. The mutation trigger identifier is then written into the mutation trigger table.

[0027] In this embodiment, the driving grayscale execution structure to undergo corresponding state changes includes: Extract the volume surge stage number, slice number, time interval identifier, and statistical evidence unit identifier from the mutation trigger identifier. Based on the statistical evidence unit identifier, read the corresponding statistical evidence unit and its distribution offset score and contribution record from the statistical evidence unit sequence. Based on the volume expansion stage number, the stage boundary record and instance mapping relationship table of the volume expansion stage are read from the gray-scale execution structure snapshot, and the stage end judgment condition in the stage boundary record and the traffic weight field in the instance mapping relationship table are jointly determined as the control input for this state transition; Establish control event records for statistical evidence units corresponding to mutation trigger identifiers. Control event records should include at least the volume ramp-up stage number, slice number, time interval identifier, distribution offset score, contribution record summary and control event timestamp, and write the control event records to the control log area of ​​the grayscale execution structure snapshot. Using the existence of a mutation trigger identifier as the state transition trigger condition, the current control state of the gray-scale execution state is divided into one of the following: continued expansion state, expansion freeze state, and version rollback state. The current expansion stage is then determined to proceed based on the stage end judgment condition in the stage boundary record. When it is determined that progress is allowed and the current control state is to continue to increase volume, a continue to increase volume instruction is generated. The continue to increase volume instruction includes at least the target volume increase stage number and the target traffic weight update identifier. Based on the instance mapping relationship table in the gray-scale execution structure, the traffic weight field is updated in stages so that the traffic proportion of the target version instance set in the next volume increase stage meets the weight allocation rules corresponding to the stage boundary record. At the same time, the updated instance mapping relationship table is fixed and a snapshot of the gray-scale execution structure of the next volume increase stage is generated. When it is determined that advancement is not allowed and the current control state is a volume release freeze state, a volume release freeze instruction is generated. The volume release freeze instruction includes at least the freeze volume release stage sequence number and the freeze start timestamp. The traffic weight field in the instance mapping relationship table is kept as the weight distribution state at the time corresponding to the mutation trigger identifier. At the same time, a freeze snapshot is generated and the freeze snapshot is written to the gray-scale execution structure snapshot set. When the current control state is the version rollback state, a version rollback instruction is generated. The version rollback instruction includes at least the rollback target version identifier, the rollback trigger stage sequence number and the rollback trigger evidence identifier. The target version instance set traffic weight field in the instance mapping relationship table is updated to the zero weight state, the baseline version instance set traffic weight field is updated to the full weight state, and the current stage state of the stage boundary record is marked as rollback complete and a rollback snapshot is generated. The volume increase command, volume increase freeze command, or version rollback command will continue to be output to the traffic scheduling component in the canary execution state. The traffic scheduling component will then perform request route updates based on the instance mapping table, so that the canary execution structure can complete the corresponding state changes under the control command.

[0028] Example: To verify the feasibility and technical effectiveness of the proposed canary release and intelligent verification method for financial business systems in a real production environment, this embodiment selects the core business system of a large financial institution as the application object. This system operates under high-concurrency business scenarios, carrying critical business logic such as account verification, transaction routing, and risk control judgment. The business request types are complex and the access patterns are diverse. Due to frequent adjustments to business rules, the system needs to complete multiple version upgrades without interrupting service. Existing canary release solutions typically monitor only a few indicators such as request success rate, error code ratio, or average response time. When a new version causes changes in the overall behavior distribution in some business paths but has not yet triggered obvious indicator alerts, anomalies are difficult to identify in a timely manner, and it is impossible to accurately determine whether the anomaly was triggered by the version rollout, thus leading to the accumulation of release risks.

[0029] During a version upgrade of this core business system, a canary release and intelligent verification method for financial business systems, as proposed in this invention, was introduced. The system simultaneously deploys the target version and the baseline version in the same production environment, and constructs isolated sets of running instances for each type of version. A unified traffic scheduling mechanism distributes real business requests to different version instances according to a set ratio, enabling the old and new versions to run in parallel under shared infrastructure conditions. The entire release process is divided into multiple progressively advancing phases, each corresponding to a fixed traffic weight distribution state. A canary execution structure is used to solidify and manage the mapping relationship between the phase boundaries and instances.

[0030] During the canary release process, the system continuously monitors business requests entering each version instance, collecting operational behavior characteristics and business result data associated with each request during request processing. Operational behavior characteristics at least cover request processing latency and processing status information, while business result data covers transaction result identifiers and core business fields. Within the defined time interval of each rollout phase, the system aligns and merges the collected data according to a uniform time granularity, forming multi-dimensional system behavior observation data reflecting the overall operational status of that rollout phase. The data corresponding to each rollout phase are then arranged into a system behavior observation sequence according to the phase order.

[0031] After the system behavior observation sequence is constructed, the system performs inter-stage distribution modeling on the data of adjacent volume expansion stages. For each pair of adjacent volume expansion stages, the system performs a rank transformation operation on the multi-dimensional system behavior observation data within the corresponding stage, converting the values ​​of each feature dimension from absolute values ​​to rank values ​​that reflect the relative order relationship within the stage, thereby eliminating the influence of dimensional differences and extreme values. After completing the rank transformation, the system performs intra-stage time segmentation on the data of adjacent volume expansion stages according to a preset time granularity, generating pairs of stage-specific time slices, and constructing slices with the same slice index position as stage comparison slice pairs to characterize the differences in behavior distribution between adjacent stages under the same time structure.

[0032] For each stage of the comparison slice pair, the system uses the data from the corresponding stage of the baseline version as the reference sample set and the data from the corresponding stage of the target version as the target sample set, and performs relative density ratio calculation based on these sample sets. In specific implementation, the system adopts the RuLSIF operation method with relative density ratio as its core. Through fixed kernel center construction rules and kernel width determination strategies, it calculates the estimated relative density ratio of the target sample distribution relative to the reference sample distribution, and generates a distribution shift score reflecting the intensity of distribution change based on this. Each distribution shift score is bound to the corresponding volume increase stage, time slice, and sample size information, and is solidified as a statistical evidence unit for subsequent analysis and control.

[0033] As the grayscale implementation continues, the system continuously generates statistical evidence units (SEM) linked to each phase, forming a sequence of SEM units in chronological order. The system uses this sequence of SEM units as the detection object, performing single mutation point detection processing on the distribution shift scores within them. In practice, the system arranges the distribution shift scores in the SEM unit sequence in chronological order and then performs a Pettitt single mutation point test based on rank statistics. This test calculates the distribution difference before and after any candidate position to determine whether a significant mutation has occurred in the system's behavior distribution at a certain location.

[0034] When the Petitt single mutation point test result indicates a significant mutation in the distribution offset score sequence at a certain position, the system identifies that position as the mutation point. It then uses the mapping relationship between statistical evidence units and the grayscale execution structure to retrieve the corresponding scaling-up stage information, generating a mutation trigger flag bound to the scaling-up stage. This mutation trigger flag serves as a necessary trigger condition for the grayscale execution state transition, outputting corresponding control commands to enable the grayscale release process to automatically adjust the scaling-up state based on changes in system behavior.

[0035] In this embodiment, as the target version's traffic gradually increases to a moderate ramp-up level, the system detects a significant increase in the distribution offset score in continuous time slices. This change is confirmed to occur near the ramp-up phase boundary using the Pettitt single mutation point test. Based on the mutation trigger identifier, a ramp-up freeze operation is automatically triggered to maintain the target version's traffic share at the current stage, preventing further anomalies. Subsequent analysis of the associated feature contribution records in the statistical evidence unit reveals that the anomalies are mainly concentrated in changes in the processing latency distribution of some transaction paths. This change stems from the extended processing paths for specific request types caused by newly introduced business verification logic in the target version. After the relevant logic is adjusted, the system re-enters the canary ramp-up process, ultimately completing the version upgrade.

[0036] To verify the technical effectiveness of the method of the present invention in this application scenario, a comparative statistical analysis was conducted on the gray-scale release process using the method of the present invention and the process using the traditional gray-scale release method. The results are summarized in the table below.

[0037] Table comparing the verification effects of gray-scale release ; As can be seen from the above embodiments, the present invention introduces a system behavior distribution offset quantization mechanism based on RuLSIF operation during the gray release process, and combines it with Pettitt single mutation point test to realize the automatic location of distribution mutations. This enables the system to identify and locate the hidden risks caused by version changes in advance under real production load, thereby effectively solving the problems of gray release verification relying on single indicators, delayed anomaly detection, and lack of evidence support for release control in the prior art. This significantly improves the security, controllability, and engineering practicality of the gray release process of financial business systems.

[0038] 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 method for canary release and intelligent verification in a financial business system, characterized in that, include: In the production environment, construct parallel-running grayscale execution states based on the target version and the baseline version, and generate grayscale execution structures; Within each ramp-up phase defined by the gray-scale execution structure, the processing of business requests in the gray-scale execution state is continuously observed, service operation behavior characteristics and corresponding business result data are collected, forming a multi-dimensional system behavior observation sequence that corresponds one-to-one with the ramp-up phase; Perform rank transformation and staged time slicing on the observation sequence of multidimensional system behavior to construct the input sequence for distribution modeling; Using the distribution modeling input sequence as the sole input, a RuLSIF operation based on relative density ratio is performed to generate a distribution shift score, which is then solidified into a statistical evidence unit bound to the volume expansion phase. Using time series composed of statistical evidence units as the detection object, perform the Pettitt single mutation point test to generate mutation triggering identifiers for the corresponding volume expansion stage; The mutation trigger flag is used as a necessary trigger condition for the transition of the gray-scale execution state. Based on the mutation trigger flag, release control instructions are output to continue scaling, freeze scaling, or rollback the version, and drive the gray-scale execution structure to undergo corresponding state changes.

2. The method for gray-scale release and intelligent verification of a financial business system according to claim 1, characterized in that, The generated grayscale execution structure includes: Create separate sets of isolated running instances for the target version and the baseline version in the same production environment, and bind a unique version identifier to each set of running instances; An instance mapping table is generated based on the set of running instances, and the instance mapping table is maintained as the instance mapping part in the gray-scale execution structure. Generate a set of volume increase stages based on the preset volume increase stage sequence, generate a corresponding stage boundary record for each volume increase stage, and write the stage boundary record into the grayscale execution structure. When starting in grayscale execution mode, the traffic weight field in the instance mapping table is initialized to the weight distribution state dominated by the baseline version, and the initial state is used as the starting point of the stage boundary of the first ramp-up phase. During the rollout phase, the traffic weight field in the instance mapping table is updated in stages based on the stage boundary records. The target version is in the rollout phase. The traffic share meets the following requirements: ； in, Indicates the target version is in stage The corresponding traffic weight value, Indicates the baseline version at stage The corresponding traffic weight value; After each traffic weight update, the instance mapping relationship table and the stage boundary record are consistent and solidified to generate a snapshot of the gray-scale execution structure for the corresponding volume ramp-up stage.

3. The method for gray-scale release and intelligent verification of a financial business system according to claim 2, characterized in that, Creating a collection of mutually isolated running instances includes: In the same production environment, perform independent deployment operations for the target version and the baseline version. Start a preset number of service running instances for each version and write version identifiers and instance identifiers to the instances. Register the target version instances and the baseline version instances as two separate sets of instance lists. Configure the traffic scheduling component to bind requests based only on the instance list and version identifier.

4. The method for gray-scale release and intelligent verification of a financial business system according to claim 1, characterized in that, And form a multi-dimensional system behavior observation sequence corresponding one-to-one with the volume increase phase, including: Read the scaling phase boundary records and instance mapping relationship table from the grayscale execution structure snapshot, determine the time interval of the current scaling phase based on the scaling phase boundary records, and determine the target version instance set and the base version instance set corresponding to the current scaling phase based on the instance mapping relationship table; During the current ramp-up phase, a request identifier is generated for each business request processed by the target version instance set and the baseline version instance set, and service operation behavior characteristics associated with the request identifier are collected during the business request processing. Within the current timeframe of the ramp-up phase, based on the request identifier or the business flow identifier associated with the request identifier, extract the business result data corresponding to the business request from the business processing results, and associate the business result data with the service operation behavior characteristics according to the request identifier; The data associated with the completed request identifier is time-aligned according to a preset time granularity. Data records falling within the same time granularity range are merged into the same time index, and multi-dimensional observation vectors are formed by combining them according to the feature dimensions under the time index. Arrange the multidimensional observation vectors according to the time index to construct the stage observation matrix corresponding to the current volume increase stage. ,in This refers to the sequence number of the volume increase phase; The stage observation matrices corresponding to each volume surge stage are arranged into a multi-dimensional system behavior observation sequence according to the volume surge stage number. .

5. The method for gray-scale release and intelligent verification of a financial business system according to claim 1, characterized in that, The input sequences for generating distribution modeling include: Based on the volume increase stage number, the stage observation matrix corresponding to the current volume increase stage and the stage observation matrix corresponding to the next volume increase stage are selected sequentially from the multi-dimensional system behavior observation sequence, and the two stage observation matrices are determined as a set of adjacent volume increase stage comparison data pairs. For each pair of adjacent volume increase phase comparison data, the observation values ​​in the observation matrices of the two phases are subjected to rank transformation according to the feature dimension, and the observation values ​​are mapped to rank values ​​that reflect the relative order relationship within the phase, thus forming rank-transformed observation data. In the stage observation data after rank transformation, the time segmentation within the stage is performed on adjacent volume expansion stages according to the preset time granularity, generating a set of staged time slices corresponding to each volume expansion stage. Based on the order of time slices in each volume expansion stage, a slice pairing relationship is established for the staged time slice sets of adjacent volume expansion stages. Time slices with the same slice number position form a set of stage comparison slice pairs. For each pair of stage comparison slices, multidimensional observation data within the slices are extracted, and a slice sample representation for distribution comparison is constructed while maintaining the consistency of the feature dimension order. At the same time, the slice sample representation is associated with the corresponding volume expansion stage number, slice number and time interval identifier. The slice sample representations generated in each adjacent volume expansion stage are aggregated in the order of volume expansion stage number and slice number to form the input sequence for distribution modeling.

6. The method for gray-scale release and intelligent verification of a financial business system according to claim 5, characterized in that, The rank transformation process for observations includes: extracting all observations under each feature dimension from the stage observation matrix corresponding to each scaling stage; sorting the observations according to their numerical values; replacing the corresponding observations with their sorted position numbers to form rank values; assigning the same rank value to the group of observations according to a preset parallel rank rule or determining a unified rank value based on their sorting position, thereby converting the stage observation matrix into a rank-transformed observation matrix composed of rank values ​​and using it as input for subsequent staged time slices and distribution modeling.

7. The method for gray-scale release and intelligent verification of a financial business system according to claim 1, characterized in that, The units for generating statistical evidence include: According to the volume increase stage number and slice number, the stage comparison slice pairs are read sequentially from the distribution modeling input sequence; Align the phase comparison slices with the volume surge phase. The sliced ​​sample representation is constructed as a reference sample matrix; Align the phase comparison slices with the volume surge phase. The sliced ​​samples are represented as the target sample matrix, and the dimensions of the reference sample matrix and the target sample matrix are verified. Handling identical values, consistent column order, and missing values; Configure a quantile index set using the reference sample matrix as the sole source. The set of quantile indexes remains fixed within the same publication control link; For each feature dimension of the reference sample matrix Calculate the quantile values ​​corresponding to the quantile index set, and construct the anchor kernel center vector one by one according to the quantile index. Anchoring kernel center vector The Dimensional reference sample matrix Dimension in quantile index The quantile values ​​below form the anchored core center set. ; The anchored core center set is used to generate center records according to fixed serialization rules, and the center record check value is calculated and written into the statistical evidence unit. Using the set of anchor kernel centers as the kernel basis function center, Gaussian kernel basis function is used to calculate the kernel similarity from the sample to the anchor kernel center. Based on the set of anchor kernel centers, the set of Euclidean distances between each pair of anchor kernel centers is calculated, and the kernel width is determined by the median of the set of Euclidean distances. If the median is 0, the kernel width is determined according to the preset lower limit rule, and the kernel width value rule identifier and kernel width are written into the statistical evidence unit. Based on the kernel width, construct reference design matrices for the reference sample matrix and the target sample matrix respectively. With the target design matrix Each row corresponds to a sample, and each column corresponds to an anchor kernel center; Weighting coefficients The regularization coefficient is fixed to the preset configuration value. Fixed by the number of samples and The numerical values ​​are determined by the explicit function, and a linear term vector is constructed. With quadratic term matrix The first-order term vector is obtained by averaging the target design matrix over the samples, and the second-order term matrix is ​​obtained by weighting the second-order term estimates of the target design matrix and the reference design matrix. Obtained by weighted synthesis; Define the relative density ratio function as: ； in, This represents the probability density corresponding to the reference sample matrix. This represents the probability density corresponding to the target sample matrix. This indicates the preset weighting coefficient; In regularization coefficient Solving the least squares closed-form solution under constraints to obtain the parameter vector The expression for solving its parameters is: ； in, It is a quadratic term matrix. For a linear term vector, The identity matrix is ​​used, and the parameter solution status identifier is written into the statistical evidence unit; For each sample in the target sample matrix, extract its corresponding kernel similarity vector. And calculate the relative density ratio estimate. If the relative density ratio estimate is less than 0, it is cut to 0 according to the fixed non-negative cutting rule and the cutting ratio is calculated. The distribution offset score is generated using the squared mean of the deviation from the baseline value by 1. And generate a summary statistical field of the relative density ratio estimate and write it into the statistical evidence unit; A list of feature groups is generated based on the feature dimension identifiers using deterministic grouping rules. And the set of dimension indices corresponding to each feature group, and write the grouping rule identifier and the set of dimension indices into the statistical evidence unit; for each feature group Calculate the median rank for each dimension of the feature group in the reference sample matrix and concatenate them to form the feature group masking vector. ; Without reconstructing the set of anchor kernel centers, redetermining the kernel width, or resolving the parameter vectors. Under the premise of each feature group Construct the target sample matrix after masking The target sample matrix that belongs to the feature group Replace the dimension value with the corresponding feature group mask vector. The remaining dimensions remain unchanged, and the same set of anchored kernel centers, kernel width, and parameter vectors are reused to calculate the shielding offset score. ; The difference between the distribution offset score and the shielding offset score is used as the feature group contribution value. The feature group contribution value and the corresponding feature group identifier are written into the contribution record; Include the scaling phase number, slice number, time interval identifier, slice pair identifier, number of reference samples, number of target samples, feature dimension verification value, and weight coefficient. Kernel width value rules identifier and kernel width, regularization coefficient Value selection rules and values, anchor kernel center set verification values, parameter vectors Checksum, parameter solution status indicator, distribution offset score The relative density ratio estimate summary statistics, pruning ratio and contribution record are written into the statistical evidence unit record to form a statistical evidence unit bound to the volume expansion stage; The records of each statistical evidence unit are aggregated according to the sequence number of the volume expansion stage and the slice number to generate a statistical evidence unit sequence and output it.

8. The method for gray-scale release and intelligent verification of a financial business system according to claim 1, characterized in that, The generation of mutation trigger identifiers includes: The statistical evidence unit sequence is sorted according to the chronological order of its time interval identifiers, resulting in a sequence of length [length missing]. An ordered sequence of evidence; The distribution offset scores are extracted one by one from the ordered evidence sequence to form an offset score sequence, and a location index table is generated simultaneously. Construct an evidence integrity vector for an ordered sequence of evidence, and for each statistical evidence unit, determine whether the following fields exist in its fixed fields: The anchored core center record check value, parameter vector check value, contribution record, pruning ratio, and relative density ratio estimate are summarized and statistically analyzed, and the number of existing fields is used as the integrity count of the statistical evidence unit. When there are scores with the same value in the offset score sequence, the scores are first grouped according to their size, and then sorted a second time within each group according to the integrity count from largest to smallest. If the integrity counts are still the same, the scores are sorted a third time according to the lexicographical order of the release stage number and the slice number in the position index table to generate a stable order of rank distribution. Rank assignment is performed on the offset score sequence based on the stable order to obtain the evidence order sequence. Each of them For position The corresponding score ranks under a stable order are used to determine the evidence order column as the unique input sequence for the Pettitt single mutation point test. Perform Pettitt rank statistic calculation on the evidence order sequence for each candidate time position. Divide the evidence sequence into a front set. With the latter part of the set And calculate the rank statistic of the candidate time positions. : ； in, The sign function is used; the rank statistic is calculated sequentially for all candidate time positions to form a rank statistic sequence; Calculate the corresponding time position for each candidate time position in the rank statistic sequence. , determine The candidate time position with the maximum value is taken as the initial candidate mutation position. Its defining expression is ; And the initial mutation candidate positions, corresponding Write candidate records in association with the location index table; Based on candidate records, perform quantitative structural consistency checks on mutation candidate points and read the initial mutation candidate positions using the position index table. The volume ramp-up stage number and slice number are obtained, and the volume ramp-up stage number and slice number corresponding to at least one preceding position and at least one following position adjacent to the initial mutation candidate position are read. It is then determined whether the initial mutation candidate position satisfies the following conditions: The initial mutation candidate position corresponds to the volume release stage number and at least one of the volume release stage numbers in the adjacent positions have a stage boundary change relationship, and the initial mutation candidate position and at least one of the slice numbers in the adjacent positions satisfy the same slice pair relationship. When the consistency check passes, the initial mutation candidate position is determined as the single mutation point position, and the corresponding scale-up stage number, slice number, time interval identifier and statistical evidence unit identifier are obtained by reverse lookup through the position index table, and the mutation location result is generated; when the consistency check fails, a no-mutation result is generated and the current Pettitt single mutation point test process is terminated. Based on the mutation location results, a mutation trigger identifier is generated and written into the mutation trigger table; the mutation trigger table is output as the trigger input for the grayscale execution state transition.

Images