A Dynamic Quality Analysis and Optimization System Method Based on the Limit Computing Power of Development Model

Abstract

This invention discloses a dynamic quality analysis and optimization system method based on the ultimate computing power of a development model, belonging to the field of software engineering process measurement and quality prediction optimization technology. It addresses the difficulties in quality decision-making caused by inconsistent and untraceable multi-source process data, static defect analysis, and the unstable and easily reversed linkage prediction of productivity and defect escape rate. The method unifies the reading of project process data and standardizes, completes, and weights the data according to a data dictionary to form a unified dataset, automatically generating and versioning the stored documents. It statistically analyzes defect severity and weighted defect quantity within a rolling time window, calculating the defect density sequence and Cpk. A two-layer linkage prediction model is constructed, with gating corrections for productivity and defect escape rate. Satisfaction regression and Monte Carlo output distribution results are used for quality decision-making.

Description

Technical Field

[0001] This invention relates to the field of software engineering process measurement and quality prediction optimization technology, and more specifically, to a dynamic quality analysis and optimization system method based on the ultimate computing power of a development model. Background Technology

[0002] In the software development and delivery process, enterprises often use process management systems, defect management systems, and delivery record files simultaneously. Inconsistent data definitions, missing data, and conflicts are common, leading to incomparable measurement results and untraceable sources. Existing practices largely rely on manual summarization and sampling to generate documents, which is inefficient and struggles to ensure completeness. Defect analysis is mostly static statistics, lacking the ability to depict fluctuations and offsets under a rolling window, and is also difficult to align with specifications. For productivity and defect escape rate prediction, traditional single-layer models are susceptible to feature coupling, resulting in adjacent window flipping and uninterpretable and unstable output. Furthermore, differences in maturity standards lead to inconsistencies in templates and classifications; without a data dictionary and version control, audit reproducibility is difficult. In addition, customer satisfaction is often statistically analyzed independently, without quantitative correlation with process quality indicators, and lacks simulation data based on historical distributions, making risk and resource assessment and improvement decision support insufficient.

[0003] To address the above problems, this invention proposes a solution. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a dynamic quality analysis and optimization system method based on the ultimate computing power of the development model, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] In a preferred embodiment, it includes:

[0007] Based on the project identifier, the system retrieves project process data from the enterprise process management data source and forms a unified project dataset. It then generates multi-level project documents based on a document template set and template engine and stores them in a fixed format.

[0008] Based on the defect field of the enterprise process management data source or the unified project dataset, dynamic statistical analysis is performed according to the time window to generate a dynamic sequence of defect density and process capability index, and to form a defect analysis result record.

[0009] After obtaining the basic productivity prediction and basic defect escape rate prediction of the first-layer model, the defect density fluctuation pattern ratio, process offset amplitude and gating coefficient are calculated by combining the defect density dynamic sequence and process capability index. Based on the offset consistency identifier variable, the deep factor set is determined and the second-layer model input vector is constructed for relearning and adding consistency constraint terms. Then, gating correction and linkage consistency constraint processing are performed and written to the training log and prediction log.

[0010] The customer satisfaction field is extracted based on the project identifier and combined with the productivity and defect escape rate fields to perform correlation analysis and modeling. Monte Carlo sampling is performed according to the personnel skill level to generate a simulated dataset and output the quantile statistics results.

[0011] In a preferred embodiment, basic project information, process metrics, and structured fields of delivery records are read from the enterprise process management data source according to the project identifier and field mapping rules, and written to the data acquisition log. Then, field type validation, unit standardization, and value range validation are performed on the read fields, and a normalization function is used to convert synonymous fields of different units into uniform values. For missing fields, default values ​​are generated according to the default value generation rules of the project data dictionary, and the log records the rule identifier. For multi-source conflict fields with the same business meaning, a weighted adjudication algorithm is performed to select authoritative values ​​according to the preset source credibility weights, and the candidate values, source identifiers, weights, and selection results are written to the log to construct a unified project dataset.

[0012] In a preferred embodiment, after obtaining the unified dataset for the project, a set of document templates is selected based on the project maturity level identifier. The templates are parsed using a template engine, placeholders are located, and the unified dataset fields are replaced and written according to the placeholder binding rules. The date format is unified, the numbering rules are unified, the table rows and columns are expanded, and the directory fields are updated and formatted to generate multi-level project documents. Then, the total number of placeholders and the number of residuals are counted to calculate the placeholder residual rate. When residuals are found, the placeholder name, template version identifier, and bound field name are recorded, and backfilling rereading and replacement writing are triggered. Finally, the document instance is written to the document repository and the project identifier, template version identifier, and generation timestamp are bound to generate a version number. The hash input string and output version number are recorded.

[0013] In a preferred embodiment, defect fields are extracted from enterprise process management data sources or project unified datasets based on project identifiers to form a defect record set and retain the source identifier of each defect data. Then, the severity level of the defects is normalized according to the severity grading rules corresponding to the project maturity standard, and default values ​​are generated for missing fields according to the project data dictionary rules and the rule identifier is recorded.

[0014] Based on this, a rolling time window is introduced. The number and proportion of defects within the window are counted according to their severity, and the weighted defect quantity is calculated by combining the preset weights and written to the analysis log. Then, the defect density is calculated by combining the output scale normalization value corresponding to the time window and forming a dynamic sequence of defect density as the window rolls, and the data source used for each sequence point is recorded. After forming the dynamic sequence of defect density, the process capability index is calculated according to the upper and lower limits of the defect density specification fixed in the project data dictionary. When the standard deviation is zero, it is replaced with the minimum distinguishable standard deviation according to the zero variance processing rule and the configuration version identifier is recorded. Finally, a defect analysis result record is generated, which includes severity distribution, weighted defect quantity, dynamic sequence of defect density and process capability index, and the input fields are accompanied by source identifiers and the configuration items are accompanied by version identifiers.

[0015] In a preferred embodiment, for cases where the basic predictions of productivity and defect escape rate reverse in adjacent time windows, and the prediction outputs are inconsistent with each other when defect density fluctuations increase and process capability index decreases, after obtaining the basic productivity prediction value and basic defect escape rate prediction value of the first-layer model, the process capability index and defect density dynamic sequence of the target time window are read from the defect dynamic statistical analysis result record, and the defect density dynamic sequence is divided into the first half and the second half according to the time sequence. The standard deviations of the two segments are calculated respectively, and the defect density fluctuation pattern ratio is calculated accordingly. At the same time, the process offset amplitude is calculated based on the process capability index and its benchmark value. Then, the process offset amplitude and the fluctuation pattern ratio are used to construct a gating coefficient through a logic function as a calculable controllable quantity for the response intensity of dynamic quantity changes.

[0016] In a preferred embodiment, a bias consistency identifier variable is constructed on the training samples to characterize whether the residuals of the two objectives are in the same or opposite direction. For each candidate feature, its correlation score with the productivity target value, the defect escape rate target value, and the bias consistency identifier variable is calculated. A candidate set that cannot be simultaneously replaced is screened out based on the non-dominated selection criterion, and a set of depth factors is determined by the criterion of maximizing the weakest correlation.

[0017] In a preferred embodiment, the deep factor set, the first-layer model prediction output, the gating coefficient, and the process offset amplitude and fluctuation pattern ratio derived from the above dynamic statistics are used to construct the second-layer model input vector. The first-layer prediction results are then relearned and optimized prediction values ​​are output. A weighted consistency constraint term that is in the same direction as the process offset amplitude constraint is introduced into the training loss to suppress the reverse change in defect escape rate prediction when the process deteriorates. After the second-layer output, a nonlinear correction linked to the productivity offset is applied to the defect escape rate prediction based on the gating coefficient to weaken the flipping of adjacent time windows. Then, the two types of prediction outputs are pruned according to the business boundaries fixed by the quantiles of the historical project process data sources. The calculation caliber of the gating coefficient, the statistical caliber of the process capability index benchmark value, the calculation caliber of the defect density fluctuation pattern ratio, the deep factor set and its source field identifier, the model version identifier, the configuration version identifier, the training and inference time range, the values ​​before and after correction, and the pruning boundary version identifier are written into the training log and the prediction log.

[0018] In a preferred embodiment, the first-layer model is used to output the basic productivity prediction value and the basic defect escape rate prediction value. The second-layer model is used to relearn the prediction results of the first-layer model and output the optimized prediction value. The defect density fluctuation morphology ratio and the process offset amplitude are derived from the defect density dynamic sequence and the process capability index. The gating coefficient is used to control the response intensity to the change of dynamic quantity. The offset consistency identifier variable is used to characterize whether the residuals of the two targets are in the same direction or opposite direction. The depth factor set is used as the input feature of the second-layer model. The input vector of the second-layer model is used to organize the depth factor set, the prediction output of the first-layer model, the gating coefficient, the process offset amplitude and the defect density fluctuation morphology ratio.

[0019] In a preferred embodiment, the customer satisfaction field is extracted by project identifier, and the productivity and defect escape rate fields for that project are also extracted. First, the data is standardized according to the project data dictionary, and default generation and weighted adjudication are performed on missing and conflicting fields respectively, and the basis is recorded. Then, the correlation between customer satisfaction and productivity and defect escape rate is calculated, and the regression correlation coefficient is solved with customer satisfaction as the dependent variable and productivity and defect escape rate as independent variables, and written into the correlation analysis log. Subsequently, productivity and defect density samples are extracted from historical project process data sources according to personnel skill level, and normal distribution is parameterized with sample mean and standard deviation. Monte Carlo sampling is performed according to the preset number of iterations to generate simulated data pairs of productivity and defect density, and the quantile statistics results are output and the version identifier of the distribution parameters used is recorded.

[0020] In a preferred embodiment, it includes: a data normalization module, a defect statistics module, a two-layer optimization module, a satisfaction simulation module, and signal connections between the modules;

[0021] The data unification module is used to obtain project process data from enterprise process management data sources based on project identifiers and form a unified project dataset. It also generates multi-level project documents based on document template sets and template engines and stores them in a fixed manner.

[0022] The defect statistics module is used to perform dynamic statistical analysis on the defect field based on the enterprise process management data source or the unified project dataset by time window, generate a dynamic sequence of defect density and process capability index and form a defect analysis result record;

[0023] The two-layer optimization module is used to calculate the defect density fluctuation pattern ratio, process offset amplitude and gating coefficient by combining the defect density dynamic sequence and process capability index after obtaining the basic productivity prediction value and basic defect escape rate prediction value of the first-layer model. Based on the offset consistency identifier variable, the deep factor set is determined and the second-layer model input vector is constructed for relearning and adding consistency constraint terms. Then, gating correction and linkage consistency constraint processing are performed and written to the training log and prediction log.

[0024] The satisfaction simulation module is used to extract the customer satisfaction field by project identifier and combine it with the productivity field and defect escape rate field to perform correlation analysis and modeling. It performs Monte Carlo sampling according to the skill level of personnel to generate a simulation dataset and outputs quantile statistics.

[0025] The technical effects and advantages of the dynamic quality analysis and optimization system method based on the ultimate computing power of the development model of the present invention are as follows:

[0026] This invention achieves traceability of field sources and reproducibility of definitions through a data dictionary and acquisition logs, and drives automatic backfilling based on placeholder residual rate, significantly reducing document preparation and review costs. It introduces a rolling window defect density sequence and Cpk to quantify fluctuations and offsets to specification limits. A two-layer model, combining gating, non-dominated factor selection, and consistency constraints, suppresses adjacent window prediction flipping, making the linkage between productivity and escape rate more stable. Satisfaction regression is superimposed with a hierarchical Monte Carlo output interval distribution, enhancing risk assessment and resource allocation capabilities for quality improvement. It also supports automatic adaptation to CMMI Level 3 / 5 differences. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the pipeline flow path for a dynamic quality analysis and optimization system method based on the ultimate computing power of a development model, according to the present invention.

[0028] Figure 2 This is a swimlane diagram illustrating the control flow of a dynamic quality analysis and optimization system method based on the ultimate computing power of a development model, as described in this invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Example

[0031] This invention discloses a dynamic quality analysis and optimization method based on the ultimate computing power of a development model, comprising:

[0032] Step 1: Unified acquisition of project process data and automatic generation of multi-level project documents;

[0033] First, such as Figure 1 As shown, for a project to be processed, a data read operation is initiated based on the project identifier. Following pre-configured field mapping rules, structured fields are read from the project process data source and written to the data acquisition log. The project process data source is an existing process management data storage carrier of the enterprise, containing basic project information data, project process measurement data, and delivery-related record data. The basic project information data includes the project name, project stage, version identifier, and module list; the project process measurement data includes stage work hours, output scale, and defect fields; and the delivery-related record data includes requirement items, design items, test items, and their status fields. Data reading is completed using either database querying or process record file parsing. During the reading process, a data acquisition log is generated synchronously. This log records at least the field name, field source identifier, read timestamp, and project identifier to identify the source of each field.

[0034] Then, consistency checks and normalization are performed on the read fields to form a unified dataset for the project. The consistency checks and normalization include at least field type verification, unit standardization, and value range verification. A normalization function is used to standardize fields with the same business meaning but different units into comparable values ​​of the same standard. This normalization function satisfies the following relationship: ;

[0035] Where x represents the original field value read from the project process data source, and u represents the unit or caliber identifier of the field. This represents the normalized field value; This function performs conversion, rounding, and formatting of field values ​​according to the unit conversion rules defined in the project data dictionary. By applying the above normalization function to each field, the same type of field in the project's unified dataset maintains the same scope and type across different sources and recording methods.

[0036] When a field is detected as missing, a default value is generated according to the default value generation rules defined in the project data dictionary, and a default generation flag and the rule identifier are recorded in the data acquisition log. When multiple source fields with the same business meaning are detected and their values ​​are inconsistent, instead of relying on manual judgment, a weighted decision algorithm is executed to determine the authoritative value. Specifically, a source credibility weight is pre-assigned to each field source. And select the authoritative value according to the following relationship: ;

[0037] Among them, set This represents the set of candidate values ​​for the same business field read from different sources, where s(i) represents the candidate values. Corresponding source identifier, This represents the trusted weight of the source identifier, and v represents the final value written to the project's unified dataset after the decision. During the decision-making process, the conflict candidate values, their respective source identifiers, their respective weights, and the final selection result are simultaneously written to the data acquisition log.

[0038] After obtaining the unified project dataset, a set of document templates is selected based on the project maturity level identifier, and template parsing and dynamic data binding are performed to generate multi-level project documents. The document template set includes different template versions adapted to CMMI Level 3 and CMMI Level 5; each template version has a pre-defined set of placeholder fields and binding rules for placeholders to fields in the unified project dataset. Template parsing and dynamic data binding are implemented using a template engine: the template document is parsed into an editable structure, placeholders are located item by item, and replacements are performed using the corresponding field values ​​in the unified project dataset; during the replacement process, formatted rendering is performed, which includes at least unified date formatting, unified numbering rules, table row and column expansion, and automatic updates to table of contents fields. As one implementation method, the template engine uses POI or XWPF to read Word templates, locate placeholders, and perform replacements.

[0039] Subsequently, a fill integrity check is performed on the generated document instances. This check is not based on manual sampling, but rather on the placeholder retention rate as the core metric. Specifically, this involves counting the total number of placeholders that should be replaced in the template. Count the number of placeholders that have not yet been replaced in the document instance. And calculate the placeholder residual rate R: ;

[0040] R is used to characterize the completeness of the field filling generated this time. When R>0, the residual placeholder name, the template version identifier, and the corresponding bound field name are written to the generation log, and the backfill operation is triggered. The backfill operation rereads the corresponding field according to the field mapping rules and performs the replacement writing again. When the reread still fails, the placeholder is retained and the pending completion mark and the failure reason description are recorded in the generation log, so that the generation result can be verified.

[0041] Finally, the generated multi-level project documents are stored in a versioned manner. Specifically, each generated document instance is written to a document repository and bound to a project identifier, template version identifier, and generation timestamp to generate a version number. The version number can be generated using a hash method to obtain a fixed-length identifier, thus avoiding conflicts and confusion between different generation batches. The data items used for hash generation all come from the project identifier, template version identifier, and generation timestamp explicitly recorded in this step, and the hash input string and output version number are recorded together in the generation log, thereby ensuring that the version number source is clear and reproducible and verifiable.

[0042] Step 2: Dynamic statistical analysis of defect data and calculation of process capability index;

[0043] Step 2 performs dynamic statistical analysis on the defect data and calculates the process capability index to quantitatively reflect the stability and deviation of the defect process. The defect data used in this step all come from the defect fields of the project process data source or the corresponding defect fields in the unified project dataset formed in Step 1, and each defect data retains its source identifier in the data acquisition log.

[0044] First, the defect record set is extracted and its fields are standardized. Specifically, defect records related to the project are read from the project process data source based on the project identifier, forming a defect record set. Each defect record includes at least a defect identifier, defect discovery time, defect module identifier, defect discovery stage identifier, and defect severity level identifier. For the defect severity level identifier, it is standardized according to pre-configured severity grading rules: when the project adopts CMMI Level 3, the severity grading rules standardize the defect severity level identifier into a three-level set; when the project adopts CMMI Level 5, the severity grading rules standardize the defect severity level identifier into a four-level set, thus achieving consistent processing of defect severity statistical caliber under different maturity standards. If a defect record is found to have missing fields, default values ​​are generated according to the default generation rules defined in the project data dictionary, and the default generation flag and rule identifier are recorded in the data acquisition log, making the defect record set computable and traceable.

[0045] Subsequently, dynamic statistical calculations of the defect severity distribution are performed. To achieve this dynamism, this step introduces the concept of a time window: a time window refers to a subset of defect records extracted within a preset time span, based on the current statistical time point; the time window is defined by a start time and an end time, and can be updated forward by a preset step size. For any subset of defect records within a time window, the number of defects at each severity level is counted, and their proportion is calculated to form the severity distribution result. To ensure that the distribution result can numerically express the quality impact of prioritizing severe defects, this step further calculates the severity-weighted defect quantity, defined as follows: ;

[0046] Where K represents the number of severity levels; when using the CMMI Level 3 standard, K=3, and when using the CMMI Level 5 standard, K=4; This represents the number of defects corresponding to the kth severity level within the time window; The weighting coefficient represents the k-th severity level. This weighting coefficient is predefined and fixed in the project data dictionary, and each weighting coefficient corresponds one-to-one with a severity level identifier. By calculating S, the number of defects across multiple severity levels is uniformly mapped to a weighted defect quantity, which characterizes the overall severity level of defects within that time window. The severity distribution statistics and weighted defect quantity calculation described above are among the direct outputs of this step and are recorded in the analysis log. The analysis log must at least include the start and end times of the time window and the number of defects at each level. Weighting coefficients The version identifier and the calculation result of the weighted defect quantity S.

[0047] After obtaining the distribution results of defect quantity and severity, defect density is calculated and a dynamic sequence is generated. Defect density characterizes the strength of defect quantity relative to output size, with the output size field taken from the project's unified dataset corresponding to the current time window. The defect density is defined as follows: ;

[0048] Where D represents the defect density; N represents the total number of defects within the time window, i.e.: Q represents the output scale value corresponding to the time window. The unit and scope of the output scale value are defined by the project data dictionary and obtained as a normalized value through the normalization function in step one. The defect density D is calculated once for each rolling update of the time window, thereby forming a dynamic sequence of defect densities arranged by time. The source identifiers of N and Q used in the calculation of each sequence point are written into the analysis log to ensure that the source of the defect density is clear and verifiable.

[0049] After generating the dynamic sequence of defect density, the process capability index is calculated to quantify the stability and deviation of the defective process relative to the quality specification limits. In this step, the quality specification limits are defined as the lower and upper limits of the defect density, denoted as LSL and USL, respectively. Both are predefined and configured by the project data dictionary. Different LSL and USL configuration versions can be used when different maturity standards or project types are adopted. The configuration version identifier is written to the analysis log to ensure consistency and traceability. This step calculates the mean μ and standard deviation of the defect density within the selected statistical window. Based on this, the process capability index Cpk is calculated: ;

[0050] Where μ represents the arithmetic mean of the dynamic defect density sequence within the statistical window; σ represents the standard deviation of the dynamic defect density sequence within the statistical window; USL represents the upper limit of the defect density specification; LSL represents the lower limit of the defect density specification; and Cpk represents the process capability index. To avoid the situation where the standard deviation is zero, which would make it uncalculateable, when a defect density is detected... In this step, the zero variance handling rule preset in the project data dictionary is followed: σ is set to the minimum distinguishable standard deviation. The To define a positive threshold explicitly in the project data dictionary, and to record the zero variance treatment marker in the analysis log. The configuration version identifier. The Cpk calculated above, along with the corresponding μ, σ, USL, LSL, and the start and end times of the statistical window, are written into the analysis log, thus ensuring that the calculation basis of the process capability index is complete and verifiable.

[0051] Finally, the analysis results of this step are output and a defect analysis result record is generated. The defect analysis result record includes at least: the number and proportion of defects at each severity level, the weighted defect quantity S, the dynamic sequence of defect density, and the process capability index Cpk; and, for each input field used for calculation, the defect analysis result record includes its source identifier, and for each configuration item, it includes its configuration version identifier.

[0052] Step 3: Predicting the linkage between productivity and defect escape rate and optimizing the two-layer model;

[0053] First, the training dataset is constructed and field definitions are standardized. Specifically, historical project records are read from the organization's historical project process data source. These records include at least the following fields: historical project identifier, historical project output scale, historical project input, historical project cycle, historical project productivity, and historical project defect escape rate. The historical project productivity field is either the ratio of output scale to input within a specified cycle or an existing enterprise metric. The historical project defect escape rate field is either the ratio of the number of defects discovered within a specified observation period after project delivery to the total number of defects before delivery or an existing enterprise metric. All read fields are processed using the same normalization function as in step one to ensure consistency in units, definitions, and types across different historical projects. The normalization rules are derived from the project data dictionary, and the version identifier of the project data dictionary used is recorded in the training log of this step. Missing fields are processed using default values ​​generated from the project data dictionary, and the default generation tag and rule identifier are recorded in the training log, thus ensuring the training dataset is trainable and traceable. The above processing generates a historical training dataset, which is then divided into an input feature set and an output target set. The input feature set is denoted as X, and the output target set includes productivity target values ​​and defect escape rate target values, denoted as X, ... and .

[0054] Subsequently, the first-layer model is trained to obtain baseline prediction results. The first-layer model employs supervised learning training, mapping the input feature set X to predicted productivity and defect escape rate values. To meet the requirement of linked prediction, the first-layer model trains separate first-layer prediction functions for productivity and defect escape rate on the same input feature set X, defined as follows: ; ; in, This represents the first-level prediction function for productivity. The first-level prediction function represents the defect escape rate; This represents the productivity prediction value output by the first-layer model. This represents the defect escape rate prediction output by the first-layer model. The first-layer model is trained by minimizing the loss function; productivity prediction uses regression loss, and defect escape rate prediction uses regression loss or probabilistic regression loss. The loss function is uniformly denoted as... and the true target value from the historical training dataset. and The parameters are optimized using these as supervisory signals. After training, the model type identifier, training parameter configuration identifier, training data time range, and error metric on the validation set of the first layer are written to the training log to indicate the model's origin and training basis.

[0055] The first-layer model in this step corresponds to the first-layer model version 1.0 structure in the document, which is to input historical data and output basic predictions.

[0056] It should be noted that when the first-layer model outputs basic predicted values ​​for productivity and defect escape rate, the effects of multiple factors in the input feature set on the two prediction targets may be coupled. Combining the dynamic sequence of defect density and the process capability index obtained from dynamic statistical analysis of defect data, it can be seen that the fluctuation and offset of these two dynamic quantities do not change monotonically as the time window rolls. An anomaly has emerged in practice: within certain time windows, the fluctuation of the defect density dynamic sequence increases and the process capability index decreases, but the basic defect escape rate prediction value decreases instead of increasing. Furthermore, the unidirectional offset relationship between the basic productivity prediction value and the basic defect escape rate prediction value reverses in adjacent time windows, making it impossible to determine which combinations of factors in the input feature set triggered this reversal, and preventing the prediction output from responding consistently to changes in the dynamic quantities.

[0057] Therefore, in this embodiment, after obtaining the basic productivity and defect escape rate predictions of the first-layer model, the construction and optimization training of the second-layer model are performed. The second-layer model takes the prediction output of the first-layer model, the deep factor selection results, and the adjustment parameters derived from the dynamic statistics of the defect process as inputs, relearns the prediction results of the first-layer model, and outputs optimized prediction values.

[0058] To ensure the adjustment parameters have a calculable source, the process capability index, defect density dynamic sequence, and standard deviation of the defect density dynamic sequence within the target time window are first retrieved from the defect dynamic statistical analysis results record. The defect density dynamic sequence within the target time window is then divided into a first half and a second half subsequence in chronological order. The standard deviations of the first and second half subsequences are calculated separately, and the defect density fluctuation pattern ratio is also calculated. :

[0059] ;

[0060] Among them, the dynamic sequence of defect density within the target time window is arranged in chronological order from early to late; the sequence is divided into a first half subsequence and a second half subsequence according to the number of sequence points. When the number of sequence points is odd, the subsequence with the earlier time order is used as the first half subsequence, and the rest are used as the second half subsequence, so as to ensure that the division rule of the first half subsequence or the second half subsequence is unique and verifiable. and Calculated separately within their respective subsequences according to the statistical definition of standard deviation, these values ​​characterize the fluctuation strength of the defect density dynamic sequence before and after the window; ε is a positive threshold to avoid a zero denominator, and is also used to ensure... It is positive to satisfy the subsequent ln( The computability prerequisite for ε is given by the model configuration file, which also records the configuration version identifier, enabling... The calculation method is traceable.

[0061] Simultaneously, the process deviation magnitude is calculated based on the process capability index. : ;

[0062] in, This represents the process capability index within the target time window. This represents the baseline value of the process capability index. The value of is specified by the statistical caliber configuration item in the model configuration file: when the statistical caliber configuration item is the median, Take the median of the process capability index from the historical training dataset; when the statistical caliber configuration item is set to a specified quantile... A specified quantile of the process capability index is taken from the historical training dataset. Both the quantile parameter and the time range of the historical training data used in the statistics are written to the training log during training, and a configuration version identifier is recorded. The source is verifiable. It is used to convert the deviation of the process capability index from the benchmark value into a calculable quantity, and to serve as the adjustment input for the subsequent gating coefficient g and the consistency constraint term.

[0063] In obtaining and Then, a gating coefficient g is constructed to control the response strength of the predicted path to the above computable quantities. The gating coefficient is calculated according to the following relationship: ;

[0064] Among them, κ is used to control the steepness of the response of the gating coefficient g to the changes in the process offset amplitude RC and the defect density fluctuation pattern ratio RD, so as to increase the suppression of abnormal reversal when the process offset amplitude and fluctuation pattern ratio are significant, and reduce the intervention intensity when the process is stable; κ is given by the model configuration file and records the configuration version identifier to ensure that the gating response caliber is traceable. Represents a logical function that satisfies , where e is a natural constant. The resulting gating coefficient g ranges from [0,1] and is updated according to the ratio of process offset amplitude to fluctuation pattern within the target time window.

[0065] Deep factor selection is then performed. To ensure that the factor selection reflects the phenomenon of the baseline productivity forecast and the baseline defect escape rate forecast shifting in the same direction within adjacent time windows, a shift consistency indicator variable is first constructed on the training samples. : ;

[0066] in, and Let represent the true productivity target value and the true defect escape rate target value of the i-th training sample, respectively. and These represent the base productivity prediction and base defect escape rate prediction values ​​output by the first-layer model for the i-th training sample, respectively. This is a sign function; it outputs +1 when the input is greater than zero, -1 when it is less than zero, and 0 when it is equal to zero. The offset consistency flag variable is used to convert residuals in the same or opposite direction into a computable target sequence.

[0067] Then, for each candidate feature Calculate the correlation scores between the correlation scores and the productivity target, the defect escape rate target, and the deviation consistency indicator variable; the correlation scores are calculated using the Pearson correlation coefficient, defined as follows: ;

[0068] in, This represents the value of the i-th training sample on the j-th candidate feature. This represents the mean of the j-th candidate feature on the training samples; This represents the target value of the i-th training sample. represents the target mean; m represents the number of training samples. We obtain... and .

[0069] To avoid compressing the multi-objective influence into a simple linear weighting, this step uses a non-dominated selection criterion to determine the set of depth factors. Specifically, if another candidate feature exists... , making , , If at least one inequality is strictly greater than, then the candidate feature is determined. Candidate features Domination: Candidate features not dominated by any candidate feature are grouped into a non-dominated candidate set. Then, within the non-dominated candidate set, a set of deep factors is determined using the weakest correlation maximization criterion: for each candidate feature, the minimum absolute value of its three correlation terms is calculated. ,according to Select the top few candidate features from largest to smallest as the depth factor set. The upper limit of the dimension of the deep factor set is given by the model configuration file and the configuration version identifier is recorded. Through the above selection, the candidate features entering the deep factor set have an explanatory power that cannot be replaced simultaneously for all three objectives, and avoid the overshadowing of flipped related factors due to strong correlation of a single objective.

[0070] After completing the depth factor selection, the input vector Z of the second-layer model is constructed, incorporating the predicted output of the first-layer model, the set of depth factors, and the gating coefficients into the input, defined as: ;

[0071] in, and These are the base productivity prediction and base defect escape rate prediction values ​​output by the first-layer model, respectively, where g is the gating coefficient. For process offset magnitude, The defect density fluctuation pattern ratio is used. A second-layer prediction function is trained, defined as follows: ;

[0072] in, This represents the second-level prediction function for productivity. The second-level prediction function represents the defect escape rate; This represents the productivity optimization prediction output by the second-layer model. This represents the optimized defect escape rate prediction output by the second-layer model. The second-layer model is trained by minimizing the loss function, using the actual productivity target value and the actual defect escape rate target value from the historical training dataset as supervision signals.

[0073] In the training of the second-layer model, to ensure that the prediction results produce a stable response to changes in the process offset magnitude and fluctuation pattern ratio, this step adds a consistency constraint term to the loss function: For each training sample, a consistency constraint term is calculated. and construct constraint terms. ,in This represents the process offset amplitude calculated within the corresponding time window for the training sample. The consistency constraint terms of each training sample are weighted by the gating coefficient g, and the average is added to the base loss and written into the loss function. This ensures that the larger the process offset amplitude, the stronger the consistency constraint, thereby suppressing the anomalous situation where the process capability index deteriorates while the defect escape rate prediction changes in the opposite direction. After training, the model type identifier of the second-layer model, the feature list of the deep factor set and its source field identifiers, the calculation method of the gating coefficient, the statistical method of the process capability index baseline value, the calculation method of the defect density fluctuation morphology ratio, the training parameter configuration identifier, the training data time range, and the validation error index are written into the training log.

[0074] After the second-layer model output, gating correction is performed to mitigate the anomalous reversal of the unidirectional offset relationship between adjacent time windows. Specifically, this involves calculating the productivity offset. The defect escape rate prediction is then subjected to a nonlinear correction based on the gating coefficient to obtain the corrected defect escape rate prediction. : ;

[0075] in, Represents the hyperbolic tangent function. To correct the scaling parameters, these parameters are provided in the model configuration file and the configuration version identifier is recorded. By making the correction term nonlinear with g, the correction strength increases when the process offset magnitude is significantly greater than the fluctuation pattern, and decreases when the process is stable, thus enabling the defect escape rate prediction to adaptively respond to changes in dynamic statistics.

[0076] Subsequently, a linkage consistency constraint is executed to ensure that the productivity forecast and the defect escape rate forecast meet the preset business consistency boundary. The linkage consistency constraint is implemented through a reasonable range trimming method: allowing ranges are defined for both the productivity forecast and the defect escape rate forecast. and The boundary values ​​are obtained from the quantile statistics of historical training data and are fixed as configuration items, with the configuration version identifier recorded in the training log. A pruning function is then executed on the predicted output: ;

[0077] Where y represents the predicted value to be pruned, and a and b represent the lower and upper bounds of the allowable range, respectively. This represents the trimming function. Calculate the final productivity forecast values ​​respectively. And the final defect escape rate prediction value And the gating coefficient g and the process offset amplitude Defect density fluctuation pattern ratio The values ​​before and after correction, as well as the boundary configuration version identifier used, are written into the prediction log to make the prediction process verifiable.

[0078] Finally, predictive inference is performed on the target project, and the prediction results are recorded. Specifically, the set of input feature fields of the target project is read from the unified project dataset formed in step one. The data is processed using the same normalization function as during the training phase to form the inference input; the inference input is then substituted into the first-layer prediction function to obtain... and Then construct the second layer input. Substituting this into the second-level prediction function yields... and The final predicted output is obtained through the pruning function. and .

[0079] Among them, RCtar and RDtar are obtained by recording the results of the dynamic statistical analysis of defects in the time window corresponding to the target project according to the same calculation method in step three, and gtar is calculated by RCtar and RDtar according to the gating coefficient calculation formula.

[0080] The final prediction output and its computation path are recorded as a prediction result record. This prediction result record includes at least: the target project identifier, the model version identifier (i.e., the first-layer model version identifier and the second-layer model version identifier), the deep factor set feature list, the input feature field source identifier, and the prediction output. and The trimming boundary configuration includes version identifiers and predicted timestamps.

[0081] Step 4: Customer satisfaction correlation analysis and Monte Carlo simulation data generation;

[0082] First, the customer satisfaction dataset and process quality indicator dataset are extracted and their definitions are standardized. Specifically, based on the project identifier, the customer satisfaction field corresponding to the delivery batch of the project is read from the customer satisfaction data source. The customer satisfaction field includes at least one or more of the following: customer satisfaction rating field or net promoter score field. Simultaneously, the defect escape rate field and productivity field corresponding to the project are read from the prediction result record in step three or the unified project dataset to form the process quality indicator dataset. The customer satisfaction field, defect escape rate field, and productivity field are respectively validated for type, unit standardization, and value range according to the project data dictionary. Missing fields are processed according to the default value generation rule, and the default generation mark and rule identifier are recorded. For conflicting fields with the same business meaning, the authoritative value is determined according to the weighted adjudication algorithm in step one, and the adjudication basis is recorded, thereby obtaining a unified version of the customer satisfaction dataset and the process quality indicator dataset.

[0083] Subsequently, a correlation analysis model was performed to examine the relationship between customer satisfaction and process quality indicators. The correlation analysis employed a combination of correlation calculation and regression fitting: first, the correlation coefficients between the customer satisfaction field and the defect escape rate field, and between the customer satisfaction field and the productivity field, were calculated to obtain the correlation results; then, a multiple linear regression model was constructed with customer satisfaction as the dependent variable and defect escape rate and productivity as independent variables to obtain the correlation coefficients. The correlation coefficients were calculated using the Pearson correlation coefficient, defined as follows: ;

[0084] Where X represents the first indicator sequence involved in the correlation calculation, and y represents the second indicator sequence involved in the correlation calculation. and Let represent the values ​​of the i-th sample, and Let X and Y represent the sequence means, N represent the sample size, and r(X,Y) represent the correlation coefficient. Calculate the correlation coefficient between customer satisfaction and defect escape rate. And the correlation coefficient between customer satisfaction and productivity. ,in The field sequence represents customer satisfaction, E represents defect escape rate, and P represents productivity. The sample source, time range, and field source identifier for correlation calculation are written into the correlation analysis log.

[0085] After the correlation calculation is completed, a multiple linear regression model is established to obtain the correlation coefficient, which is defined as follows: ;

[0086] in, This represents the value for the customer satisfaction field, P represents the value for the productivity field, and E represents the defect escape rate field. Represents the intercept term coefficient. This represents the regression coefficient of productivity on customer satisfaction. This represents the regression coefficient of defect escape rate on customer satisfaction. Represents the residual term. Regression coefficients. The least squares method was used to solve the problem, and the number of samples used, field source identifiers, and solution error statistics were written into the association analysis log.

[0087] After completing the correlation analysis modeling, Monte Carlo simulation data generation is performed to obtain distributed simulation results for productivity and defect data. The Monte Carlo simulation constructs the input distribution using a hierarchical parameterization of personnel skill levels: personnel skill levels include two categories: ordinary personnel skill levels and skilled personnel skill levels. Productivity and defect density samples corresponding to ordinary personnel skill levels, and those corresponding to skilled personnel skill levels, are extracted from historical project process data sources. Statistical fitting is performed on the samples to obtain the productivity and defect density distribution parameters for ordinary personnel skill levels, and for skilled personnel skill levels. To avoid introducing undefined new distribution types, this step uniformly models the simulated inputs for productivity and defect density as normal distributions, using the sample mean and sample standard deviation as distribution parameters; the productivity distribution parameter for ordinary personnel skill levels is denoted as... The productivity distribution parameter of skilled personnel skill level is denoted as The defect density distribution parameter for the skill level of ordinary personnel is denoted as... The defect density distribution parameter of skilled personnel's skill level is denoted as The above mean and standard deviation were obtained from historical samples, and their sample range and field source identifiers were written into the simulation parameter log.

[0088] Subsequently, Monte Carlo sampling and result calculation were performed. The number of simulation iterations was set to m, where m is a preset positive integer and recorded in the simulation parameter log; for each iteration... Based on the skill level of the personnel, random samples of productivity and defect density are drawn to generate the simulated productivity values ​​for this iteration. Compared with the simulated value of defect density The sampling process satisfies the following relationship: ; ; in, denoted by 'normal distribution', g represents the skill level category identifier, with values ​​ranging from ordinary skill level to skilled skill level. and These represent the mean and standard deviation of the productivity distribution for this skill level category, respectively. and These represent the mean and standard deviation of the defect density distribution for this personnel skill level category, respectively. For each iteration... and Create simulated data pairs and write them into the simulated dataset; the simulated dataset shall at least record the iteration number, personnel skill level category identifier, simulated productivity value, simulated defect density value, and version identifier of the distribution parameters used.

[0089] Finally, a distributed statistical output is generated based on the simulated dataset. Specifically, this involves generating a sequence of simulated productivity values ​​from the simulated dataset. With the sequence of simulated defect density values Quantile statistics are calculated to provide prediction intervals. The thresholds corresponding to the quantiles, the number of iterations (m) used in their calculation, the personnel skill level category identifier, and the distribution parameter version identifier are written into the simulation results log. Quantile calculation is performed using a sorting and positional method: after sorting the sequence to obtain an ordered sequence, values ​​are taken as quantiles at preset quantile positions.

[0090] This invention also proposes a dynamic quality analysis and optimization system based on the ultimate computing power of the development model, such as... Figure 2 As shown, it includes: a data normalization module, a defect statistics module, a two-layer optimization module, a satisfaction simulation module, and signal connections between the modules;

[0091] The data unification module is used to obtain project process data from enterprise process management data sources based on project identifiers and form a unified project dataset. It also generates multi-level project documents based on document template sets and template engines and stores them in a fixed manner.

[0092] The defect statistics module is used to perform dynamic statistical analysis on the defect field based on the enterprise process management data source or the unified project dataset by time window, generate a dynamic sequence of defect density and process capability index and form a defect analysis result record;

[0093] The two-layer optimization module is used to calculate the defect density fluctuation pattern ratio, process offset amplitude and gating coefficient by combining the defect density dynamic sequence and process capability index after obtaining the basic productivity prediction value and basic defect escape rate prediction value of the first-layer model. Based on the offset consistency identifier variable, the deep factor set is determined and the second-layer model input vector is constructed for relearning and adding consistency constraint terms. Then, gating correction and linkage consistency constraint processing are performed and written to the training log and prediction log.

[0094] The satisfaction simulation module is used to extract the customer satisfaction field by project identifier and combine it with the productivity field and defect escape rate field to perform correlation analysis and modeling. It performs Monte Carlo sampling according to the skill level of personnel to generate a simulation dataset and outputs quantile statistics.

[0095] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0096] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0097] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and inventive constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0098] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0099] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0100] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dynamic quality analysis and optimization method based on the ultimate computing power of a development model, characterized in that, include: Based on the project identifier, the system retrieves project process data from the enterprise process management data source and forms a unified project dataset. It then generates multi-level project documents based on a document template set and template engine and stores them in a fixed format. Based on the defect field of the enterprise process management data source or the unified project dataset, dynamic statistical analysis is performed according to the time window to generate a dynamic sequence of defect density and process capability index, and to form a defect analysis result record. After obtaining the basic productivity prediction and basic defect escape rate prediction of the first-layer model, the defect density fluctuation pattern ratio, process offset amplitude and gating coefficient are calculated by combining the defect density dynamic sequence and process capability index. Based on the offset consistency identifier variable, the deep factor set is determined and the second-layer model input vector is constructed for relearning and adding consistency constraint terms. Then, gating correction and linkage consistency constraint processing are performed and written to the training log and prediction log. The customer satisfaction field is extracted based on the project identifier and combined with the productivity and defect escape rate fields to perform correlation analysis and modeling. Monte Carlo sampling is performed according to the personnel skill level to generate a simulated dataset and output the quantile statistics results.

2. The dynamic quality analysis and optimization method based on the ultimate computing power of the development model according to claim 1, characterized in that: Based on the project identifier, the system reads basic project information, process metrics, and structured fields of delivery records from the enterprise process management data source according to field mapping rules and writes them to the data acquisition log. Then, it performs field type validation, unit standardization, and value range validation on the read fields and uses a normalization function to convert synonymous fields with different units into uniform values. For missing fields, it generates default values ​​according to the default value generation rules of the project data dictionary and records the rule identifier in the log. For multi-source conflict fields with the same business meaning, it performs a weighted decision algorithm according to the preset source credibility weight to select authoritative values ​​and writes the candidate values, source identifiers, weights, and selection results to the log to build a unified dataset for the project.

3. The dynamic quality analysis and optimization method based on the ultimate computing power of the development model according to claim 2, characterized in that: After obtaining the unified dataset for the project, a set of document templates is selected based on the project maturity level identifier. The templates are parsed using a template engine, placeholders are located, and written according to the placeholder binding rules. The unified dataset fields are replaced and written, and the date format, numbering rules, table row and column expansion, and directory field updates are performed to format and render multi-level project documents. Then, the total number of placeholders and the number of remnants are counted to calculate the placeholder remnant rate. When remnants are found, the placeholder name, template version identifier, and bound field name are recorded, and backfilling rereading and replacement writing are triggered. Finally, the document instance is written to the document repository and bound with the project identifier, template version identifier, and generation timestamp to generate a version number. The hash input string and output version number are recorded.

4. The dynamic quality analysis and optimization method based on the ultimate computing power of the development model according to claim 3, characterized in that: Based on the project identifier, defect fields are extracted from enterprise process management data sources or project unified datasets to form a defect record set and the source identifier of each defect data is retained. Then, the severity level of the defects is standardized according to the severity grading rules corresponding to the project maturity standard, and default values ​​are generated for missing fields according to the project data dictionary rules and the rule identifier is recorded. Based on this, a rolling time window is introduced. The number and proportion of defects within the window are counted according to their severity, and the weighted defect quantity is calculated by combining the preset weights and written to the analysis log. Then, the defect density is calculated by combining the output scale normalization value corresponding to the time window and forming a dynamic sequence of defect density as the window rolls, and the data source used for each sequence point is recorded. After forming the dynamic sequence of defect density, the process capability index is calculated according to the upper and lower limits of the defect density specification fixed in the project data dictionary. When the standard deviation is zero, it is replaced with the minimum distinguishable standard deviation according to the zero variance processing rule and the configuration version identifier is recorded. Finally, a defect analysis result record is generated, which includes severity distribution, weighted defect quantity, dynamic sequence of defect density and process capability index, and the input fields are accompanied by source identifiers and the configuration items are accompanied by version identifiers.

5. The dynamic quality analysis and optimization method based on the ultimate computing power of the development model according to claim 4, characterized in that: To address the situation where the first-layer model exhibits an inverted relationship between the baseline predictions of productivity and defect escape rate within adjacent time windows, and inconsistent predicted outputs when defect density fluctuations increase and process capability index decreases, after obtaining the baseline productivity and defect escape rate predictions of the first-layer model, the process capability index and defect density dynamic sequence for the target time window are retrieved from the defect dynamic statistical analysis results record. The defect density dynamic sequence is then divided into a first half and a second half according to time sequence, and the standard deviations of the two segments are calculated respectively. Based on this, the defect density fluctuation pattern ratio is calculated. Simultaneously, the process offset amplitude is calculated based on the process capability index and its benchmark value. Finally, the process offset amplitude and fluctuation pattern ratio are used to construct a gating coefficient through a logic function as a calculable controllable quantity for the response intensity to changes in dynamic quantities.

6. The dynamic quality analysis and optimization method based on the ultimate computing power of the development model according to claim 5, characterized in that: On the training samples, a bias consistency identifier variable is constructed to characterize whether the residuals of the two objectives are in the same or opposite direction. For each candidate feature, the correlation score between it and the productivity target value, the defect escape rate target value, and the bias consistency identifier variable is calculated. Based on the non-dominated selection criterion, a candidate set that cannot be simultaneously replaced is screened out, and the set of depth factors is determined by the criterion of maximizing the weakest correlation.

7. The dynamic quality analysis and optimization method based on the ultimate computing power of the development model according to claim 6, characterized in that: Based on this, the deep factor set, the first-layer model prediction output, the gating coefficient, and the process offset amplitude and fluctuation pattern ratio derived from the above dynamic statistics are jointly constructed into the second-layer model input vector. The first-layer prediction results are relearned and optimized prediction values ​​are output. A weighted consistency constraint term with the same direction constraint as the process offset amplitude is introduced into the training loss to suppress the reverse change of defect escape rate prediction when the process deteriorates. After the second-layer output, a nonlinear correction linked to the productivity offset is applied to the defect escape rate prediction based on the gating coefficient to weaken the flipping of adjacent time windows. Then, the two types of prediction outputs are pruned according to the business boundary solidified by the quantile of the historical project process data source. The calculation caliber of the gating coefficient, the statistical caliber of the process capability index benchmark value, the calculation caliber of the defect density fluctuation pattern ratio, the deep factor set and its source field identifier, the model version identifier, the configuration version identifier, the training and inference time range, the values ​​before and after correction, and the pruning boundary version identifier are written into the training log and the prediction log.

8. The dynamic quality analysis and optimization method based on the ultimate computing power of the development model according to claim 7, characterized in that: The first-layer model is used to output the basic productivity prediction value and the basic defect escape rate prediction value. The second-layer model is used to relearn the prediction results of the first-layer model and output the optimized prediction value. The defect density fluctuation pattern ratio and the process offset amplitude are derived from the defect density dynamic sequence and the process capability index. The gating coefficient is used to control the response intensity to the change of dynamic quantity. The offset consistency identifier variable is used to characterize whether the residuals of the two targets are in the same direction or opposite direction. The depth factor set is used as the input feature of the second-layer model. The input vector of the second-layer model is used to organize the depth factor set, the prediction output of the first-layer model, the gating coefficient, the process offset amplitude and the defect density fluctuation pattern ratio.

9. The dynamic quality analysis and optimization method based on the ultimate computing power of the development model according to claim 8, characterized in that: Extract the customer satisfaction field from the project identifier, and extract the productivity and defect escape rate fields for that project. First, standardize the data according to the project data dictionary, and perform default generation and weighted decision-making for missing and conflicting fields respectively, and record the basis. Then, calculate the correlation between customer satisfaction and productivity and defect escape rate, and solve the regression correlation coefficient with customer satisfaction as the dependent variable and productivity and defect escape rate as independent variables, and write it into the correlation analysis log. Subsequently, extract productivity and defect density samples from historical project process data sources according to personnel skill level, parameterize the normal distribution with sample mean and standard deviation, perform Monte Carlo sampling according to the preset number of iterations to generate simulated data pairs of productivity and defect density, output the quantile statistics results, and record the version identifier of the distribution parameters used.

10. A dynamic quality analysis and optimization system based on the ultimate computing power of a development model, used to implement the dynamic quality analysis and optimization method based on the ultimate computing power of a development model as described in any one of claims 1-9, characterized in that, include: Data normalization module, defect statistics module, two-layer optimization module, satisfaction simulation module, and signal connections between modules; The data unification module is used to obtain project process data from enterprise process management data sources based on project identifiers and form a unified project dataset. It also generates multi-level project documents based on document template sets and template engines and stores them in a fixed manner. The defect statistics module is used to perform dynamic statistical analysis on the defect field based on the enterprise process management data source or the unified project dataset by time window, generate a dynamic sequence of defect density and process capability index and form a defect analysis result record; The two-layer optimization module is used to calculate the defect density fluctuation pattern ratio, process offset amplitude and gating coefficient by combining the defect density dynamic sequence and process capability index after obtaining the basic productivity prediction value and basic defect escape rate prediction value of the first-layer model. Based on the offset consistency identifier variable, the deep factor set is determined and the second-layer model input vector is constructed for relearning and adding consistency constraint terms. Then, gating correction and linkage consistency constraint processing are performed and written to the training log and prediction log. The satisfaction simulation module is used to extract the customer satisfaction field by project identifier and combine it with the productivity field and defect escape rate field to perform correlation analysis and modeling. It performs Monte Carlo sampling according to the skill level of personnel to generate a simulation dataset and outputs quantile statistics.

Images