Consulting design process management and control system

Abstract

The application discloses a consultation design process management and control system and relates to the technical field of computer-aided project management; structural confusion and low efficiency caused by complex task and personnel coordination relationship and lack of quantitative control means in the traditional consultation design process are solved. The application gathers project management logs, collaborative editing records, personnel skill tags and historical cases through a data center module, extracts low-entropy process fragments from historical successful projects and generates an initial process template with a collaborative entropy value lower than an intervention threshold through a cold start template generation module, constructs a heterogeneous hypergraph at a fixed time interval and calculates the collaborative entropy value of the current running process based on a normalized Laplacian matrix through a collaborative entropy real-time monitoring module, and selects the optimal control action from a preset action set and executes it through a pre-trained value network when the entropy value is over the limit through an entropy regularization dynamic reconstruction module, so that the process structure quantitative monitoring and adaptive optimization control of the whole life cycle of the consultation design project are realized.

Description

Technical Field

[0001] This invention belongs to the field of computer-aided project management technology, specifically a consulting and design process control system. Background Technology

[0002] The consulting and design industry is a typical knowledge-intensive service industry, encompassing engineering consulting, architectural design, and Consulting, management consulting, and other sub-sectors all revolve around projects as their core production unit. This involves collaboration among multidisciplinary and multi-role personnel to complete the entire process, from needs assessment and solution design to internal review, client confirmation, and deliverables. Unlike standardized production lines in manufacturing, consulting and design projects are highly non-standardized: each project has significantly different needs, scale, technical approaches, and personnel configurations; the dependencies between tasks are complex and dynamically changing during project execution; design solutions undergo frequent iterations; and collaboration among personnel is characterized by its temporary, interdisciplinary, and high-frequency nature.

[0003] The management of consulting and design projects has long faced unique challenges. On the one hand, the intricate coupling relationships between tasks, between tasks and personnel, and between personnel create a constantly evolving collaborative network, making it difficult for managers to accurately assess the current process status from a holistic perspective. On the other hand, the degree of structural disorder in the process—i.e., the lack of effective quantitative measurement tools due to fragmented tasks, redundant communication paths, and uneven workloads—means that managers typically rely on personal experience and periodic reports for coarse-grained judgments. Furthermore, when localized disruptions or bottlenecks occur in the process, there is a lack of systematic decision support methods regarding what intervention measures to take, when to intervene, and at what cost to restore order. These factors collectively contribute to frequent delays, quality defects, and cost overruns in consulting and design projects.

[0004] In recent years, although project management software, online collaborative editing platforms and instant messaging tools have been widely used in consulting and design firms, these tools have accumulated a large amount of process data, providing the possibility for data-driven process control. However, existing systems mostly focus on the functional implementation of individual tools or post-event statistical analysis, lacking a set of quantifiable and adaptive process structure control technology solutions that run through the entire project lifecycle.

[0005] The following problems exist in the existing technology:

[0006] Traditional consulting and design process control systems often rely on data recording from a single tool, lacking the ability to uniformly aggregate and standardize multi-source heterogeneous data such as project management logs, collaborative editing records, personnel skill tags, and historical project cases. This results in inconsistent data definitions and semantic fragmentation, making it difficult to support quantitative analysis across the entire process.

[0007] In the initial stage of new consulting and design projects, existing systems often rely on the personal experience of managers or directly apply general templates to formulate initial process templates. They lack the technical means to generate low collaborative entropy starting templates based on the smooth-running process segments in historical successful projects, resulting in a high risk of structural disorder at the start of the project.

[0008] The existing system lacks dynamic quantification of the structural disorder caused by the intertwining and coupling of tasks and personnel during the operation of consulting and design projects. Managers cannot know the collaborative health status of the current process in real time, and it is difficult to discover and locate problems such as process fragmentation, broken personnel collaboration, or uneven workload in a timely manner.

[0009] When structural chaos occurs in the consulting and design process, the existing system lacks the means to adaptively select the optimal control action based on quantitative indicators. Managers usually take intervention measures based on intuition and are unable to scientifically weigh the control costs against the entropy reduction effect, which can easily lead to over-intervention or under-intervention. Summary of the Invention

[0010] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes a consulting design process control system to solve the above-mentioned technical problem.

[0011] The first aspect of the present invention provides a consulting design process control system, comprising the following modules:

[0012] The data platform module is used to collect and standardize the processing of multi-source heterogeneous data throughout the entire lifecycle of consulting and design projects. The multi-source heterogeneous data includes at least project management logs, collaborative editing records, personnel skill tags, and structured records of historical consulting and design project cases.

[0013] The cold start template generation module is used to extract low-entropy process segments with collaborative entropy values ​​lower than a preset collaborative entropy intervention threshold from historical consulting and design project cases when a new consulting and design project is initiated. Using the structural characteristics of the low-entropy process segments as a reference, an initial process template with a collaborative entropy value lower than the preset collaborative entropy intervention threshold is generated according to the requirements of the new consulting and design project. The initial process template includes the serial and parallel dependencies between tasks, the pre-assigned personnel roles for each task, and the preset positions of key collaborative checkpoints.

[0014] The collaborative entropy real-time monitoring module is used to collect collaborative events at fixed time intervals during the operation of the consulting and design project, construct a heterogeneous hypergraph in real time, and calculate the collaborative entropy value of the current operation process of the consulting and design project based on the normalized Laplace matrix of the heterogeneous hypergraph.

[0015] The entropy regularization dynamic reconstruction module has a built-in pre-trained approximate dynamic programming value network and a preset set of process intervention actions. When the collaborative entropy value of the current operation process of the consulting and design project exceeds the collaborative entropy intervention threshold, a current process state vector is constructed based on the task completion ratio, personnel load distribution, and the collaborative entropy value of the current operation process. Using the approximate dynamic programming value network, for each candidate action in the process intervention action set, the expected long-term cumulative cost consisting of the action execution cost and the collaborative entropy value as a state penalty term is calculated. The candidate action with the minimum expected long-term cumulative cost is selected as the optimal control action, and an execution instruction is output to change the topology and personnel configuration of the current operation process of the consulting and design project.

[0016] Preferably, the data platform module includes a standardization engine, a skill vector encoder, and a historical case reconstructor;

[0017] The standardization engine includes a verb mapping table and a quintuple constructor. The verb mapping table stores the mapping relationship between action verbs from different sources and their corresponding verbs in the standard action set. The quintuple constructor outputs three types of standardized quintuples based on the mapping relationship. All three types of quintuples carry a timestamp and a payload attribute.

[0018] The skill vector encoder includes a fixed-dimensional skill classification system and a collaborative filtering inference engine. The collaborative filtering inference engine is connected to the skill classification system and infers and completes the missing dimension values ​​in the skill vectors generated based on the skill classification system.

[0019] The historical case refactoring unit includes a task graph structuring unit, an event sorting unit, and an encapsulation unit. The task graph structuring unit transforms the task decomposition structure in historical consulting and design project cases into a task graph composed of task nodes and dependency edges. The event sorting unit arranges the collaborative event records in historical consulting and design project cases into a standardized collaborative event sequence in ascending order of timestamps. The encapsulation unit encapsulates the task graph, the standardized collaborative event sequence, and the performance tags of historical consulting and design project cases into a unified data object.

[0020] Preferably, the cold start template generation module includes:

[0021] The low-entropy process segment acquisition unit is used to select cases with performance tags as successful projects from the historical consulting and design project cases, divide each successful project case into continuous time segments according to a fixed time window, construct a historical segment heterogeneous hypergraph for each time segment and calculate its historical segment synergistic entropy value, and mark the historical segment heterogeneous hypergraph corresponding to the time segment with a historical segment synergistic entropy value lower than the synergistic entropy intervention threshold as a low-entropy process segment; each low-entropy process segment is accompanied by its task graph subgraph, personnel allocation snapshot, and the eigenvalue spectrum distribution of the historical segment normalized Laplace matrix of the historical segment heterogeneous hypergraph.

[0022] Preferably, the cold start template generation module further includes:

[0023] The template generation unit incorporates a text encoder, a condition generator network, a hypergraph reconstructor network, and a spectral alignment checker.

[0024] The text encoder uses a multi-head self-attention mechanism to extract semantic encoding vectors from the requirements features of a newly launched consulting and design project; the condition generator network uses the semantic encoding vectors as conditional variables and injects conditional information into the hidden layer features through a conditional batch normalization layer to generate target domain representation vectors; the hypergraph reconstructor network maps the target domain representation vectors to the initial heterogeneous hypergraph of the newly launched consulting and design project.

[0025] The spectral alignment checker calculates the Wasserstein distance between the eigenvalue spectrum distribution of the initial normalized Laplacian matrix of the initial heterogeneous hypergraph and the eigenvalue spectrum distribution of the normalized Laplacian matrix of the historical fragments of the heterogeneous hypergraph. When the Wasserstein distance is lower than a preset distance threshold, the hypergraph structure resolver extracts the serial-parallel dependencies of task nodes, the hyperedge connections between tasks and personnel, and the collaborative checkpoints formed by multiple hyperedge connections from the initial heterogeneous hypergraph, and combines them into an initial process template.

[0026] Preferably, the collaborative entropy real-time monitoring module includes:

[0027] The collaborative event acquisition unit is used to acquire collaborative events generated within the current time window from the project management log and the collaborative editing record at preset fixed time intervals. Each collaborative event includes the associated task identifier, the identifier of the personnel involved, the event type, the timestamp, the duration of the interaction, and the identifier of the deliverable operated.

[0028] The super-edge merging unit is used to merge collaborative events that are associated with the same task identifier and whose timestamp interval is lower than a preset time interval threshold into a collaborative activity super-edge. Each collaborative activity super-edge connects the task node corresponding to the task identifier and the personnel node corresponding to the personnel identifier.

[0029] The matrix construction unit is used to construct the current vertex set by forming the task nodes corresponding to all task identifiers and the personnel nodes corresponding to all personnel identifiers appearing within the current time window, and to form the current hyperedge set by forming all collaborative activity hyperedges, thus constructing the association matrix. Construct a heterogeneous hypergraph for the current execution flow using the current vertex set and the current hyperedge set. The membership relationships between vertices and hyperedges in the heterogeneous hypergraph for the current execution flow are determined by the incidence matrix. express;

[0030] The weight calculation unit is used to, for each deliverable identified by a deliverable identifier involved in a hyperedge, query the parameter transmission frequency between each deliverable from the pre-stored design structure matrix and take the average as the information coupling degree; for the personnel involved in the hyperedge, obtain the skill vector of each person from the personnel skill tags, and obtain the task requirement vector of the task associated with the hyperedge from the project management log, calculate the cosine similarity between the skill vector and the task requirement vector and take the average as the average experience matching degree; the weight of each hyperedge is calculated by dividing the product of the information coupling degree and the interaction duration by the sum of the average experience matching degree and the preset positive constant, and then using the weight diagonal matrix. express, The diagonal elements are the weights of the corresponding hyperedges.

[0031] Preferably, the collaborative entropy real-time monitoring module further includes:

[0032] Entropy calculation unit, used for calculating entropy based on the correlation matrix and the weight diagonal matrix Calculate the vertex degree matrix and hypermarginality matrix ,in The diagonal elements are the sum of the weights of the hyperedges to which each vertex belongs. The diagonal elements represent the number of vertices contained in each hyperedge; calculate the Laplacian matrix of the symmetric normalized hypergraph. And calculate the collaborative entropy value of the current operating process of the consulting and design project. ,in To and Identity matrices of the same dimension Represents the determinant of a matrix.

[0033] Preferably, the approximate dynamic programming value network in the entropy regularization dynamic reconstruction module By minimizing the loss function Training achieved The formula is:

[0034]

[0035] in, To extract from the completed historical consulting and design project cases according to a preset time step The sampled process state vector, In order to be in The actual actions to be performed To execute The process state vector that is then transferred to the next sampling time. In order to be in Next action The cost of one immediate execution The equivalent cost coefficient per unit of collaborative entropy. for The collaborative entropy value under the following conditions; For the target value network, the parameters of the target value network. Periodic from Replication; Expectation The training sample set is composed of triplets of process state vectors, actions, and process state vectors at the next sampling time, generated from the historical consulting and design project cases.

[0036] Preferably, the entropy regularization dynamic reconstruction module includes:

[0037] The entropy regularization dynamic reconstruction module, during online decision-making, selects from a preset action set... The optimal control action is selected from the candidates, and for each candidate action... Calculate the expected long-term cumulative cost The calculation formula is:

[0038]

[0039] in, This is the current process state vector. In the state Next action The cost of executing a single action The unit collaborative entropy equivalent cost coefficient is... This represents the collaborative entropy value of the current running process. For the preset time step, This is the pre-trained approximate dynamic programming value network. To perform the action The next state vector after transition, Let be the expectation of the state transition distribution;

[0040] Elected The smallest candidate action is selected as the optimal control action. ,Right now .

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] This invention collects and standardizes the aforementioned multi-source heterogeneous data through a data platform module, maps action verbs from different sources to a standard action set, parses collaborative events into standardized quintuples, vectorizes and completes personnel skill tags, and reorganizes historical cases into unified data objects. This provides a high-quality data foundation with semantic consistency and structural uniformity for the entire system, enabling the effective organization and consumption of process data throughout the entire lifecycle of consulting and design projects.

[0043] This invention uses a cold start template generation module to extract low-entropy process segments with collaborative entropy values ​​below a threshold from historical successful project cases within a fixed time window. Using the task graph subgraph, personnel allocation snapshot, and normalized Laplace matrix eigenvalue spectral distribution attached to the low-entropy process segments as references, an initial process template aligned with the historical low-entropy structure in terms of spectral distribution is generated through spectral alignment verification. This ensures that the initial collaborative entropy value at the start of a new project is naturally lower than the collaborative entropy intervention threshold, reducing the probability of early project chaos from the source.

[0044] This invention obtains real-time collaborative events from project management logs and collaborative editing records at fixed time intervals through a collaborative entropy real-time monitoring module, constructs a heterogeneous hypergraph composed of the current vertex set and the current hyperedge set, and calculates the collaborative entropy value of the current running process based on the normalized Laplace matrix of the heterogeneous hypergraph. This quantifies the structural coupling state of tasks and personnel into a dimensionless scalar, providing an objective metric for process health that can be calculated in real time and compared horizontally.

[0045] This invention utilizes an entropy regularization dynamic reconstruction module to calculate the expected long-term cumulative cost of each candidate action in a preset action set when the collaborative entropy value exceeds the intervention threshold. This cost consists of the action execution cost and the collaborative entropy penalty term. The optimal control action that minimizes this cost is selected, and the execution instruction is automatically output to change the topology and personnel configuration of the current operation process. This achieves adaptive closed-loop control that automatically achieves the optimal trade-off between control costs and process orderliness. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the module flow of the present invention. Detailed Implementation

[0047] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0048] Please see Figure 1 This invention is a consulting design process control system, comprising the following modules:

[0049] The data platform module is used to collect and standardize the processing of multi-source heterogeneous data throughout the entire lifecycle of consulting and design projects. The multi-source heterogeneous data includes at least project management logs, collaborative editing records, personnel skill tags, and structured records of historical consulting and design project cases.

[0050] The cold start template generation module is used to extract low-entropy process segments with collaborative entropy values ​​lower than a preset collaborative entropy intervention threshold from historical consulting and design project cases when a new consulting and design project is initiated. Using the structural characteristics of the low-entropy process segments as a reference, an initial process template with a collaborative entropy value lower than the preset collaborative entropy intervention threshold is generated according to the requirements of the new consulting and design project. The initial process template includes the serial and parallel dependencies between tasks, the pre-assigned personnel roles for each task, and the preset positions of key collaborative checkpoints.

[0051] The collaborative entropy real-time monitoring module is used to collect collaborative events at fixed time intervals during the operation of the consulting and design project, construct a heterogeneous hypergraph in real time, and calculate the collaborative entropy value of the current operation process of the consulting and design project based on the normalized Laplace matrix of the heterogeneous hypergraph.

[0052] The entropy regularization dynamic reconstruction module has a built-in pre-trained approximate dynamic programming value network and a preset set of process intervention actions. When the collaborative entropy value of the current operation process of the consulting and design project exceeds the collaborative entropy intervention threshold, a current process state vector is constructed based on the task completion ratio, personnel load distribution, and the collaborative entropy value of the current operation process. Using the approximate dynamic programming value network, for each candidate action in the process intervention action set, the expected long-term cumulative cost consisting of the action execution cost and the collaborative entropy value as a state penalty term is calculated. The candidate action with the minimum expected long-term cumulative cost is selected as the optimal control action, and an execution instruction is output to change the topology and personnel configuration of the current operation process of the consulting and design project.

[0053] Specifically, the data platform module continuously aggregates full lifecycle data, including project management logs, collaborative editing records, personnel skill tags, and historical consulting and design project cases, from multiple heterogeneous tools such as project management software and collaborative editing platforms. Through a standardization engine, action verbs from different sources are uniformly mapped to a standard action set, and collaborative events are parsed into standardized quintuples carrying timestamps and load attributes. Through a skill vector encoder, personnel skill tags are transformed into fixed-dimensional skill vectors, and a collaborative filtering inferencer is used to infer and complete sparse dimensions. Through a historical case reconstructor, historical projects are transformed into a unified data object containing directed acyclic graphs, standardized collaborative event sequences, and performance tags. When a new consulting and design project is initiated, the cold start template generation module extracts low-entropy process segments with collaborative entropy values ​​below the intervention threshold from historical successful projects within a fixed time window. Using the task graph subgraph, personnel allocation snapshot, and eigenvalue distribution of the normalized Laplacian matrix of the historical segments as structural references, the module extracts semantic encoding vectors from the new project's requirements features through a text encoder. These vectors are then used by a condition generator network to generate target domain representation vectors, which are mapped to an initial heterogeneous hypergraph by a hypergraph reconstructor network. After the spectral alignment checker calculates the Wasserstein distance between the initial heterogeneous hypergraph and the spectral distribution of the historical segment heterogeneous hypergraph and passes the check, the hypergraph structure parser extracts the serial and parallel dependencies between tasks, the pre-assigned personnel roles for each task, and the preset positions of key collaborative checkpoints. These are then combined into an initial process template with a collaborative entropy value below the intervention threshold and output to the project management tool. During project operation, the collaborative entropy real-time monitoring module retrieves real-time collaborative events from project management logs and collaborative editing records at fixed time intervals. The hyperedge merging unit groups events with the same related tasks and similar timing into collaborative activity hyperedges. A heterogeneous hypergraph of the current workflow is constructed using the current vertex set and the current hyperedge set, with its structural information represented by an association matrix and a weighted diagonal matrix. The entropy calculation unit calculates the symmetric normalized hypergraph Laplacian matrix and the collaborative entropy value, pushing this entropy value to the project management dashboard to reflect the degree of structural disorder in the workflow in real time. When the collaborative entropy value exceeds the intervention threshold, the entropy regularization dynamic reconstruction module is triggered. Based on the task completion ratio, personnel load distribution, and collaborative entropy value of the current workflow, a state vector is constructed. Using a pre-trained value network, the expected long-term cumulative cost, consisting of the action execution cost and the collaborative entropy penalty term, is calculated for each candidate action in the preset action set. The optimal control action is selected and automatically pushed to the project management tool through the instruction distribution interface to change the workflow topology and personnel configuration, completing the adaptive closed-loop control of perception, decision-making, and execution.

[0054] In one embodiment of the present invention, the data platform module includes a standardization engine, a skill vector encoder, and a historical case reconstructor;

[0055] The standardization engine includes a verb mapping table and a quintuple constructor. The verb mapping table stores the mapping relationship between action verbs from different sources and their corresponding verbs in the standard action set. The quintuple constructor outputs three types of standardized quintuples based on the mapping relationship. All three types of quintuples carry a timestamp and a payload attribute.

[0056] The skill vector encoder includes a fixed-dimensional skill classification system and a collaborative filtering inference engine. The collaborative filtering inference engine is connected to the skill classification system and infers and completes the missing dimension values ​​in the skill vectors generated based on the skill classification system.

[0057] The historical case refactoring unit includes a task graph structuring unit, an event sorting unit, and an encapsulation unit. The task graph structuring unit transforms the task decomposition structure in historical consulting and design project cases into a task graph composed of task nodes and dependency edges. The event sorting unit arranges the collaborative event records in historical consulting and design project cases into a standardized collaborative event sequence in ascending order of timestamps. The encapsulation unit encapsulates the task graph, the standardized collaborative event sequence, and the performance tags of historical consulting and design project cases into a unified data object.

[0058] Specifically, the standardization engine extracts project management logs from project management software and collaborative editing records from collaborative editing platforms through pre-defined adapter interfaces. The extracted raw events are then fed into the standardization engine's internal verb mapping table. This verb mapping table is a pre-defined relational mapping table used to uniformly map action verbs expressing the same semantics but with different syntax from different project management software or collaborative editing platforms to a fixed set of standard actions. For example, project management tools... Mark the start of the task as "Start," tools Mark it as "Start", Tools The verb mapping table will map all instances of "in progress" to the standard action "task started"; in collaborative editing platforms, "edit," "modify," and "update" are mapped to "document editing." The mapping table is constructed based on interface surveys and manual annotations of common consulting and design tools, and can be expanded as new tools are integrated.

[0059] The quintuple constructor receives the raw event mapped by the verb mapping table and parses it into a standardized quintuple form. Each quintuple has the following structure: [Head Entity Identifier, Relationship Type, Tail Entity Identifier, Timestamp, Payload Attribute]. The head and tail entity identifiers are globally unique identifiers, representing specific task or personnel nodes. The relationship type is selected from three types of relationships in the standard action set: task assignment, collaborative editing, and instant messaging reference. The timestamp is accurate to milliseconds and uses Coordinated Universal Time (UTC). The payload attribute extends the storage of additional information about the event in key-value pairs and must include the fields "interaction duration" and "list of deliverable identifiers operated on".

[0060] According to the mapping rules used by the quintuple constructor, three types of normalized quintuples are output, all of which carry a timestamp and a payload attribute.

[0061] The first type of 5-tuple corresponds to a task assignment relationship. The header entity is the task node identifier, the footer entity is the personnel node identifier, and the relationship type is task assignment. For example, a project management log record "task..." "Assigned to architect Zhang Gong", after mapping, generates […]. Task assignment, Zhang's personnel identification, 2025-06-15 Duration: 0, Deliverable Identifier: []}】, The duration is 0 because the assignment action itself does not consume working time.

[0062] The second type of quintuple corresponds to a collaborative editing relationship. Both the header and footer entity identifiers are personnel node identifiers, indicating that two personnel engaged in editing interaction on the same deliverable; the relationship type is collaborative editing. For example, the collaborative editing record reveals that structural engineer Li and architectural designer Wang edited the same drawing within 30 minutes of each other. This generates [Li's personnel identifier, collaborative editor, Wang's personnel identifier, timestamp of the midpoint between their editing times, {duration: 1800 seconds, deliverable identifier: [ 】

[0063] The third type of quintuple corresponds to an instant messaging reference relationship. The header entity identifier is the person node identifier, and the tail entity identifier is the task node identifier, indicating that a person explicitly mentions a task in the message; the relationship type is instant messaging reference. For example, if the message "Please pay attention" is retrieved from the instant messaging log... If the sender of the "Calculation Progress" is Engineer Zhao, then generate [Engineer Zhao's Personnel Identifier, Instant Messaging Reference], Message sending timestamp, {duration: 0, deliverable identifier: []}

[0064] The skill vector encoder is responsible for converting personnel skill tags into fixed-dimensional real-valued vectors and appropriately completing any missing skill dimensions due to data sparsity. Internally, it includes a collaborative filtering inference engine and a fixed-dimensional skill classification system. The skill classification system adopts a skill tree structure commonly used in the consulting and design industry, categorizing all possible skills into... One standard skill dimension, The default value is 128. Each person is initialized with an original skill vector based on their actual skill tags. Each dimension of the vector corresponds to a standard skill. If the person clearly possesses this skill, the initial value is set to a proficiency value between 0 and 1. If not marked, it is left blank for the time being.

[0065] The collaborative filtering inference engine is integrated into the skill classification system to infer and complete the missing dimensions in the original skill vectors. The completion algorithm first calculates the cosine similarity of the current user with all other users whose skills are fully labeled across the known skill dimensions, and then selects the user with the highest similarity. A neighbor, The default value is 30. Then, for each missing dimension of the current personnel, use this... The weighted average of the proficiency levels of the neighbors in each dimension is used as the inferred value, and the weights are the similarity scores. For example, structural engineer Zhang has a proficiency of 0.9 in the "seismic design" dimension and 0.5 in the "building information modeling" dimension, but is missing in the "wind engineering" dimension. We find the five neighbors whose proficiency levels in the "wind engineering" dimension are most similar to his, with proficiency levels of 0.8, 0.7, 0.6, 0.9, and 0.7 respectively, corresponding to similarity weights of 0.92, 0.88, 0.85, 0.91, and 0.87. Therefore, Zhang's inferred value for "wind engineering" is a weighted average of approximately 0.746. After completion, a complete personnel skill vector is formed.

[0066] The historical case refactoring tool extracts completed and archived historical consulting and design project cases from the historical project archive database, transforming them into unified data objects containing task diagram structured units, event sequencing units, and encapsulation units. Regarding the processed historical consulting and design project case data, it is obtained by batch importing the company's existing project archive files during the initial system deployment. After the system continues to run, new cases are supplemented by the system at the end of the project's entire lifecycle, through unified archiving of the project's full standardized 5-tuple sequence output by the standardization engine, the participant skill vector snapshots output by the skill vector encoder, and the original task breakdown structure data and performance evaluation data exported from the project management software.

[0067] The task graph structured unit receives raw task decomposition structure data from historical consulting and design project cases and transforms it into a directed acyclic graph (DAG). Each node represents a task, with attributes including a unique task identifier, name, planned and actual start and end dates, task type (e.g., design, review, delivery), and a skills requirement vector. Dependencies between tasks are transformed into directed edges, with types including finish-start, start-start, finish-finish, and start-finish. Edge attributes record the dependency type and lag time.

[0068] The event sorting unit arranges all standardized collaborative events (i.e., the five-tuples output by the standardization engine) generated throughout the entire lifecycle of historical consulting and design project cases in ascending order of timestamps, forming a standardized collaborative event sequence. For events with the same timestamp, a secondary sort is further performed based on the priority of the relationship type, with the priority order from high to low as follows: task assignment relationship, collaborative editing relationship, and instant messaging reference relationship.

[0069] The encapsulation unit integrates directed acyclic graphs, standardized collaborative event sequences, and performance tags from historical consulting and design project cases into a unified data object. Performance tags are used to distinguish between successful and unsuccessful projects, with tag values ​​being binary; for example, a project that simultaneously meets three conditions—actual project duration not exceeding 110% of the planned duration, customer satisfaction rating of "excellent" or "good," and no major design rework—is marked as a successful project; otherwise, it is considered an unsuccessful project. The encapsulated unified data object uses a unique project identifier as its primary key.

[0070] In one embodiment of the present invention, the cold start template generation module includes:

[0071] The low-entropy process segment acquisition unit is used to select cases with performance tags as successful projects from the historical consulting and design project cases, divide each successful project case into continuous time segments according to a fixed time window, construct a historical segment heterogeneous hypergraph for each time segment and calculate its historical segment synergistic entropy value, and mark the historical segment heterogeneous hypergraph corresponding to the time segment with a historical segment synergistic entropy value lower than the synergistic entropy intervention threshold as a low-entropy process segment; each low-entropy process segment is accompanied by its task graph subgraph, personnel allocation snapshot, and the eigenvalue spectrum distribution of the historical segment normalized Laplace matrix of the historical segment heterogeneous hypergraph.

[0072] Specifically, in the offline preprocessing stage, cases with performance tags indicating successful projects are selected from historical consulting and design project cases. For each successful project case, its entire lifecycle is divided into multiple consecutive time segments according to a fixed time window. The length of the fixed time window is 5% of the total project duration, with a minimum limit of 1 day. For example, for a project with a total duration of 120 days, the window length is 6 days, resulting in 20 time segments; for a project with a total duration of 30 days, 5% is 1.5 days, not less than the minimum limit, and the window length is 1 day, resulting in 30 time segments. If the remaining duration of the last time segment after division is less than a complete window, the remaining portion is merged into the previous time segment for processing.

[0073] For each time segment, all collaborative event data occurring within that segment are extracted to construct a historical segment heterogeneous hypergraph. The collaborative event data originates from the standardized collaborative event sequence of this successful project case. Standardized collaborative events falling within the current time segment are selected based on their timestamps and then processed according to the same hypergraph construction rules as during online monitoring. Specifically, a historical vertex set is constructed using all task nodes corresponding to all task identifiers and all personnel nodes corresponding to all personnel identifiers appearing within the time segment. Collaborative events associated with the same task identifier and whose timestamp intervals are below a preset time interval threshold are grouped into a collaborative activity hyperedge. Each collaborative activity hyperedge connects the task node corresponding to that task identifier and the personnel node corresponding to the relevant personnel identifier. All merged collaborative activity hyperedges constitute the historical hyperedge set. This yields the historical segment association matrix for that time segment. The rows correspond to vertices in the historical vertex set, and the columns correspond to hyperedges in the historical hyperedge set. If a vertex belongs to a hyperedge, the corresponding element is 1; otherwise, it is 0. A historical fragment heterogeneous hypergraph consists of a historical vertex set and a historical hyperedge set. The membership relationship between vertices and hyperedges is determined by… express.

[0074] Calculate the weight for each collaborative activity hyperedge. For the deliverable identifiers involved in a hyperedge, query the parameter transfer frequency between these deliverables from the pre-stored design structure matrix, and take the arithmetic mean as the information coupling degree. The pre-stored design structure matrix is ​​a phalanx, This represents the total number of deliverable types in the design structure matrix, with rows and columns indexed by deliverable identifiers, and elements... Indicates the delivered goods For the delivered goods The frequency of parameter transmission is calculated from historical project data. For example, if the architectural design drawings... Each time the structural calculation sheet is revised, On average, it needs to be updated 2.3 times. For the personnel involved in the hyperedge, the skill vector of each person is obtained from the personnel skill vectors stored in the system, and the task requirement vector of the task associated with the hyperedge is obtained from the project management log. The cosine similarity between the skill vector and the task requirement vector of each person is calculated, and the arithmetic mean of these similarities is taken as the average empirical matching degree. The duration of the interaction is recorded as... The default positive number is (Taking a value such as 0.01), then the weight of the hyperedge is... The weights of all hyperedges constitute the historical segment weight diagonal matrix. Its diagonal elements are the weights of each hyperedge, and its off-diagonal elements are 0.

[0075] After obtaining the historical fragment correlation matrix and historical segment weight diagonal matrix Then, calculate the vertex degree matrix of the historical fragment. and historical fragment hyper-edge degree matrix ,in The The diagonal elements are That is, the first The sum of the weights of all the hyperedges to which each vertex belongs; The The diagonal elements are That is, the first The number of vertices contained in a hyperedge. Calculate the Laplacian matrix of the symmetric normalized hypergraph for a history fragment. ,in To and An identity matrix of the same dimension has a dimension equal to the total number of vertices in the heterogeneous hypergraph of that historical segment. , This is the sum of the number of active task nodes and personnel nodes within that time segment, a dynamic value that varies with the scale of the process. Based on Calculate the historical segment co-entropy value of this time segment. .because For a positive semi-definite matrix, its eigenvalues ​​are... ,therefore The eigenvalue is 1+ The determinant is greater than 0.

[0076] Will With the preset collaborative entropy intervention threshold Compare. If Then, the heterogeneous hypergraph of the historical segments corresponding to that time segment is marked as a low-entropy process segment; if If the segment fails, discard it. Analyze the entropy distribution of each segment across all historically successful projects, and take the upper quartile of this distribution to determine the optimal value. For example, it can be set to 2.0.

[0077] Each tagged low-entropy process segment requires three types of data. The first type is a task graph subgraph for that time segment, extracted from the complete task graph of the successful project, showing the task nodes of active tasks within that time segment and the dependency edges between them. The second type is a snapshot of personnel allocation for that time segment, recording the personnel IDs and roles actually assigned to each active task within that time segment. For example, the task "Structural Calculation Document Compilation" is assigned personnel IDs... The role is that of a structural engineer. The third type of supplementary data is the eigenvalue spectrum distribution of the normalized Laplacian matrix of the historical fragment of the heterogeneous hypergraph. This spectrum distribution is obtained by: applying the constructed symmetric normalized hypergraph Laplacian matrix... Perform eigenvalue decomposition to obtain all eigenvalues ​​and sort them in ascending order; divide the eigenvalue range [0,1] into equal parts. Each interval The default value is 30; the proportion of eigenvalues ​​falling within each interval to the total number of eigenvalues ​​is used to form a vector representation of the spectral distribution. ,in For the first The frequency of characteristic values ​​within each interval.

[0078] In one embodiment of the present invention, the cold start template generation module further includes:

[0079] The template generation unit incorporates a text encoder, a condition generator network, a hypergraph reconstructor network, and a spectral alignment checker.

[0080] The text encoder uses a multi-head self-attention mechanism to extract semantic encoding vectors from the requirements features of a newly launched consulting and design project; the condition generator network uses the semantic encoding vectors as conditional variables and injects conditional information into the hidden layer features through a conditional batch normalization layer to generate target domain representation vectors; the hypergraph reconstructor network maps the target domain representation vectors to the initial heterogeneous hypergraph of the newly launched consulting and design project.

[0081] The spectral alignment checker calculates the Wasserstein distance between the eigenvalue spectrum distribution of the initial normalized Laplacian matrix of the initial heterogeneous hypergraph and the eigenvalue spectrum distribution of the normalized Laplacian matrix of the historical fragments of the heterogeneous hypergraph. When the Wasserstein distance is lower than a preset distance threshold, the hypergraph structure resolver extracts the serial-parallel dependencies of task nodes, the hyperedge connections between tasks and personnel, and the collaborative checkpoints formed by multiple hyperedge connections from the initial heterogeneous hypergraph, and combines them into an initial process template.

[0082] Specifically, the text encoder employs a pre-trained language model based on a multi-head self-attention mechanism. This model has undergone initial training on a general corpus before system deployment, with pre-set weight parameters. It operates directly in inference mode during runtime, without requiring online training. Its input consists of the requirement characteristics of a newly initiated consulting and design project. These characteristics originate from the keywords in the requirement document and the preliminary list of milestone task names filled in during project initiation in the project management tool. For example, the requirement keywords for an architectural design project might be high-rise office building, seismic fortification intensity of 7 degrees, and green building level 2. The text encoder first maps each requirement keyword and task name into a word embedding vector. Then, it calculates the contextual dependencies between words using a multi-head self-attention mechanism. Finally, an average pooling layer aggregates the implicit representations of all words into a fixed-length semantic encoding vector, typically with dimensions of 256 or 512.

[0083] The conditional generator network generates a target domain representation vector using a semantic encoding vector as a conditional variable. Each hidden layer within the network employs a conditional batch normalization layer, injecting conditional information into the distribution parameters of the hidden layer features. The scaling and offset parameters of the traditional batch normalization are output separately from the semantic encoding vector through independent fully connected layers. The dimension of the target domain representation vector is typically 128 or 256.

[0084] The hypergraph reconstructor network maps the target domain representation vector to an initial heterogeneous hypergraph for a newly initiated consulting and design project. The initial heterogeneous hypergraph is a graph structure consisting of an initial vertex set and an initial hyperedge set. The initial vertex set contains all task nodes parsed from the requirement features and all personnel nodes determined from the pool of available personnel. The pool of available personnel, based on the industry type and required skills of the new project, selects all currently eligible personnel and their skill vectors from personnel skill tag data. Each hyperedge in the initial hyperedge set represents a collaborative activity, connecting the task nodes and personnel nodes participating in that activity. The structural information of the initial heterogeneous hypergraph consists of an initial node feature matrix and an initial association matrix. The two matrices provide a complete representation. Each row of the initial node feature matrix corresponds to the feature embedding vector of a vertex, recording the attribute information of that vertex; Each row corresponds to a vertex, and each column corresponds to a hyperedge. If a vertex belongs to a hyperedge, the corresponding element is 1; otherwise, it is 0. This records the topological connection relationship between vertices and hyperedges. The initial node feature matrix and... All are obtained by decoding the target domain representation vector from the hypergraph reconstructor network through a multi-layer graph convolutional decoder.

[0085] The spectral alignment checker is responsible for quality verification of the generated initial heterogeneous hypergraph. This checker first uses the same method as the method used in the low-entropy process fragment acquisition unit to calculate the spectral distribution vector of the normalized Laplacian matrix eigenvalues ​​of the initial heterogeneous hypergraph, thus calculating the spectral distribution vector of the heterogeneous hypergraph. Simultaneously, the spectral alignment checker reads the eigenvalue spectral distribution of the historical segment normalized Laplace matrix of each low-entropy process segment from the accompanying data input from the low-entropy process segment acquisition unit. Then, calculate. With each The Wasserstein distance between the two discrete distributions. and The distance to Vivasserstein is ,in and Two distributions in the th Cumulative frequency over each interval The interval width, The default value is 30. The spectral alignment checker takes the minimum Wasserstein distance between the current initial heterogeneous hypergraph and all low-entropy process segments, and compares it with the preset distance threshold. Compare them. The calibration is performed by statistically analyzing the distribution of the Wasserstan distance between any two low-entropy process segments in historical successful projects, using the upper quartile of this distribution as the default value. If the minimum Wasserstan distance is lower than... If the verification passes, the initial heterogeneous hypergraph is sent to the hypergraph structure resolver; otherwise, the spectral alignment verifier triggers the condition generator network to resample, adjust the random noise at the input, and regenerate the target domain representation vector and the initial heterogeneous hypergraph until the verification passes.

[0086] After successful validation, the hypergraph structure parser extracts three components that constitute the initial process template from the initial heterogeneous hypergraph. The first component is the serial-parallel dependencies between tasks, which are checked... The existence of a directed hyperedge connecting two task nodes determines their relationship. If a directed hyperedge exists, it indicates a sequential dependency, with the direction indicating the execution order. Two task nodes belonging to at least one collaborative activity hyperedge but not connected by a directed hyperedge are considered parallel. The second item is the pre-assigned personnel roles for each task. For each task node, a search is performed... The task node includes all personnel nodes connected by its hyperedges. The cosine similarity of the feature embeddings between the task node and each personnel node is calculated. The personnel skill vectors are checked to ensure they meet the minimum threshold requirement for the task skill requirement vector. The personnel with the highest similarity and meeting the skill requirements are selected as pre-assigned personnel, and their roles are determined based on the common role types they exhibit in historical personnel assignment snapshots. The third item is the preset location of key collaborative checkpoints, which are retrieved. A hyperedge that simultaneously connects at least two task nodes and at least two personnel nodes is identified as a key collaborative checkpoint. The set of task nodes connected to it is extracted as the scope of the covered tasks, and the set of personnel nodes connected to it is extracted as the list of participating roles. The checkpoint is located at the completion time of the last task in the covered task set.

[0087] In one embodiment of the present invention, the collaborative entropy real-time monitoring module includes:

[0088] The collaborative event acquisition unit is used to acquire collaborative events generated within the current time window from the project management log and the collaborative editing record at preset fixed time intervals. Each collaborative event includes the associated task identifier, the identifier of the personnel involved, the event type, the timestamp, the duration of the interaction, and the identifier of the deliverable operated.

[0089] The super-edge merging unit is used to merge collaborative events that are associated with the same task identifier and whose timestamp interval is lower than a preset time interval threshold into a collaborative activity super-edge. Each collaborative activity super-edge connects the task node corresponding to the task identifier and the personnel node corresponding to the personnel identifier.

[0090] The matrix construction unit is used to construct the current vertex set by forming the task nodes corresponding to all task identifiers and the personnel nodes corresponding to all personnel identifiers appearing within the current time window, and to form the current hyperedge set by forming all collaborative activity hyperedges, thus constructing the association matrix. Construct a heterogeneous hypergraph for the current execution flow using the current vertex set and the current hyperedge set. The membership relationships between vertices and hyperedges in the heterogeneous hypergraph for the current execution flow are determined by the incidence matrix. express;

[0091] The weight calculation unit is used to, for each deliverable identified by a deliverable identifier involved in a hyperedge, query the parameter transmission frequency between each deliverable from the pre-stored design structure matrix and take the average as the information coupling degree; for the personnel involved in the hyperedge, obtain the skill vector of each person from the personnel skill tags, and obtain the task requirement vector of the task associated with the hyperedge from the project management log, calculate the cosine similarity between the skill vector and the task requirement vector and take the average as the average experience matching degree; the weight of each hyperedge is calculated by dividing the product of the information coupling degree and the interaction duration by the sum of the average experience matching degree and the preset positive constant, and then using the weight diagonal matrix. express, The diagonal elements are the weights of the corresponding hyperedges.

[0092] Specifically, the collaborative event acquisition unit retrieves collaborative events generated within the current time window from the project management log and collaborative editing records at preset fixed time intervals. A typical fixed time interval is 1 hour. Each collaborative event includes the associated task identifier, the identifiers of the personnel involved, the event type, the timestamp, the duration of the interaction, and the identifier of the delivered item. The definitions of each field are consistent with the definitions of the corresponding fields in the standardized 5-tuple in the data platform module.

[0093] The superedge merging unit receives all collaborative events within the current time window output by the collaborative event acquisition unit, and merges collaborative events with the same task identifier and timestamp intervals lower than a preset time interval threshold into a single collaborative activity superedge. A typical preset time interval threshold is 30 minutes. Each collaborative activity superedge connects the task node corresponding to that task identifier to the personnel nodes corresponding to the personnel identifiers involved in all collaborative events merged into that superedge. If the same personnel node appears multiple times, only one connection is retained.

[0094] The matrix construction unit receives all collaborative activity hyperedges output by the hyperedge merging unit and constructs a heterogeneous hypergraph for the current running process. The construction method is as follows: the current vertex set is formed by the task nodes corresponding to all task identifiers and the personnel nodes corresponding to all personnel identifiers appearing within the current time window; the current hyperedge set is formed by all collaborative activity hyperedges formed through merging. Based on the current vertex set and the current hyperedge set, an association matrix is ​​constructed. The current vertex set contains vertices. The current set of superedges contains 100 superedges. ,but for Matrix, if the first The vertex belongs to the first vertex. If there is a super edge, then Otherwise, it is 0. The heterogeneous hypergraph of the current running process consists of the current set of vertices and the current set of hyperedges.

[0095] The weight calculation unit calculates the weight for each hyperedge, and the weights of all hyperedges constitute the total weight. Weight diagonal matrix , its first The diagonal element is the first Weight of a superedge All off-diagonal elements are 0. The weight of each superedge is... Among them, information coupling degree The average empirical matching degree is obtained by querying the parameter transfer frequency between deliverables involved in the hyperedge from the pre-stored design structure matrix and taking the average value. The mean value is obtained by taking the cosine similarity between the skill vectors of the personnel involved in the hyperedge and the task requirement vector. The duration of the interaction is extracted from the load attribute of the collaborative events contained in the superedge, and the sum is taken when there are multiple events. The default value is 0.01. The calculation method for the above parameters is the same as that for the historical segment superedge weight in the low-entropy process segment acquisition unit, the only difference being that the data source is the real-time collaborative event of the currently running project.

[0096] In one embodiment of the present invention, the collaborative entropy real-time monitoring module further includes:

[0097] Entropy calculation unit, used for calculating entropy based on the correlation matrix and the weight diagonal matrix Calculate the vertex degree matrix and hypermarginality matrix ,in The diagonal elements are the sum of the weights of the hyperedges to which each vertex belongs. The diagonal elements represent the number of vertices contained in each hyperedge; calculate the Laplacian matrix of the symmetric normalized hypergraph. And calculate the collaborative entropy value of the current operating process of the consulting and design project. ,in To and Identity matrices of the same dimension Represents the determinant of a matrix.

[0098] Specifically, the entropy calculation unit first relies on the correlation matrix. and weight diagonal matrix Calculate the vertex degree matrix and hypermarginality matrix Then, the Laplacian matrix of the symmetric normalized hypergraph is calculated according to the formula. and co-entropy The steps for constructing and calculating the above matrices are similar to those in the low-entropy process fragment acquisition unit, which uses the historical fragment association matrix. and historical segment weight diagonal matrix Calculate the collaborative entropy value of historical fragments The steps are the same, so they will not be repeated here.

[0099] Cooperative entropy The output is sent in two directions. First, it's pushed to the project management dashboard, displaying the current level of structural disorder and its changing trends in the form of a time-series curve. Second, when... Exceeding the threshold of collaborative entropy intervention At that time, an intervention request is proactively sent to the entropy regularization dynamic reconstruction module, carrying the current process state vector. The current process state vector consists of three components. The first component is the task completion percentage of each task, obtained by extracting the executed hours and planned total hours of each task at the current time window deadline from the project management log, dividing the executed hours by the planned total hours to obtain the completion percentage for each task; if a task has not yet started, the completion percentage is 0. The second component is the personnel load distribution, obtained by counting the number of different task identifiers associated with each personnel identifier from all collaborative events collected by the collaborative event acquisition unit within the current time window; this number is used as the personnel's load value within the current time window; a load value of 0 indicates that the personnel are not currently assigned any active tasks. The third component is the current collaborative entropy value. .

[0100] In one embodiment of the present invention, the approximate dynamic programming value network in the entropy regularization dynamic reconstruction module By minimizing the loss function Training achieved The formula is:

[0101]

[0102] in, To extract from the completed historical consulting and design project cases according to a preset time step The sampled process state vector, In order to be in The actual actions to be performed To execute The process state vector that is then transferred to the next sampling time. In order to be in Next action The cost of one immediate execution The equivalent cost coefficient per unit of collaborative entropy. for The collaborative entropy value under the following conditions; For the target value network, the parameters of the target value network. Periodic from Replication; Expectation The training sample set is composed of triplets of process state vectors, actions, and process state vectors at the next sampling time, generated from the historical consulting and design project cases.

[0103] Specifically, for each historical consulting and design project case, according to the preset time step... Perform sampling. A typical value is 1 day. At each sampling point, the process state vector for that sampling point is extracted from the standardized collaborative event sequence and project management log contained in the unified data object corresponding to the historical consulting and design project case. The process state vector consists of three components: the task completion ratio of each task, the personnel workload distribution of each person, and the collaboration entropy value at that moment. The task completion ratio is obtained from the project management log as the ratio of executed hours to the planned total hours. The personnel workload distribution is obtained by counting the number of different tasks associated with each person in the collaboration events within the previous time window. The collaboration entropy value is calculated by the collaboration entropy real-time monitoring module in historical playback mode based on the project data up to that moment. Simultaneously, the actions actually performed at that sampling moment are obtained from the project management log of this historical consulting and design project case. This action refers to actual process adjustment operations recorded in the project management log, such as task reordering, personnel reassignment, or checkpoint insertion actually performed by managers. The immediate execution cost of this action... The result is the sum of two parts: basic working time loss determined by the action type and personnel changeover cost. Both basic working time loss and personnel changeover cost are statistically determined from the working time records and personnel allocation records of historical projects. The execution data is obtained from the unified data object of this historical consulting and design project case. The process state vector at the next sampling time .Will Store it as a training sample triple in the training sample set.

[0104] Value Network It is a fully connected neural network. This is the set of weight and bias parameters for the network; the network input is the process state vector. The output is a scalar, representing the value from... The estimated long-term cumulative cost of starting with the optimal strategy. parameters By minimizing the loss function Conduct training. middle, This represents the expectation of all triples in the training sample set. The equivalent cost coefficient of unit collaborative entropy. The calibration method is as follows: During periods when the collaborative entropy value in historical projects exceeded the intervention threshold, the project delay duration and rework volume were statistically analyzed. The sum of the equivalent time cost calculated from the delay duration and the equivalent time cost calculated from the rework volume was divided by the increase in the collaborative entropy value during that period, and the average of all historical statistical results was taken as the calibrated value. For example, it can be calibrated as the cost of 4 people per unit of entropy. Process state vector The cooperative entropy value under the given conditions. For value network The output value, For target value network right The output value. Network structure and Completely identical, parameters Periodic from The replication cycle is typically executed once every 100 training iterations.

[0105] During training, for each triple in the training sample set, the sum of the immediate execution cost and the state penalty term is calculated once. In addition to the target value network Output value Subtracting the value network's impact Output value The time-series difference error is obtained; the square of the time-series difference error is used as the loss of the triplet, and the average of the losses of all triplets is taken to obtain the result. Updated using gradient descent. ,make Gradually reduce. Training continues until... Convergence is achieved when the mean of the loss function decreases by less than a preset convergence threshold for value network training over several consecutive training rounds. , The value is set to one percent of the mean of the loss function in the early stages of training. For example, if the mean loss in the early stages of training is 10.0, then... Take 0.1.

[0106] In one embodiment of the present invention, the entropy regularization dynamic reconstruction module includes:

[0107] The entropy regularization dynamic reconstruction module, during online decision-making, selects from a preset action set... The optimal control action is selected from the candidates, and for each candidate action... Calculate the expected long-term cumulative cost The calculation formula is:

[0108]

[0109] in, This is the current process state vector. In the state Next action The cost of executing a single action The unit collaborative entropy equivalent cost coefficient is... This represents the collaborative entropy value of the current running process. For the preset time step, This is the pre-trained approximate dynamic programming value network. To perform the action The next state vector after transition, Let be the expectation of the state transition distribution;

[0110] Elected The smallest candidate action is selected as the optimal control action. ,Right now .

[0111] Specifically, construct the current process state vector. The current process state vector consists of three components. The method of obtaining each component is the same as the method of constructing the state vector in the training samples during the pre-training of the value network. The only difference is that the data source is the real-time data of the currently running consulting and design project rather than historical archived data.

[0112] From the pre-set process intervention action set Iterate through each candidate action Process intervention action set Several executable process structure modification operations are pre-defined, each corresponding to a specific topology adjustment or personnel configuration change method. Typical candidate actions include: changing two parallel tasks to a serial relationship, changing two serial tasks to a parallel relationship, adding a reviewer to a specified task, transferring a task from one person to another, and inserting a mandatory alignment meeting between two stages. The action set can be expanded or reduced according to the management standards of consulting and design firms.

[0113] For each candidate action Calculate the expected long-term cumulative cost ,exist In the calculation formula, In order to be in Next action The cost of action execution is obtained by adding the basic working time loss determined by the action type and the personnel switching overhead, which is the same as the calculation method of the immediate execution cost in the pre-training stage of the value network. This represents the collaborative entropy value of the current running process. To perform the action The state vector that is then transferred to the next process state vector. Let the expected value of the state transition distribution be denoted as . Statistical analysis of historical data shows that: clustering is performed on all the next states transitioned to after performing the same type of action in similar states in the historical training sample set, and the cluster center is used as the expected value, or the transition result is fitted with a Gaussian distribution and the mean is taken.

[0114] Complete the action set After considering all candidate actions, select the one that makes the action. The smallest candidate action is selected as the optimal control action. Optimal control action These are converted into specific execution instructions and pushed to the project management tool via an instruction distribution interface. The mapping between instruction types and actions is pre-defined in the action set. For example, if... To "the mission" With the task "Change from parallel to serial" means the instruction dispatch interface sends instructions to the project management tool. Call, modify task With the task The dependencies between tasks are changed from the original start-start constraint to a finish-start constraint, and the affected task owners are notified. For the "task" If "add reviewers" is selected, the task assignment table will be automatically updated and task cards will be pushed to the newly added personnel.

[0115] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A consulting design process control system, characterized in that, Includes the following modules: The data platform module is used to collect and standardize the processing of multi-source heterogeneous data throughout the entire lifecycle of consulting and design projects. The multi-source heterogeneous data includes at least project management logs, collaborative editing records, personnel skill tags, and structured records of historical consulting and design project cases. The cold start template generation module is used to extract low-entropy process segments with synergistic entropy values ​​lower than the preset synergistic entropy intervention threshold from the historical consulting and design project cases when a new consulting and design project is initiated; and to generate an initial process template with synergistic entropy values ​​lower than the preset synergistic entropy intervention threshold based on the structural characteristics of the low-entropy process segments and the demand characteristics of the new consulting and design project, using the structural characteristics of the low-entropy process segments as a reference. The initial process template includes the serial and parallel dependencies between tasks, the pre-assigned personnel roles for each task, and the preset locations of key collaborative checkpoints; The collaborative entropy real-time monitoring module is used to collect collaborative events at fixed time intervals during the operation of the consulting and design project, construct a heterogeneous hypergraph in real time, and calculate the collaborative entropy value of the current operation process of the consulting and design project based on the normalized Laplace matrix of the heterogeneous hypergraph. The entropy regularization dynamic reconstruction module has a built-in pre-trained approximate dynamic programming value network and a preset set of process intervention actions. When the collaborative entropy value of the current operation process of the consulting and design project exceeds the collaborative entropy intervention threshold, a current process state vector is constructed based on the task completion ratio, personnel load distribution, and the collaborative entropy value of the current operation process. Using the approximate dynamic programming value network, for each candidate action in the process intervention action set, the expected long-term cumulative cost consisting of the action execution cost and the collaborative entropy value as a state penalty term is calculated. The candidate action with the minimum expected long-term cumulative cost is selected as the optimal control action, and an execution instruction is output to change the topology and personnel configuration of the current operation process of the consulting and design project.

2. The consulting design process control system according to claim 1, characterized in that, The data platform module includes a standardization engine, a skill vector encoder, and a historical case reconstructor. The standardization engine includes a verb mapping table and a quintuple constructor. The verb mapping table stores the mapping relationship between action verbs from different sources and their corresponding verbs in the standard action set. The quintuple constructor outputs three types of standardized quintuples based on the mapping relationship. All three types of quintuples carry a timestamp and a payload attribute. The skill vector encoder includes a fixed-dimensional skill classification system and a collaborative filtering inference engine. The collaborative filtering inference engine is connected to the skill classification system and infers and completes the missing dimension values ​​in the skill vectors generated based on the skill classification system. The historical case refactoring unit includes a task graph structuring unit, an event sorting unit, and an encapsulation unit. The task graph structuring unit transforms the task decomposition structure in historical consulting and design project cases into a task graph composed of task nodes and dependency edges. The event sorting unit arranges the collaborative event records in historical consulting and design project cases into a standardized collaborative event sequence in ascending order of timestamps. The encapsulation unit encapsulates the task graph, the standardized collaborative event sequence, and the performance tags of historical consulting and design project cases into a unified data object.

3. The consulting design process control system according to claim 1, characterized in that, The cold start template generation module includes: The low-entropy process segment acquisition unit is used to select cases with performance tags as successful projects from the historical consulting and design project cases, divide each successful project case into continuous time segments according to a fixed time window, construct a historical segment heterogeneous hypergraph for each time segment and calculate its historical segment synergistic entropy value, and mark the historical segment heterogeneous hypergraph corresponding to the time segment with a historical segment synergistic entropy value lower than the synergistic entropy intervention threshold as a low-entropy process segment; each low-entropy process segment is accompanied by its task graph subgraph, personnel allocation snapshot, and the eigenvalue spectrum distribution of the historical segment normalized Laplace matrix of the historical segment heterogeneous hypergraph.

4. The consulting design process control system according to claim 3, characterized in that, The cold start template generation module further includes: The template generation unit incorporates a text encoder, a condition generator network, a hypergraph reconstructor network, and a spectral alignment checker. The text encoder uses a multi-head self-attention mechanism to extract semantic encoding vectors from the requirements features of a newly launched consulting and design project; the condition generator network uses the semantic encoding vectors as conditional variables and injects conditional information into the hidden layer features through a conditional batch normalization layer to generate target domain representation vectors; the hypergraph reconstructor network maps the target domain representation vectors to the initial heterogeneous hypergraph of the newly launched consulting and design project. The spectral alignment checker calculates the Wasserstein distance between the eigenvalue spectrum distribution of the initial normalized Laplacian matrix of the initial heterogeneous hypergraph and the eigenvalue spectrum distribution of the normalized Laplacian matrix of the historical fragments of the heterogeneous hypergraph. When the Wasserstein distance is lower than a preset distance threshold, the hypergraph structure resolver extracts the serial-parallel dependencies of task nodes, the hyperedge connections between tasks and personnel, and the collaborative checkpoints formed by multiple hyperedge connections from the initial heterogeneous hypergraph, and combines them into an initial process template.

5. The consulting design process control system according to claim 1, characterized in that, The collaborative entropy real-time monitoring module includes: The collaborative event acquisition unit is used to acquire collaborative events generated within the current time window from the project management log and the collaborative editing record at preset fixed time intervals. Each collaborative event includes the associated task identifier, the identifier of the personnel involved, the event type, the timestamp, the duration of the interaction, and the identifier of the deliverable operated. The super-edge merging unit is used to merge collaborative events that are associated with the same task identifier and whose timestamp interval is lower than a preset time interval threshold into a collaborative activity super-edge. Each collaborative activity super-edge connects the task node corresponding to the task identifier and the personnel node corresponding to the personnel identifier. The matrix construction unit is used to construct the current vertex set by forming the task nodes corresponding to all task identifiers and the personnel nodes corresponding to all personnel identifiers appearing within the current time window, and to form the current hyperedge set by forming all collaborative activity hyperedges, thus constructing the association matrix. Construct a heterogeneous hypergraph for the current execution flow using the current vertex set and the current hyperedge set. The membership relationships between vertices and hyperedges in the heterogeneous hypergraph for the current execution flow are determined by the incidence matrix. express; The weight calculation unit is used to, for each deliverable identified by a deliverable identifier involved in a hyperedge, query the parameter transmission frequency between each deliverable from the pre-stored design structure matrix and take the average as the information coupling degree; for the personnel involved in the hyperedge, obtain the skill vector of each person from the personnel skill tags, and obtain the task requirement vector of the task associated with the hyperedge from the project management log, calculate the cosine similarity between the skill vector and the task requirement vector and take the average as the average experience matching degree; the weight of each hyperedge is calculated by dividing the product of the information coupling degree and the interaction duration by the sum of the average experience matching degree and the preset positive constant, and then using the weight diagonal matrix. express, The diagonal elements are the weights of the corresponding hyperedges.

6. The consulting design process control system according to claim 5, characterized in that, The collaborative entropy real-time monitoring module also includes: Entropy calculation unit, used for calculating entropy based on the correlation matrix and the weight diagonal matrix Calculate the vertex degree matrix and hypermarginality matrix ,in The diagonal elements are the sum of the weights of the hyperedges to which each vertex belongs. The diagonal elements represent the number of vertices contained in each hyperedge; calculate the Laplacian matrix of the symmetric normalized hypergraph. And calculate the collaborative entropy value of the current operating process of the consulting and design project. ,in To and Identity matrices of the same dimension Represents the determinant of a matrix.

7. The consulting design process control system according to claim 1, characterized in that, The approximate dynamic programming value network in the entropy regularization dynamic reconstruction module By minimizing the loss function Training achieved The formula is: in, To extract from the completed historical consulting and design project cases according to a preset time step The sampled process state vector, In order to be in The actual actions to be performed To execute The process state vector that is then transferred to the next sampling time. In order to be in Next action The cost of one immediate execution The equivalent cost coefficient per unit of collaborative entropy. for The collaborative entropy value under the following conditions; For the target value network, the parameters of the target value network. Periodic from Replication; Expectation The training sample set is composed of triplets of process state vectors, actions, and process state vectors at the next sampling time, generated from the historical consulting and design project cases.

8. The consulting design process control system according to claim 7, characterized in that, The entropy regularization dynamic reconstruction module includes: The entropy regularization dynamic reconstruction module, during online decision-making, selects from a preset action set... The optimal control action is selected from the candidates, and for each candidate action... Calculate the expected long-term cumulative cost The calculation formula is: in, This is the current process state vector. In the state Next action The cost of executing a single action The unit collaborative entropy equivalent cost coefficient is... This represents the collaborative entropy value of the current running process. For the preset time step, This is the pre-trained approximate dynamic programming value network. To perform the action The next state vector after transition, Let be the expectation of the state transition distribution; Elected The smallest candidate action is selected as the optimal control action. ,Right now .

Images