Enterprise business process oriented rapid modeling generation method

Abstract

The application discloses a quick modeling generation method for enterprise business processes, relates to the technical field of business process modeling, and comprises the following steps: after receiving a business process modeling request, performing structure similarity analysis, comparing a current process structure input with path vectors of a plurality of process templates, judging structure coincidence degree and conditional node semantic consistency, and identifying whether the case belongs to a plurality of template combined process generation; in the case of identifying the plurality of template combined process generation, extracting each path trigger frequency original value corresponding to a conditional judgment node, a frequency sampling time range, a sample quantity, a collection strategy label and a path trigger distribution density, constructing a normalized offset vector, and outputting a normalized difference result. The application solves the problem of inconsistent path frequency statistics caliber in the multi-template process combined modeling, and realizes the normalized evaluation and dynamic optimization setting of a default path.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of business process modeling, in particular to a rapid modeling generation method for enterprise business processes. BACKGROUND

[0002] The rapid modeling generation for enterprise business processes refers to quickly constructing a modeling scheme that meets the actual business process requirements of an enterprise through information technology means, especially automation, visualization, templating, rule-driven or intelligent algorithms, to support the digitalization, systematization and automation management of business processes. At present, the existing rapid modeling generation technology for enterprise business processes is mainly realized through the following ways: first, a visual modeling interface is provided to enable business personnel to define process nodes, process paths and logical relationships through graphical operations (such as drag and drop); second, a rich process template library and industry general business components are built into the system to support users to customize and modify based on existing templates, thereby speeding up the modeling; third, with the help of rule engines or business rule configuration platforms, business logic is automatically recognized and process structure is dynamically generated; in addition, some technologies also combine natural language processing or knowledge graphs to support automatic construction of business process models through semantic understanding. The entire modeling generation process usually includes the following links: business requirement collection and analysis, modeling template matching, process node automatic layout, logical relationship configuration, visual process display, model verification and optimization, and model deployment and execution environment docking. Through the integration and optimization of the above links, the rapid modeling generation technology greatly improves the efficiency and accuracy of enterprise process online, reduces the dependence on professional developers, and realizes the "what you see is what you get" business process digital construction mode.

[0003] The existing technology has the following shortcomings:

[0004] In the process of merging multiple sources of enterprise business process templates for modeling, when there are the same conditional judgment nodes but inconsistent subsequent processing paths in different templates, the system usually calculates the weight according to the trigger frequency of each path in the historical process instances, and automatically generates the default branch path. Due to the different historical data collection caliber of each template, the statistical period, sample size and source system of the trigger frequency are inconsistent, resulting in that the path frequency under the same condition is in different dimension systems, and after direct merging, it cannot truly reflect the actual usage proportion of each path in the business. In this case, the existing rapid modeling generation technology for enterprise business processes cannot effectively process the frequency data of each path according to the normalized difference between the trigger frequencies of the conditional paths in the process of merging multiple templates to generate a process, so as to accurately generate the default branch path of the conditional node, which easily leads to the deviation of the default path from the actual main process behavior in the generated process model, affecting the accuracy of process execution and the rationality of business logic.

[0005] The above information disclosed in the BACKGROUND section merely to enhance the understanding of the background of the present disclosure, and therefore it can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0006] The object of the present application is to provide a fast modeling generation method for enterprise business processes to solve the problems in the background.

[0007] In order to achieve the above object, the present application provides the following technical solution: a fast modeling generation method for enterprise business processes, specifically comprising the following steps:

[0008] S1, after receiving a business process modeling request, performing structure similarity analysis, comparing the path vector of the current process structure input with a plurality of process templates, judging the structure coincidence degree and the condition node semantic consistency, and identifying whether it belongs to the case of generating a process by merging multiple templates;

[0009] S2, in the case of generating a process by merging multiple templates, extracting the path trigger frequency original value, frequency sampling time range, sample number, acquisition strategy label and path trigger distribution density corresponding to the condition judgment node, constructing a normalized offset vector, and outputting a normalized difference result;

[0010] S3, inputting the normalized difference result into the normalized transformation function family constructed by the deep residual network, combining the period compensation factor, the sampling variance adjustment factor and the distribution balance tensor, performing nonlinear normalization processing on the original frequency of each path, and generating a frequency vector sequence with unified dimension;

[0011] S4, based on the frequency vector sequence with unified dimension, fusing the path historical execution success rate, the upstream and downstream cooperative stability and the behavior deviation rate, generating a condition path default selection evaluation index, using a guided attention mechanism to score and sort each path, and forming a path selection sequence;

[0012] S5, according to the score value of the path selection sequence, performing dynamic setting of the default path through the segmented interval mapping control logic, and updating the path evaluation index through the process running feedback, realizing dynamic regulation and adaptive optimization of the default path.

[0013] Preferably, S1 specifically comprises the following steps:

[0014] S101, after receiving a business process modeling request, performing structure similarity analysis, comparing the path vector of the current process structure input with a plurality of process templates, judging the structure coincidence degree and the condition node semantic consistency, and identifying whether it belongs to the case of generating a process by merging multiple templates;

[0015] S102, performing structure similarity analysis, by inputting the path vector of the current process structure into the cosine similarity calculation and node field semantic vector comparison with the path vector in the plurality of process templates, generating a structure coincidence degree score set and a conditional node semantic consistency score set;

[0016] S103, calculating the number of process templates in the structure coincidence degree score set whose coincidence degree exceeds the first preset threshold, and calculating the number of conditional node pairs in the conditional node semantic consistency score set whose semantic similarity exceeds the second preset threshold, when both exceed the corresponding template merging judgment threshold, determining that the current process structure input belongs to the case of multiple template merging generation process.

[0017] Preferably, S102 specifically comprises:

[0018] The path vector of the current process structure input and the path vector in the plurality of process templates are normalized, the vector angle similarity between the two paths is obtained by cosine similarity calculation, and the result is stored in the structure coincidence degree score set;

[0019] The text content of the conditional judgment field in the current process structure input is input into the double-channel semantic embedding model to generate the node field semantic vector, and the Euclidean distance calculation is performed with the conditional node semantic vector in the plurality of process templates to construct the conditional node semantic consistency score set;

[0020] The structure coincidence degree score set and the conditional node semantic consistency score set are respectively statistically aggregated according to the template dimension to generate the structure coincidence degree score set and the conditional node semantic consistency score set.

[0021] Preferably, S2 specifically comprises the following steps:

[0022] S201, in the case of identifying multiple template merging generation processes, performing path association retrieval on each conditional judgment node, extracting the trigger frequency original value of the corresponding path based on the source template record, and simultaneously obtaining the frequency sampling time range, sample number, collection strategy label and path trigger distribution density corresponding to each path, and constructing a structured path frequency information set;

[0023] S202, time scale calibration is performed on the frequency sampling time range in the structured path frequency information set, and the trigger frequency original values statistically calculated in different time periods are periodically aligned; combined with the statistical caliber difference marked by the sample number and the collection strategy label, the standard frequency reference system is uniformly mapped, and based on the path trigger distribution density, a normalized offset vector reflecting the path sampling intensity and timing distribution offset is constructed;

[0024] S203, dimensionally expanding and aggregating the normalized offset vectors, calculating the deviation degree of each path in each normalized dimension, outputting the normalized difference result through a vector comparison function, and forming a set of normalized difference indicators representing the normalized difference between multiple paths in the statistical feature level.

[0025] Preferably, S203 specifically includes the following steps:

[0026] Dimensionally expanding the normalized offset vector corresponding to each path, respectively extracting four types of dimensions, i.e., dimension offset, time period offset, sample number weight offset, and trigger distribution density offset, and organizing the normalized features of each path into a set of multi-dimensional vectors with a unified structure;

[0027] Based on the multi-dimensional vector set, calculating the difference value between each type of offset dimension between all paths in pairs, and combining Mahalanobis distance and cosine angle measurement to construct a normalized offset degree matrix between path pairs, thereby realizing comprehensive comparison and quantization between different offset types;

[0028] Integrating the normalized offset degree matrix through a weighted aggregation strategy, setting aggregation weights combined with the statistical properties of the path context, outputting the normalized difference result, and forming a structured normalized difference indicator set, which is used as an input parameter in the path frequency normalization process and path evaluation mechanism.

[0029] Preferably, S3 specifically includes the following steps:

[0030] After encoding the normalized difference result into a multi-dimensional statistical feature matrix, inputting it into a family of normalization transformation functions constructed based on a deep residual network structure, each residual unit of the network is composed of a linear mapping layer, a normalization layer, and an activation mapping layer, and adopting cross-layer connection to maintain the integrity of difference information in the multi-level mapping process;

[0031] Constructing the period compensation factor, the sampling variance adjustment factor, and the distribution balance tensor into a normalization mapping weight matrix, respectively, and fusing and encoding them with the normalized difference result in the channel dimension, participating in the feature transformation process in the deep residual network through multi-dimensional conditional injection, so that the nonlinear normalization calculation can maintain the original difference expression ability of the path frequency while offsetting the structural offset caused by the statistical dimension;

[0032] Taking the normalized result output by the deep residual network as a nonlinear transformation function, mapping and transforming each original frequency value of each path item by item, obtaining a frequency vector sequence with a unified dimension, each frequency vector corresponding to a conditional path, maintaining structural consistency in the sequence to support subsequent path sorting and evaluation tasks.

[0033] Preferably, S4 specifically includes the following steps:

[0034] S401, the uniform dimension frequency vector sequence is normalized coded with the historical execution success rate of the corresponding path, the upstream and downstream cooperative stability and the behavior deviation rate respectively, and the features are spliced in the path dimension to construct a multi-dimensional evaluation vector set that integrates the path historical execution success rate, the upstream and downstream cooperative stability and the behavior deviation rate, which is used to form a path comprehensive representation;

[0035] S402, the path comprehensive representation is input into the path scoring network constructed by the guided attention mechanism, the key dimensions in the integrated features are dynamically weighted and calculated by setting the path level context attention prior in the guided attention mechanism, a conditional path default selection evaluation index of each path is generated, which represents the default recommended priority of the path in the current modeling scenario;

[0036] S403, according to the path score value contained in the conditional path default selection evaluation index, all paths are compared numerically and sorted from high to low according to the score value, and a path selection sequence is output, each position in the path selection sequence uniquely corresponds to a conditional path, and serves as the input basis for the default path configuration and execution control strategy.

[0037] Preferably, S402 is specifically:

[0038] Embedding each dimension feature in the path comprehensive representation into a unified tensor space, and constructing a local dependence graph of each path in the process structure through a context window, which is used to extract path level context attention prior information, representing the interaction dependence degree of different paths in the structure layer;

[0039] Based on the path level context attention prior information, a fusion type guided attention network structure is constructed, each key dimension in the path comprehensive representation is taken as the basic input of the attention weight generation, the dynamic weighting is applied to different dimensions through dot product attention calculation, and a high-order attention representation reflecting the multi-dimensional advantage features of the path is formed;

[0040] The high-order attention representation is output as a single channel path score value through a fully connected mapping layer, which is used as a conditional path default selection evaluation index, representing the default recommended priority of the path in the current modeling scenario according to the score value, and supporting the subsequent sequence use in the path sorting and strategy making task.

[0041] Preferably, S5 is specifically:

[0042] The score value of the path selection sequence is used to construct a segmented interval mapping control logic, the score value is divided into multiple priority segments according to the numerical interval, and each priority segment corresponds to a default path setting policy label; during the execution of the process instance, the position of the path score value associated with the current condition judgment node in the path selection sequence is matched with the corresponding interval segment according to the control logic, and the default path of the current instance of the node is set to realize the dynamic setting of the default path;

[0043] The real-time triggering frequency of each condition path during the running of the process, the path execution result and the cooperative change state of the upstream and downstream dependent nodes are collected, and the behavior deviation rate of the path in different time windows is statistically modeled to form a feedback feature data set corresponding to the path;

[0044] Based on the feedback feature data set, a path evaluation index update model is constructed, the process running feedback data is input into the model, an updated path evaluation index is generated, and the original index result is replaced, and the update result is synchronized to the path score network and the path selection sequence, a closed-loop feedback channel is constructed, and dynamic regulation and adaptive optimization of the default path are realized.

[0045] In the above technical solution, the technical effects and advantages provided by the present application are as follows:

[0046] 1、The present application realizes accurate identification and effective fusion of modeling scenarios with structural overlap but path configuration difference in multiple source business process templates by constructing a multi-template process merging judgment mechanism and a path frequency normalization process. By introducing a structure coincidence degree score and a condition node semantic consistency double analysis mechanism, it can be judged whether the process structure input is derived from the combination of multiple templates, and the logical conflict of the process caused by misjudgment is avoided. At the same time, in view of the inconsistency of statistical dimensions such as frequency collection period, sample quantity and collection strategy between different templates, a normalization offset vector construction method is designed, combined with a normalization transformation function family constructed by a deep residual network, to perform nonlinear normalization mapping on the original path frequency, unify the dimension system, ensure the data basis of the subsequent path evaluation link to be comparable and consistent, and significantly improve the accuracy and rationality of process modeling.

[0047] 2、The application fuses multi-dimensional evaluation indexes such as historical execution success rate, upstream and downstream coordination stability and behavior deviation rate, constructs a path score network through a guided attention mechanism, realizes efficient generation of conditional path default selection evaluation indexes while maintaining the ability of context semantic understanding, can dynamically sort and prioritize the paths under each conditional node, and improves the rationality and business fit of the default path setting. In the actual operation process of the flow, the system can also construct a path evaluation index update model according to real-time feedback, form a closed-loop regulation and control mechanism from modeling, execution to feedback, have the ability of continuous learning and self-adaptive optimization, solve the problems of static rule solidification and lack of feedback update in traditional flow modeling, and thus realize a more efficient, more intelligent and more accurate enterprise flow modeling generation system. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0049] Figure 1 The flowchart of the rapid modeling generation method for enterprise business process of the present application. DETAILED DESCRIPTION

[0050] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations may, however, be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these example implementations are provided so that this disclosure will be thorough and complete, and will fully convey the inventive aspects of the example implementations to those skilled in the art.

[0051] The present application provides a rapid modeling generation method for enterprise business process as shown in Figure 1 The specific steps of the rapid modeling generation method for enterprise business process of the present application include the following steps:

[0052] S1, after receiving the business process modeling request, performing structure similarity analysis, comparing the path vector of the current process structure input with a plurality of process templates, judging the structure coincidence degree and the conditional node semantic consistency, and identifying whether it belongs to the case of generating a process by merging a plurality of templates;

[0053] In this embodiment, S1 specifically includes the following steps:

[0054] S101, after receiving the business process modeling request, uniformly vector encoding the node order, path connection relationship and conditional judgment field content in the process structure input, generating the path vector of the current process structure input, and performing structure similarity analysis;

[0055] After receiving the service process modeling request, the core components in the process structure input need to be extracted first, including node order, path connection relationship and condition judgment field content. The node order refers to the arrangement order of each business activity node in the logical execution chain in the process, the path connection relationship represents the connection mode and flow path between the nodes, and the condition judgment field content is used to describe the judgment logic of the decision node, such as judgment amount, approval level, etc. In order to realize unified vector coding, the above three types of information need to be structured converted. First, the node order is converted to sequence coding through topological sorting; second, the path connection relationship is represented by the adjacency matrix of the directed edge between the nodes, and then the structure representation vector is extracted by using the graph neural network encoder; finally, the condition judgment field content is semantically embedded by using the word vector model to generate the semantic representation of each judgment field. The three types of vectors are connected or fused in a unified dimensional space to generate the path vector representing the current process structure input, which is used for subsequent structure similarity analysis. The path vector can comprehensively reflect the execution logic structure and decision semantic features of the process, ensuring good consistency and matching basis for the structure comparison between multiple process templates.

[0056] S102, performing structure similarity analysis, by calculating the cosine similarity of the path vector of the current process structure input and the path vector in the multiple process templates, and comparing the node field semantic vectors, a structure coincidence degree score set and a condition node semantic consistency score set are generated;

[0057] S103, calculating the number of process templates in the structure coincidence degree score set whose coincidence degree exceeds the first preset threshold, and calculating the number of condition node pairs in the condition node semantic consistency score set whose semantic similarity exceeds the second preset threshold, when both exceed the corresponding template merging judgment threshold, it is determined that the current process structure input belongs to the case of multiple template merging generation process.

[0058] After the aggregation of the structure coincidence score set and the condition node semantic consistency score set is completed, it is necessary to judge whether the current process structure input is derived from the fusion of multiple process templates based on the statistical results. The specific implementation manner is as follows: first, in the structure coincidence score set, all process templates compared with the current process structure input are searched, and the number of templates with a score higher than a first preset threshold is counted, which is used to evaluate how many templates the current process structure significantly coincides with in the structure layer; then, in the condition node semantic consistency score set, the number of condition node pairs with a semantic similarity higher than a second preset threshold is calculated, which is used to measure whether the current process has the cross performance of multiple template semantic characteristics in the judgment logic. When both statistical results exceed the respective defined template merging judgment threshold, it indicates that the current process structure not only coincides with multiple templates in the path structure, but also has multiple consistencies in the business semantics of the decision node, which meets the judgment standard of process fusion. Through the judgment mechanism, false judgments caused by single structure or semantic coincidence can be effectively avoided, and the accuracy and business representativeness of template fusion identification can be ensured.

[0059] In the process judgment mechanism, the first preset threshold is a score limit set for the structure coincidence score, which is usually determined based on the high section of the path vector cosine similarity statistical distribution, to ensure that the templates counted indeed have high structure coincidence; the second preset threshold is a semantic distance upper limit set for the condition node semantic consistency score, and its value reflects the similar lower limit requirement between node semantics, which is usually set by parameter tuning according to the labeled data set of the semantic embedding model; the template merging judgment threshold is a set of composite condition thresholds, which is used to jointly constrain the statistical results of the structure coincidence and semantic consistency two dimensions to prevent false judgments as template fusion situations when only one of the two dimensions is met. The three thresholds are set and optimized through cross-validation in the model training process, which can dynamically adapt to the process modeling needs in different enterprise scenarios and improve the universality and precision of the merging judgment.

[0060] In this embodiment, S102 specifically comprises:

[0061] The path vectors of the current process structure input and the path vectors in the multiple process templates are normalized, the vector angle similarity between two paths is calculated through cosine similarity, and the result is stored in the structure coincidence score set;

[0062] Before comparing the path vector of the current process structure input with the path vectors in the plurality of process templates, first, all vectors need to be normalized to make them comparable in the same dimensional space. The purpose of normalization is to eliminate the interference of different path vectors due to different dimension distributions or numerical ranges, so that the similarity comparison of each path vector in the unit vector space is more accurate. After normalization, the cosine similarity calculation method is used to compare the path vector of the current process structure input with the path vector in each process template one by one. Cosine similarity is mainly used to measure the angle between two vectors. The smaller the angle, the higher the similarity. For example, the current process path vector represents a purchase approval process, and the encoded vector is highly consistent with the path vector of a certain template in multiple node dimensions. Only the end approval node has structural differences, so the cosine similarity value will be close to 1, indicating that the two structures are highly consistent. On the contrary, if another template path vector represents a stock allocation process, its node structure is quite different from the purchase process, and the similarity value is significantly lower. All the similarity values calculated will be stored in the structure coincidence degree score set in turn. This set is used to measure the similarity between the current process structure and each template in the path structure layer, and to provide a quantitative basis for subsequent structure judgment. This process ensures the accuracy and objectivity of the structure comparison in the process modeling process, and helps to achieve high-quality multi-template structure fusion.

[0063] The text content of the condition judgment field in the current process structure input is input into the double-channel semantic embedding model to generate a node field semantic vector, and the Euclidean distance is calculated with the condition node semantic vectors in the plurality of process templates to construct a condition node semantic consistency score set;

[0064] In processing the condition judgment field in the current process structure input, the text content of each judgment field needs to be semantically encoded first. The encoding structure adopted is a two-channel semantic embedding model, which consists of a context encoding channel and a field attribute encoding channel. The context encoding channel is used to capture the context meaning of the field in the complete judgment statement, for example, in "whether the approval amount is greater than 10,000 yuan", the "amount" is constrained by the business semantics of "approval"; the field attribute encoding channel is used to model the position, decision weight and historical usage frequency of the field in the business process and other attribute characteristics. The vectors generated by the two channels will be fused in the embedding space to form the final node field semantic vector. This vector will be used to calculate the Euclidean distance between the corresponding condition nodes in multiple process templates, which is used to evaluate the closeness of the semantics. For example, the current judgment field is "the payment amount is greater than 5,000 yuan", and the corresponding judgment field in a process template is "the payment amount exceeds 5,000", although the words are different, but in the two-channel semantic embedding model, the semantic vectors are similar due to the similarity of the context and field attributes, the Euclidean distance is very close, which shows that the semantics are highly consistent; on the contrary, if the template field is "the inventory quantity is less than 100", its semantic vector will be quite different from "the payment amount", the corresponding Euclidean distance is larger, indicating that the semantics are not consistent. All the calculated Euclidean distance values will form a set of condition node semantic consistency scores, which are used for subsequent analysis of semantic matching relationship between process structures to ensure the semantic accuracy and consistency in the structure fusion process.

[0065] The structure coincidence degree score set and the condition node semantic consistency score set are respectively statistically aggregated according to the template dimension to generate the structure coincidence degree score set and the condition node semantic consistency score set.

[0066] After the similarity calculation between the path structure and the semantic vector is completed, all scoring data needs to be aggregated according to the dimensions of the process template, so as to generate a structure coincidence degree score set and a condition node semantic consistency score set that can be used for template merging judgment. First, for the structure coincidence degree score set, the similarity scores of all path pairs between the current process structure input and each process template are traversed, the path similarity values belonging to the same template are gathered together, and the mean value, maximum value and distribution standard deviation and other statistical characteristics are calculated to generate the structure coincidence degree score record corresponding to the template. The score record not only reflects the overall structure similarity level, but also reveals the significance degree of local path coincidence. Secondly, for the condition node semantic consistency score set, the Euclidean distance scores between each condition judgment field in the current process structure input and all condition nodes in each template are aggregated, a group of node pairs with close semantic distance are selected, and the semantic similarity scores of these node pairs are aggregated in the template dimension, and the mean value, minimum value and semantic concentration degree and other statistical characteristics are generated. For example, if there are five paths in a template that are highly coincident with the current process structure, and three condition nodes are close in semantics to the nodes in the current process, the template will be given a higher structure coincidence degree score and semantic consistency score in the two score sets. This aggregation process effectively enhances the comparability of the scores at the template level, ensures that the subsequent template merging judgment has sufficient structure and semantic basis, and at the same time avoids the risk of misleading the overall judgment result by a single high score.

[0067] S2, in the case of identifying multiple template merging generation processes, extracting the path trigger frequency original value corresponding to each condition judgment node, the frequency sampling time range, the sample number, the acquisition strategy label and the path trigger distribution density, constructing a normalized offset vector, and outputting a normalized difference result;

[0068] In this embodiment, S2 specifically includes the following steps:

[0069] S201, in the case of identifying multiple template merging generation processes, performing path association retrieval on each condition judgment node, extracting the trigger frequency original value of the corresponding path based on the source template record, and simultaneously acquiring the frequency sampling time range, the sample number, the acquisition strategy label and the path trigger distribution density corresponding to each path, and constructing a structured path frequency information set;

[0070] In the case of identifying multiple template merging generation processes, it is necessary to locate each path connected to the conditional judgment node from all the fused process templates, and to establish a correspondence between these paths and their historical source templates. By traversing the template data structure, the path association mapping table is constructed according to the node identifier and the path direction information, and then the trigger frequency original value recorded in the historical process instance of each path is extracted. Subsequently, each path also needs to be dimensionally supplemented, respectively collecting its corresponding frequency sampling time range, the number of samples participating in the statistics, the collection strategy label adopted and the distribution density of path triggering on the time axis. The above multi-dimensional statistical information is uniformly packaged as a structured path frequency information set for the next normalization calculation. The purpose of this is to ensure that before normalization, the frequency data of all paths has traceability of source and quantifiability of statistical properties, thereby providing sufficient data basis for dimension elimination and bias evaluation between different source paths.

[0071] The "conditional judgment node" refers to a judgment type logic node in the process with multiple branch outputs, such as nodes that branch according to fields such as amount, region, and approval role; "path association retrieval" is to establish a one-to-one mapping between the conditional node and all subsequent paths through directed graph analysis of process structure information, for path-level frequency information extraction; "source template record" refers to the running data collected during modeling or execution of the original process template, including the number of times each path is triggered in actual business; "frequency sampling time range" represents the statistical period of the path frequency data, which determines the time representativeness of the frequency value; "sample number" refers to the number of valid process instances participating in the calculation of path triggering frequency statistics, which affects the stability and confidence of the frequency data; "collection strategy label" is used to describe the collection method of the data and its applicable scenario, such as whether it is continuous sampling, whether it is collected after excluding abnormal data, etc.; "path triggering distribution density" is used to measure the concentration degree of path triggering in different time periods, such as uniform triggering or peak concentrated triggering, which determines the time stability of path behavior; "structured path frequency information set" is the packaging of all the above multi-dimensional information, forming a data structure convenient for vectorization processing, providing basic input for the construction of subsequent normalization bias.

[0072] S202, time scale calibration is performed on the frequency sampling time range in the structured path frequency information set, the trigger frequency original values statistically in different time periods are periodically aligned; combined with the statistical caliber differences marked by the sample number and the collection strategy label, the trigger frequency original values are uniformly mapped to a standard frequency reference system, and a normalization bias vector reflecting the sampling intensity and time sequence distribution bias of the path is constructed based on the path triggering distribution density;

[0073] To ensure the comparability of frequency data from different source paths, the frequency sampling time range of each path in the structured path frequency information set needs to be calibrated first. The specific implementation is to uniformly convert the statistical period of the original frequency data, recalculate the frequency value according to the standard time granularity (such as week, ten or month), and ensure the alignment of frequency data under different time windows. Then, according to the difference between the sample size and the statistical caliber marked in the collection strategy label, a standard frequency reference system is designed to map the frequency values of each path to a unified standard. In this process, the weight adjustment factor considering the sample sufficiency and data stability is used to control the deviation caused by the inconsistency of sample size. Then, the time distribution characteristics represented by the path trigger distribution density are used to construct a feature vector that can describe the path sampling intensity and timing offset. Finally, combined with the alignment results and the feature vector, a normalized offset vector is generated through a feature fusion mechanism to quantify the differences of each path in the statistical dimension. The purpose of the above processing is to realize the unified expression of cross-source, cross-period and cross-caliber data, and provide a high-consistency data basis for subsequent path frequency normalization and path evaluation index generation.

[0074] The "frequency sampling time range" refers to the time period covered by the frequency data of each path, which determines the periodicity and timeliness of the frequency value; the "time scale calibration" is to adjust the frequency value of different statistical time granularity to a unified unit time period, to solve the problem of non-comparison of frequency between paths; the "period alignment" is the re-slicing processing of the original frequency data, which makes it have a unified time window feature; the "sample size" is used to measure the stability of frequency data statistics, the more samples, the more stable the result; the "collection strategy label" describes the contextual attributes of data collection method, such as whether to use event trigger, whether to filter abnormal behavior or whether to intercept based on log system, reflecting the reliability of frequency data; the "statistical caliber difference" refers to the inconsistency of frequency value calculation method caused by different collection strategies, which must be mapped into the standard frequency system through rules to eliminate structural errors; the "standard frequency reference system" is a frequency expression system built on unified caliber and unified time scale, which is used to accept all path frequency data; the "path trigger distribution density" reflects the time concentration of path trigger behavior in the statistical period, and identifies whether the behavior is concentrated in a specific period or cycle; the "normalized offset vector" is a feature vector formed after the fusion and normalization of path frequency data in multiple dimensions of dimension, time, sample and behavior density, which is used to support subsequent path normalization transformation and evaluation modeling.

[0075] S203, the normalized offset vector is processed by dimension expansion and aggregation, the deviation degree of each path in each normalized dimension is calculated, the normalized difference result is output through a vector comparison function, and a normalized difference index set representing multiple paths in the statistical feature level is formed.

[0076] In the scenario of merging multiple process templates, in order to identify the differences between different paths in the triggering frequency statistical characteristics, it is necessary to first perform dimension expansion operation on the normalized offset vector, and decompose the composite dimension in the original normalized offset vector into multiple independent dimension variables, such as dimension offset, time period offset, sample number weight offset and trigger distribution density offset. Then, the multi-dimensional offset index set of each path is aggregated to construct a unified multi-dimensional path feature representation, ensuring that all paths are comparable in the same dimension space. Based on these unified structure feature vectors, the difference between each pair of paths is calculated, and the difference in each dimension is quantified by using a vector comparison function, such as measuring the separability of statistical feature distribution by Mahalanobis distance and reflecting the consistency of offset direction by cosine angle. The difference result output by the comparison function will form a normalized difference score set, which fully expresses the deviation intensity of the path in the statistical dimension. Finally, a quantifiable normalized difference index set of multiple paths in the statistical feature level is generated by aggregation processing, which provides data support for subsequent path frequency adjustment and default path selection based on difference quantity, and ensures the rationality, accuracy and executability of the process model in the merging scenario.

[0077] In this embodiment, S203 is specifically:

[0078] The normalized offset vector corresponding to each path is dimensionally expanded, and four types of dimensions, dimension offset, time period offset, sample number weight offset and trigger distribution density offset, are extracted respectively, and the normalized features of each path are organized into a set of multi-dimensional vectors with unified structure.

[0079] In order to realize the unified comparison of multiple path frequency data in the dimension of statistical characteristics, the dimension expansion operation is needed for the normalized offset vector corresponding to each path. The specific implementation is: first, the various information contained in the normalized offset vector is parsed according to the pre-defined dimension, and the dimension offset reflecting the inconsistency of frequency units, the time period offset reflecting the difference of statistical period, the sample number weight offset reflecting the difference of sample number, and the trigger distribution density offset reflecting the time sequence difference of path behavior distribution are extracted respectively. Each type of offset value is converted into a calculable standard vector segment through a numerical mapping strategy, and is marked through a dimension label. Subsequently, the four types of offset vectors are spliced according to the unified structure rule to form a multi-dimensional vector set with fixed vector sequence, consistent field label and unified numerical interval, so as to support the subsequent calculation of path deviation degree and difference quantization analysis. The core goal of the operation is to decompose the normalized features of multiple source path frequency data into structured, multi-dimensional and comparable mathematical expression formats, so as to support the unified measurement and difference modeling of different paths in the statistical attribute level, and improve the accuracy and algorithm stability of path merging modeling. Through dimension expansion, different normalized influence factors are explicitly separated, providing strong structural support and semantic clarity for the aggregation and comparison of normalized offset degree in the next step.

[0080] Based on the multi-dimensional vector set, the pair-wise calculation difference of each type of offset dimension between all paths is performed, and the Mahalanobis distance and cosine angle measure are combined to construct the normalized offset degree matrix between path pairs, so as to realize the comprehensive comparison and quantization between different offset types;

[0081] To realize the difference quantification of different paths in multiple types of normalized dimensions, it is necessary to perform pairwise calculation on each type of offset dimension between paths based on the set of multi-dimensional vectors. First, for the four types of dimensions, dimension offset, time period offset, sample number weight offset, and trigger distribution density offset, the numerical difference in each dimension is calculated according to the two-by-two combination of paths. For example, for path A and path B, in the sample number weight offset dimension, if the normalized value of path A is 0.8 and the normalized value of path B is 0.5, then the difference is 0.3. To further measure the statistical significance of these differences in the overall offset space, the Mahalanobis distance is applied to calculate the covariance standardized distance between different path vectors, thereby revealing the relative outlying degree of the offset in the overall feature distribution. At the same time, the cosine angle measure is introduced to capture the consistency of the direction of different path vectors, representing the comprehensive deviation direction of each type of offset under weight combination. After fusing the distance value and the angle similarity index of each pair of paths in the four dimensions, a multi-dimensional structure of path pair normalization offset degree matrix is formed. For example, path A and path B, path A and path C correspond to different matrix units, representing the offset degree and similarity strength in each dimension. The significance of constructing this matrix lies in that it can not only identify the overall difference level between paths, but also analyze the dominant factors of different normalization dimensions in the offset, providing fine-grained and structured input support for subsequent weight evaluation and path selection ranking, ensuring the objectivity and consistency of process modeling results in the scenario of diversified data sources.

[0082] The normalized offset degree matrix is integrated through a weighted aggregation strategy, the aggregation weight is set in combination with the statistical characteristics of the path context, the normalized difference result is output, and a structured normalized difference index set is formed, which is used as an input parameter in the path frequency normalization processing and path evaluation mechanism.

[0083] To effectively integrate the normalized deviation degree matrix between path pairs into a set of indicators that can be used for subsequent analysis, a set of weighted aggregation strategies need to be constructed to ensure that the contribution of each deviation dimension to the overall difference can be reasonably measured. First, based on the deviation values of each path pair in the normalized deviation degree matrix in different dimensions, the statistical distribution characteristics of each dimension in the full path network are extracted, such as deviation mean, deviation variance, and deviation density distribution. Then, combined with the statistical characteristics information in the path context, including the frequency of the path in the original template, the average stay time, the importance of business decision, etc. Data indicators are set for different deviation dimensions to give deviation indicators business awareness ability. For example, for a high-frequency path pair, if its dimensional deviation is small and the time period deviation is significant, the aggregation weight of the time period deviation can be increased, thereby amplifying the difference expression of the path pair in time consistency. Then, by performing vector dot product on the multi-dimensional deviation values of each path pair and the corresponding aggregation weight, a set of quantized difference scores is output, forming a structured normalized difference result set. This set not only retains the detailed levels of the normalized differences between paths, but also realizes the adjustment of the difference degree based on the context semantics, providing accurate and dynamic input basis for subsequent frequency data normalization processing and path selection strategy, effectively improving the accuracy and business logic adaptation degree of process modeling output in the multi-template fusion modeling scenario.

[0084] S3, input the normalized difference result into the normalized transformation function family constructed by the deep residual network, combine the period compensation factor, the sampling variance adjustment factor and the distribution balance tensor, and perform nonlinear normalization processing on the original frequency of each path to generate a uniform dimensional frequency vector sequence;

[0085] In this embodiment, S3 is specifically:

[0086] After the normalized difference result is encoded into a multi-dimensional statistical feature matrix, it is input into the normalized transformation function family constructed based on the structure of the deep residual network. Each residual unit of the network is composed of a linear mapping layer, a normalization layer and an activation mapping layer, and uses cross-layer connection to maintain the integrity of the difference information in the multi-level mapping process.

[0087] In the process of constructing the normalized transformation function family, it is necessary to first encode the normalized difference results into a structured multi-dimensional statistical feature matrix, which contains the offset degree information of each path in different normalization dimensions. To enhance the non-linear expression ability of feature extraction, the matrix is input into the normalized transformation function family constructed based on the deep residual network structure. The network performs deep information fusion processing through multi-layer residual stacking mechanism, and uses cross-layer connection structure to perform residual transmission on the original difference features, effectively preventing gradient dissipation and information loss caused by deep non-linear mapping. This method can preserve the original feature expression of the normalized offset while extracting multi-level features, realize deep modeling of complex offset characteristics, and is beneficial to the subsequent precise non-linear normalization processing of path frequency.

[0088] When the multi-dimensional statistical feature matrix is input into the network, the numerical characteristics of each path in the normalized difference results in multiple dimensions such as dimensional offset, period difference, sampling bias and distribution density are preserved. In the deep residual network, each residual unit is composed of three types of key structures: first, the linear mapping layer is used to preliminarily compress the input dimension and extract the significant feature mode between paths; second, the normalization layer is used to standardize the path feature distribution to avoid gradient instability caused by feature drift; third, the activation mapping layer uses a non-linear function to enhance the feature discrimination and improve the network's ability to express complex patterns. The cross-layer connection between each layer realizes the direct addition of input and output, avoiding the over-deformation of input features by deep structure. This structure balances the information preservation and non-linear expression ability of the entire normalized transformation function family, providing reliable support for frequency normalization modeling.

[0089] The period compensation factor, the sampling variance adjustment factor and the distribution balance tensor are respectively constructed as the normalized mapping weight matrix, and are fused and encoded with the normalized difference results in the channel dimension, participating in the feature transformation process in the deep residual network through multi-dimensional conditional injection, so that the non-linear normalization calculation can offset the structural offset caused by statistical dimensions while preserving the original difference expression ability of path frequency;

[0090] To improve the accuracy and adaptability of frequency normalization processing under the multi-source process template, the period compensation factor, the sampling variance adjustment factor and the distribution balance tensor are respectively constructed as the normalization mapping weight matrix. These weight matrices are used to dynamically adjust the feature response intensity in the normalization transformation process. These matrices are fused and encoded in the channel dimension and the normalization difference result, that is, the original offset information and the correction adjustment parameter are embedded in the multi-dimensional input channel, thereby forming a complete feature input set. Through multi-dimensional conditional injection, the fused input is embedded into each layer feature transformation process in the deep residual network, so that each mapping can perceive the interference intensity of the statistical conditions suffered by the path frequency, and realize the hedging and compensation of structural offset. This processing method can reduce the normalization distortion risk caused by inconsistent data statistical dimensions while ensuring the original difference expression ability of path frequency.

[0091] The period compensation factor is used to measure the time distribution offset caused by different sampling periods of each path frequency data, and the factor is constructed as a standardized period mapping matrix to correct the systematic deviation in the time dimension. The sampling variance adjustment factor reflects the stability of the path frequency statistical sample, and by constructing a standard deviation adjustment matrix consistent with the path dimension, the path feature values with high volatility are weighted and inhibited. The distribution balance tensor is used to identify the unevenness of the distribution of each path in the statistical space, and to generate a high-order tensor structure reflecting the sparsity and local density. In the fusion encoding process, the three matrices are merged with the normalization difference matrix in the form of additional channels, and through the multi-dimensional conditional injection mechanism, each network residual block considers the relationship between the original frequency difference and the compensation factor when performing mapping calculation, ensuring that the output frequency vector can achieve high-fidelity and structure-consistent nonlinear normalization expression.

[0092] The normalization result output by the deep residual network is used as a nonlinear transformation function to map and transform each path's original frequency value item by item, obtaining a sequence of frequency vectors with uniform dimensions, each frequency vector corresponding to a conditional path, maintaining structural consistency in the sequence to support subsequent path ordering and evaluation tasks.

[0093] In order to realize the unified expression of the conditional path frequency in different source templates, the output result of the deep residual network is taken as a nonlinear transformation function, and the original frequency value of each path is mapped item by item. This transformation process not only corrects the original frequency value, but also fuses the period compensation factor, the sampling variance adjustment factor and the multi-dimensional correction factor introduced by the distribution balance tensor, so that the frequency data under different sources and different statistical indicators can be mapped to a unified expression space. In specific implementation, the trained residual network transformation function is applied to the original frequency input of each path, and the normalized frequency vector is output to form a vector sequence of multiple conditional path frequencies. The sequence keeps the consistency of the path structure while eliminating the dimension inconsistency caused by the difference in sampling conditions, thereby providing a unified and comparable quantitative basis for subsequent path ordering, weight fusion and default path evaluation.

[0094] The deep residual network outputs a high-dimensional nonlinear mapping result, and each output corresponds to the frequency normalization value of a specific conditional path under multiple offset dimension adjustments. These normalization values no longer directly represent the original number of times the path is triggered, but are normalized expressions after being modified and calculated after fusing the time period, sampling density and distribution structure. When they are used for item-by-item mapping of the original frequency, each path frequency is adjusted through a nonlinear function, which contains the original offset information channel maintained by the residual connection structure, so that even if the frequency span difference is huge, the mapping can still fall into a unified frequency scale. The final frequency vector sequence is arranged in the order of the conditional nodes, each item represents the normalized frequency of a path, and the structure dimension is strictly aligned with the process template, avoiding path order dislocation or data index mismatch, thereby constructing a standardized input space required for path ordering and evaluation.

[0095] S4, based on the frequency vector sequence of the unified dimension, fusing the historical execution success rate of the path, the upstream and downstream cooperative stability and the behavior deviation rate, generating a conditional path default selection evaluation index, using a guided attention mechanism to score and sort each path to form a path selection sequence;

[0096] In this embodiment, S4 specifically includes the following steps:

[0097] S401, the frequency vector sequence of the unified dimension is normalized and coded with the historical execution success rate of the corresponding path, the upstream and downstream cooperative stability and the behavior deviation rate, respectively, and the features are spliced in the path dimension to construct a multi-dimensional evaluation vector set fusing the historical execution success rate of the path, the upstream and downstream cooperative stability and the behavior deviation rate, which is used to form a path comprehensive representation;

[0098] When the uniform dimension frequency vector sequence is fused with the path history execution success rate, the upstream and downstream collaborative stability, and the behavior deviation rate, the four types of index data need to be normalized first. The frequency vector sequence has been unified in dimension and can be directly involved in subsequent feature splicing. The path history execution success rate is calculated by counting the actual execution completion times and the total trigger times of each path in the past process instances, calculating the stable output proportion, and using the minimum-maximum scaling method to map to the standardized interval. The upstream and downstream collaborative stability is evaluated according to the execution time deviation and behavior consistency rate of the nodes at both ends of the path and the adjacent nodes, and the structure dependence score is generated and normalized. The behavior deviation rate is constructed by comparing the deviation degree of the current execution characteristics of the path and the historical behavior template to reflect the stability of the path execution behavior. After normalization, the relative expression ability of the path in the overall feature space is retained. The four types of normalized features are spliced according to the path dimension to form a multi-dimensional evaluation vector, which constitutes a complete path comprehensive representation.

[0099] The construction process of the above path comprehensive representation ensures that all input features are involved in subsequent calculations at the same order of magnitude and vector structure, avoiding the dominant influence of value differences in a certain dimension on the model results. The uniform dimension processing of the frequency vector ensures the comparability of the path activity, and the history execution success rate reflects the actual feasibility and preference of the path. The upstream and downstream collaborative stability provides a reference for the logical stability of the process structure, and the behavior deviation rate is used to measure the consistency risk of the path behavior. The multi-dimensional vector set after feature splicing not only covers the structure, behavior and statistical three types of core indicators in dimension, but also has a consistent embedded format at the path level, providing a structured input basis for the subsequent fine evaluation and path scoring of the guided attention mechanism.

[0100] S402, input the path comprehensive representation into the path scoring network constructed by the guided attention mechanism, set the path-level context attention prior, guide the attention mechanism to dynamically weight the key dimensions in the fused features, and generate a conditional path default selection evaluation index for each path, representing the default recommended priority of the path in the current modeling scenario;

[0101] S403, according to the path score value contained in the conditional path default selection evaluation index, numerically compare all paths and sort them from high to low according to the score value, output the path selection sequence, and each position in the path selection sequence uniquely corresponds to a conditional path, which is used as the input basis for the default path configuration and execution control strategy.

[0102] In the path decision generation process, each conditional path has been assigned a path score value at the model construction, which is derived from the dynamic weighting expression of multi-dimensional features by the fusion guide attention mechanism, reflecting the comprehensive priority of the path in the current business process modeling environment. In order to form a path selection sequence with clear ordering relationship, the score values of all paths need to be compared and processed uniformly. In specific operation, the score values are sorted in descending order after being composed into a vector set, and the sorting result is the path selection sequence. In this sequence, each index position corresponds to a specific conditional path, and the mutual exclusivity is maintained to ensure that there is no path repetition or score conflict.

[0103] The output result of score sorting not only defines the default recommended order of paths in structure, but also has practical application value in strategy execution. The system will set the default flow path of each conditional judgment node in the process based on the path selection sequence, and automatically refer to the highest priority path in the sorting result as the default decision path when the execution engine triggers the judgment node. At the same time, this sorting can also be used as the input basis for the configuration module and feedback optimization mechanism to realize the adjustability and adaptability of the process decision logic. The comparison, sorting and unique mapping mechanism of the score value ensures the stability, rationality and controllability of the path selection, providing strong support for the enterprise business process modeling task.

[0104] In this embodiment, S402 is specifically:

[0105] Embedding each dimension feature in the path comprehensive representation into a unified tensor space, and constructing a local dependency graph of each path in the process structure through a context window, which is used to extract path-level context attention prior information and represent the interactive dependency degree of different paths at the structure level;

[0106] In the process of synthesizing the dimensional characteristics in the path representation, the frequency vector, execution success rate, cooperative stability and behavior deviation rate corresponding to each path need to be embedded into a unified tensor space through linear transformation first to ensure dimensional compatibility and semantic alignment in subsequent calculations. After embedding, for the context position of each path in the enterprise business process structure, the connection relationship and adjacency between the path and its predecessor, successor and concurrent branch are scanned using the context window mechanism. The local dependency graph is constructed by path connection topology, trigger timing and co-occurrence frequency. The graph takes paths as nodes and dependency relationships as edges, and the edge weight represents the structural coupling degree between paths. In specific calculations, the co-trigger probability and time offset cooperation degree indicators of the path in multiple process instances are referred to. Through the graph, the context attention prior information of each path in its local structure range can be effectively extracted, representing the interaction dependency strength of the path in the overall structure of the business process. This design not only supplements the interaction information of path features at the process structure level, but also provides a structure-driven dynamic weighting basis for subsequent guided attention mechanisms, thereby achieving a path priority evaluation with more context awareness.

[0107] Based on the path-level context attention prior information, a fusion guided attention network structure is constructed, which takes each key dimension in the path comprehensive representation as the basis input for attention weight generation. Through dot product attention calculation, dynamic weighting is applied to different dimensions to form a high-order attention representation reflecting the multi-dimensional advantage features of the path.

[0108] To realize intelligent evaluation of path priority, a fusion guided attention network structure is first constructed, in which the path-level context attention prior information is embedded into the attention generation module of the network as a guide signal. The guided attention network receives the path comprehensive representation corresponding to each path as input, and embeds each key dimension in the path comprehensive representation, including the unified dimension frequency vector, historical execution success rate, upstream and downstream cooperative stability and behavior deviation rate, to form a feature matrix. Subsequently, based on the context dependency relationship of the path, the dot product attention mechanism is applied to the feature matrix to calculate the weight of each feature dimension for the importance of the current path in the modeling scenario, dynamically adjust the representation strength of each feature dimension, and finally output the aggregated high-order attention representation to reflect the comprehensive advantage of the path under multiple evaluation indicators.

[0109] In this process, the key dimensions in the path synthesis representation play a core role in the integration of multi-source data. The unified dimension frequency vector is used to describe the breadth of path use under different template sources, the historical execution success rate is used to reflect the stable reliability of the path in business operation, the upstream and downstream coordination stability is used to measure the execution continuity and compatibility of the path with the front and rear nodes, and the behavior deviation rate is used to depict the behavior difference or abnormal trend in the path execution. The fusion of these characteristic dimensions not only has independent business meaning, but also can form a balanced expression effect through weight self-adaptive adjustment in the attention mechanism. Finally, through the weighted aggregation of the high-order attention mechanism, a path priority representation with context awareness ability, structure dependence characteristics and multi-index fusion expression ability can be generated, which lays an accurate basis for default path recommendation and subsequent sorting strategy.

[0110] The high-order attention representation is output as a single-channel path score value through a fully connected mapping layer, which is used as a conditional path default selection evaluation index. The score value represents the default recommended priority of the path in the current modeling context, and supports the serialization in the subsequent path sorting and strategy formulation tasks.

[0111] In the path priority evaluation process, the high-order attention representation carries the comprehensive information of multi-dimensional fusion features such as frequency, success rate, coordination stability and behavior deviation. In order to convert these high-dimensional embedding representations into score results with sorting ability, a fully connected mapping layer is needed to reduce the dimension of the high-order attention representation. In specific implementation, the high-order attention vector corresponding to each path is first input into the parameterized fully connected neural network, and through the combination of linear transformation and nonlinear activation function, the multi-dimensional features are compressed into a single-dimensional scalar output. The scalar is the score value of the path, which is used to represent the recommended priority of the path as the default selection path in the current modeling context. The score value not only has quantifiable sorting characteristics, but also maintains the sensitivity to context structure and the expression ability of feature fusion, and can be directly used as a conditional path default selection evaluation index for the generation of subsequent path selection sequence and the configuration input of process execution control strategy, thereby supporting intelligent path recommendation and dynamic adaptation of decision nodes in the process model.

[0112] S5, according to the score value of the path selection sequence, the default path is dynamically set through the segmented interval mapping control logic, and the path evaluation index is updated through the process running feedback, realizing the dynamic regulation and adaptive optimization of the default path.

[0113] In this embodiment, S5 is specifically:

[0114] The score value of the path selection sequence is used to construct a segmented interval mapping control logic, the score value is divided into multiple priority segments according to the numerical interval, and each priority segment corresponds to a default path setting policy label; during the execution of the process instance, the position of the path score value associated with the current condition judgment node in the path selection sequence is matched with the corresponding interval segment, and then the default path of the current instance of the node is set according to the control logic, thereby realizing the dynamic setting of the default path;

[0115] The score value of the path selection sequence is used to construct a segmented interval mapping control logic, the score value is divided into multiple priority segments according to the numerical interval, and each priority segment corresponds to a default path setting policy label; during the execution of the process instance, the position of the path score value associated with the current condition judgment node in the path selection sequence is matched with the corresponding interval segment, and then the default path of the current instance of the node is set according to the control logic, thereby realizing the dynamic setting of the default path of the path selection sequence.

[0116] In this control logic, the score value is a conditional path default selection evaluation index obtained from the attention mechanism, which is usually in a continuous real number interval, so a standard segmented mapping system needs to be constructed to ensure that different path score values can be stably assigned to the set interval segment. The priority segment is a discrete expression of the execution advantage and disadvantage ordering result of the path, and the default path setting policy label is an execution configuration parameter bound to the process control layer, which determines how to automatically select the path at a specific judgment node in a specific instance. The automatic transition of the path logic state is realized through the numerical mapping between the path score value and the priority segment, thereby avoiding the rigidity problem caused by static setting. Finally, this mechanism allows the process model to have the ability to adjust in real time according to the path advantage and disadvantage evaluation result, and to guarantee the intelligent dynamic regulation and control ability of path selection in the face of business flow strategy adjustment, data offset fluctuation and other scenarios.

[0117] The real-time trigger frequency of each condition path during the process running, the path execution result and the cooperative change state of the upstream and downstream dependent nodes are collected, and the behavior offset rate of the path in different time windows is statistically modeled to form a feedback feature data set corresponding to the path;

[0118] In the process of running, in order to realize the dynamic update of path evaluation index, the system needs to continuously collect the key feedback indicators of each conditional path in actual execution. First, based on the real-time log monitoring mechanism, record the path trigger frequency of each conditional judgment node in different process instances in the process, accurate to the timestamp level; Then, combined with the execution result state of the path, such as success, failure, interruption, etc., extract the path execution effect information. And further analyze the dependence change between each path and its upstream and downstream nodes, capture the state change behavior in collaborative execution. On this basis, a sliding time window mechanism is constructed to statistically aggregate the behavior characteristics of the path in multiple time periods, focusing on extracting the trigger volatility, execution stability and linkage change of adjacent node behavior patterns of each path in each time window, and then forming a multi-dimensional feedback feature data set. This data set not only describes the current execution state of the path, but also reflects the stability and adaptability changes in the time evolution process.

[0119] Real-time trigger frequency reflects the activity and response frequency of the path, which is a quantitative indicator of dynamic usage intensity; The path execution result as a performance feedback benchmark provides an important basis for judging the reliability of path execution; The cooperative change state of upstream and downstream dependent nodes is used to measure whether the path in the process is linked or deviated from its surrounding structure, to identify unstable factors within the process. The behavior deviation rate model calculates the variation degree of frequency and result of the path in multiple time windows to comprehensively evaluate its short-term and long-term performance trends. The final feedback feature data set will be input into the evaluation module in the form of structured tensor, providing dynamic support for subsequent path priority adjustment, default path optimization and strategy backtracking mechanisms. By capturing and modeling the path behavior feedback in real time, the process control strategy is refined, perceptible and adaptively upgraded.

[0120] Based on the feedback feature data set, an updated path evaluation index update model is constructed, the process running feedback data is input into the model, the updated path evaluation index is generated, and the original index result is replaced. At the same time, the update result is synchronized to the path scoring network and the path selection sequence, a closed-loop feedback channel is constructed, and the dynamic regulation and adaptive optimization of the default path are realized.

[0121] To realize the dynamic regulation and adaptive optimization of the default path configuration, the feedback feature data set collected in real time during the process running needs to be input into the path evaluation index updating model to complete the intelligent adjustment of the path priority. The specific process includes: first, based on the feedback feature data set, an input tensor is constructed, which includes the real-time trigger frequency trend, execution result stability index, upstream and downstream node coordination state change value and behavior deviation rate statistical parameters of each path; then the tensor is input into the path evaluation index updating model, which automatically learns the dynamic evolution mode of path performance by comparing the deviation degree of current running feedback and historical evaluation value, and then calculates the updated path evaluation index; then the index is used to replace the original index result and is transmitted into the path scoring network and path selection sequence synchronously, so as to complete the update of the path priority and realize the closed-loop regulation of the whole path decision chain. Through this mechanism, the recommended order of the path will be adjusted adaptively with the change of the running data, enhancing the response flexibility of the system to the complex business environment and the intelligent level of the process modeling.

[0122] The path evaluation index updating model is a dynamic modeling mechanism that integrates feedback feature learning and priority adjustment, including multiple core features. First, the feedback feature data set covers multiple-dimensional behavior indicators such as trigger frequency, execution result, coordination state and deviation rate in the construction process, forming a panoramic description of the path running state; second, the updating model usually uses recursive neural structure or residual learning module inside, combining historical index benchmark and feedback deviation amount for multi-scale error fitting and dynamic weight adjustment; its output result is the updated path evaluation index, which quantitatively expresses the current recommended credibility and priority order of the path; finally, by synchronizing the updated evaluation index result to the path scoring network and path selection sequence in real time, it ensures that the model updating and path recommendation mechanism are in consistent logical state. This process builds a closed-loop structure of "evaluation-recommendation-feedback-re-evaluation", enabling the default path configuration mechanism to have self-learning, self-adjusting and long-term evolution ability, effectively improving the accuracy and robustness of process intelligent modeling.

[0123] The above-described embodiments can be implemented in part or in whole through software, hardware, firmware or any combination thereof. When implemented in software, the above-described embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When loaded and executed by a computer, the computer instructions or computer programs cause the computer to perform all or part of the processes or functions described in the embodiments of the present application. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable apparatus. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through a wired or wireless (for example, infrared, wireless, microwave, etc.) manner. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. containing one or more available medium collections. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state disk.

[0124] It should be understood that the size of the sequence number of each process described above in various embodiments of the present application does not mean the order of execution, and the execution order of each process should be determined by its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0125] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be realized in electronic hardware, or a combination of computer software and electronic hardware. Whether the functions are performed in hardware or software depends on the specific application and design 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 the present application.

[0126] In several embodiments provided in the present application, it should be understood that the disclosed system and method can be implemented in other ways. For example, the above-described embodiments are only illustrative, for example, the division of units is only a logical function division, and actual implementation can have another division manner, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed units can be indirect coupling or communication connection through some interfaces, devices or units, and can be electrical, mechanical or other forms.

[0127] The units described as separate components may or may not be physically separate, and the components displayed as units may or may not be physical units, that is, may be located in one place, or may also be distributed to multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0128] In addition, each functional unit in each embodiment of the present application can be integrated into a processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit.

[0129] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for rapid modeling generation oriented towards enterprise business processes, characterized in that, Specifically comprising the following steps: S1, after receiving the service process modeling request, performing structure similarity analysis, comparing the path vector of the current process structure input with the path vector of a plurality of process templates, judging the structure coincidence degree and the consistency of the conditional node semantics, and identifying whether it belongs to the case of generating a process by merging a plurality of templates; S2, in the case of identifying the process generated by merging a plurality of templates, extracting the original value of the trigger frequency of each path corresponding to the conditional judgment node, the frequency sampling time range, the sample number, the acquisition strategy label and the path trigger distribution density, constructing a normalized offset vector, and outputting a normalized difference result; S3, inputting the normalized difference result into the normalized transformation function family constructed by the deep residual network, combining the period compensation factor, the sampling variance adjustment factor and the distribution balance tensor, performing nonlinear normalization processing on the original frequency of each path, and generating a frequency vector sequence with unified dimensions; S4, based on the frequency vector sequence with unified dimensions, fusing the path historical execution success rate, the upstream and downstream cooperative stability and the behavior deviation rate, generating a conditional path default selection evaluation index, scoring and sorting each path by using a guided attention mechanism, and forming a path selection sequence; S5, according to the score value of the path selection sequence, dynamically setting the default path by using the segmented interval mapping control logic, and updating the path evaluation index through the process running feedback, realizing the dynamic regulation and adaptive optimization of the default path.

2. The enterprise business process oriented rapid modeling generation method according to claim 1, characterized in that, S1 specifically comprises the following steps: S101, after receiving the service process modeling request, performing structure similarity analysis, comparing the path vector of the current process structure input with the path vector of a plurality of process templates, judging the structure coincidence degree and the consistency of the conditional node semantics, and identifying whether it belongs to the case of generating a process by merging a plurality of templates; S102, performing structure similarity analysis, calculating the cosine similarity between the path vector of the current process structure input and the path vector of a plurality of process templates, and comparing the node field semantic vectors, to generate a structure coincidence degree score set and a conditional node semantic consistency score set; S103, calculating the number of process templates with a coincidence degree exceeding a first preset threshold in the structure coincidence degree score set, and calculating the number of conditional node pairs with a semantic similarity exceeding a second preset threshold in the conditional node semantic consistency score set, when both exceed the corresponding template merging judgment threshold, it is determined that the current process structure input belongs to the case of generating a process by merging a plurality of templates.

3. The enterprise business process oriented rapid modeling generation method according to claim 2, characterized in that, S102 specifically comprises: Normalizing the path vector of the current process structure input and the path vector of a plurality of process templates, calculating the vector angle similarity between each two paths by cosine similarity, and storing the result in the structure coincidence degree score set; Inputting the text content of the conditional judgment field in the current process structure into a double-channel semantic embedding model to generate a node field semantic vector, and calculating the Euclidean distance between the node field semantic vector and the conditional node semantic vector in a plurality of process templates to construct a conditional node semantic consistency score set; Statistically aggregating the structure coincidence degree score set and the conditional node semantic consistency score set by template dimension to generate the structure coincidence degree score set and the conditional node semantic consistency score set.

4. The enterprise business process oriented rapid modeling generation method of claim 1, wherein, S2 specifically comprises the following steps: S201, in the case of identifying the multiple template merging generation process, performing path association retrieval on each conditional judgment node, extracting the trigger frequency original value of the corresponding path based on the source template record, and simultaneously obtaining the frequency sampling time range, sample quantity, acquisition strategy label and path trigger distribution density corresponding to each path, and constructing a structured path frequency information set; S202, performing time scale calibration on the frequency sampling time range in the structured path frequency information set, aligning the trigger frequency original values counted in different time periods in a cycle, combining the statistical aperture differences marked by the sample quantity and the acquisition strategy label, and uniformly mapping to a standard frequency reference system, and constructing a normalized offset vector reflecting the path sampling intensity and timing distribution offset based on the path trigger distribution density; S203, performing dimension expansion and aggregation processing on the normalized offset vector, calculating the deviation degree of each path in each normalized dimension, outputting the normalized difference result through a vector comparison function, and forming a normalized difference index set that can quantitatively represent the statistical feature level of multiple paths.

5. The enterprise business process oriented rapid modeling generation method according to claim 4, characterized in that, S203 specifically comprises: Dimensionally expanding the normalized offset vector corresponding to each path, respectively extracting four types of dimensions of dimensional offset, time period offset, sample quantity weight offset and trigger distribution density offset, and organizing the normalized features of each path into a multi-dimensional vector set of a unified structure; Based on the multi-dimensional vector set, calculate the difference value between each type of offset dimension between all paths, and use the Mahalanobis distance and cosine angle measurement to construct a normalized offset degree matrix between path pairs, thereby realizing comprehensive comparison and quantization between different offset types; Integrate the normalized offset degree matrix through a weighted aggregation strategy, set the aggregation weight according to the statistical characteristics of the path context, output the normalized difference result, and form a structured normalized difference index set, which is used as an input parameter in the path frequency normalization processing and path evaluation mechanism.

6. The enterprise business process oriented rapid modeling generation method of claim 1, wherein, S3 specifically comprises: After encoding the normalized difference result into a multi-dimensional statistical feature matrix, input it into the normalized transformation function family constructed based on the deep residual network structure, each residual unit of the network is composed of a linear mapping layer, a normalization layer and an activation mapping layer, and uses cross-layer connection to keep the integrity of difference information in the multi-level mapping process; Construct the period compensation factor, the sampling variance adjustment factor and the distribution balance tensor into the normalized mapping weight matrix respectively, and fuse and encode them with the normalized difference result in the channel dimension, participate in the feature transformation process in the deep residual network through multi-dimensional conditional injection, so that the nonlinear normalization calculation can maintain the expression ability of the original difference of the path frequency while offsetting the structural deviation caused by the statistical dimension; The normalized result output by the deep residual network is used as a nonlinear transformation function to map and transform each path's original frequency value item by item, and a frequency vector sequence of a unified dimension is obtained, each frequency vector corresponds to a conditional path, and the structure is consistent in the sequence to support subsequent path sorting and evaluation tasks.

7. The enterprise business process oriented rapid modeling generation method of claim 1, wherein, S4 specifically comprises the following steps: S401. Normalize and encode the frequency vector sequence with the historical execution success rate, upstream and downstream coordination stability and behavior deviation rate of the corresponding path, and perform feature concatenation on the path dimension to construct a multi-dimensional evaluation vector set that integrates the historical execution success rate, upstream and downstream coordination stability and behavior deviation rate of the path, which is used to form a comprehensive path representation. S402. Input the comprehensive path representation into the path scoring network constructed by the guided attention mechanism. By setting the path-level contextual attention prior, guide the attention mechanism to dynamically weight the key dimensions in the fused features and generate the conditional path default selection evaluation index for each path, which represents the default recommendation priority of the path in the current modeling scenario. S403. Based on the path score values ​​included in the conditional path default selection evaluation index, perform numerical comparisons on all paths and sort them from high to low score values, outputting a path selection sequence. Each position in the path selection sequence uniquely corresponds to a conditional path and serves as the input basis for the default path configuration and execution control strategy.

8. The enterprise business process oriented rapid modeling generation method of claim 7, wherein, S402 specifically refers to: The features of each dimension in the comprehensive path representation are embedded into a unified tensor space, and a local dependency graph of each path in the process structure is constructed through a context window. This graph is used to extract prior information of path-level context attention and represent the degree of interaction dependency of different paths at the structural level. Based on path-level contextual attention prior information, a fusion-guided attention network structure is constructed. The key dimensions in the comprehensive path representation are used as the basic input for generating attention weights. Dynamic weights are applied to different dimensions through dot product attention calculation to form a high-order attention representation that reflects the multidimensional advantages of the path. The higher-order attention representation is output as a single-channel path score through a fully connected mapping layer, which serves as the evaluation index for the default selection of conditional paths. The score value represents the default recommendation priority of the path in the current modeling context and supports serialization in subsequent path ranking and policy formulation tasks.

9. The enterprise business process oriented rapid modeling generation method of claim 1, wherein, S5 specifically refers to: Based on the score value of the path selection sequence, a segmented interval mapping control logic is constructed. The score value is divided into multiple priority segments according to the numerical range. Each priority segment corresponds to a default path setting strategy label. When the process instance is executed, the position of the path score value associated with the node in the path selection sequence is matched with the corresponding interval segment based on the current conditions. Then, the default path of the current instance of the node is set according to the control logic, so as to realize the dynamic setting of the default path. The system collects the real-time trigger frequency of each condition path, the path execution results, and the collaborative change status of upstream and downstream dependent nodes during the data collection process. It also statistically models the behavior offset rate of the path in different time windows to form a set of feedback feature data corresponding to the path. A path evaluation index update model is constructed based on the feedback feature data set. The process operation feedback data is input into the model to generate an updated path evaluation index and replace the original index result. At the same time, the updated result is synchronized to the path scoring network and the path selection sequence to build a closed-loop feedback path and realize the dynamic control and adaptive optimization of the default path.

Images