Intelligent task decomposition and optimization method and system based on deep learning

Abstract

The present application relates to the technical field of resource management, in particular to an intelligent task decomposition and optimization method and system based on deep learning, comprising the following steps: based on the project task decomposition list and the collaboration record, the dependency relationship of each task is identified and the node connection is established, and then according to the human resource collaboration communication record, the flow direction of problem seeking help and solving is extracted. The present application accurately depicts the knowledge dependency degree between tasks in the coupling mode of tasks and expert resources by establishing the close association between the task node dependency relationship and the problem seeking help flow direction of human resource collaboration, improves the matching accuracy between task decomposition and resource allocation, reduces the risk of idle or overload of expert resources during task implementation; through the deep traversal of the network path dependence connection, the transmission influence of task load fluctuation on the execution time of downstream tasks is quantified, the real propagation law of load disturbance between task nodes is revealed, and the resource allocation is more in line with the actual demand of tasks.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of resource management, in particular to an intelligent task decomposition and optimization method and system based on deep learning. BACKGROUND

[0002] The technical field of resource management refers to a technical system that takes human resources, equipment, knowledge, funds, time and other resources involved in project execution as the core research object, uses system analysis, optimization algorithm, information technology and other methods to accurately configure and dynamically adjust the allocation strategy of resources, so as to improve resource utilization efficiency, reduce cost risk, and realize the optimization of project management efficiency.

[0003] The prior art adopts a static task decomposition mode and an empirical resource allocation method in actual operation, which leads to rough identification of the dependency relationship between tasks and difficulty in dynamically adjusting the dependency relationship according to the actual execution situation, difficulty in accurately measuring the knowledge demand matching degree between tasks and expert resources, and easy imbalance of utilization of expert resources. At the same time, the specific quantitative analysis of the influence of task load variation on subsequent tasks is often ignored, so that the sudden disturbance generated during project execution is difficult to identify and effectively handle in time, resulting in a decrease in resource utilization efficiency. Therefore, improvement is needed. SUMMARY

[0004] The purpose of the present application is to solve the shortcomings in the prior art, and an intelligent task decomposition and optimization method and system based on deep learning are proposed.

[0005] In order to achieve the above-mentioned purpose, the present application adopts the following technical scheme, an intelligent task decomposition and optimization method based on deep learning, comprising the following steps:

[0006] Based on the project task decomposition list and the collaboration record, the dependency relationship of each task is identified and the node connection is established, and then the flow direction of problem seeking help and solving is extracted according to the human resource collaboration communication record, and a task knowledge coupling network is established;

[0007] Based on the task knowledge coupling network, the dependency connection of each execution path in the network is traversed to generate a dependency influence strength set, and the transmission influence value is taken as the weight of the corresponding dependency edge of the task knowledge coupling network based on the dependency influence strength set, and a project critical path weight graph is constructed;

[0008] Based on the project critical path weight graph and the set time window, a unit delay disturbance is applied to each task node in turn, and a full node disturbance propagation effect set is obtained, and based on the full node disturbance propagation effect set, a statistical expectation is calculated to obtain a project structure disturbance sensitivity index;

[0009] According to the project structure disturbance sensitivity index, an abnormally sensitive task decomposition node is identified, the required expert resource is determined by calling the task knowledge coupling network, the combined cost of task execution time, expert waiting time and path sensitivity under different scheduling schemes is calculated, a candidate execution sequence cost list is generated, and the sequence with the lowest combined cost is selected as the knowledge resource optimization scheduling sequence based on the candidate execution sequence cost list.

[0010] Preferably, the task knowledge coupling network acquisition step is:

[0011] Based on the project task decomposition list and the collaboration record, the task identifiers of all tasks in the project task decomposition list are extracted one by one, the direct dependency order between each task identifier and other task identifiers is matched and determined according to the dependency relationship between the task identifiers explicitly marked in the collaboration record, and the task node dependency relationship is obtained.

[0012] Based on the task node dependency relationship, the problem seeking and problem solving records in the human resource collaboration communication record are called, the identity information of the problem initiator and the problem solver is parsed from each communication record, the direction of the problem from the initiator to the solver is identified according to the communication time order appearing in the communication record, the problem flow direction in the communication record is determined one by one, and the problem seeking and solving flow direction is obtained.

[0013] Based on the problem seeking and solving flow direction, the expert identity information of the expert resource as the problem solver in the communication record is extracted, the task nodes associated in the problem seeking and solving flow direction are associated with the corresponding expert identity information one by one, the connection relationship between the task nodes and the expert resource nodes is gradually established, and the task knowledge coupling network is formed.

[0014] Preferably, the dependency influence strength set acquisition step is:

[0015] Based on the task knowledge coupling network, a deep learning process is used to call each complete execution path from the starting task node to the terminal task node in the task knowledge coupling network one by one, traverse all the task nodes on the execution path, extract the work load state data of each task node one by one, calculate the absolute value of the difference between the work load state data of each task node and the work load state data of the previous task node in the execution path, and form a set of absolute values of the load state differences between the nodes.

[0016] Based on the set of absolute values of the inter-node load state difference, according to the direction of the dependent connection between each task node in the task knowledge coupled network, the influence degree of each absolute value of the load state difference on the execution time of the directly adjacent downstream task node is analyzed item by item, the numerical relationship between each absolute value of the load state difference and the change of the execution time of the corresponding downstream task node is quantified respectively, and the inter-node transmission influence value is calculated to form the set of dependent influence strengths.

[0017] Preferably, the step of obtaining the project critical path weight graph is:

[0018] Based on the set of dependent influence strengths, each inter-node transmission influence value in the set of dependent influence strengths is matched with the dependent connection relationship between the corresponding task nodes in the task knowledge coupled network, the inter-node transmission influence value is assigned to the weight of the dependent connection between the corresponding task nodes, the dependent connection weight of the task knowledge coupled network is updated item by item, and the project critical path weight graph is formed.

[0019] Preferably, the step of obtaining the set of full-node disturbance propagation effects is:

[0020] Based on the project critical path weight graph and the set time window, the original execution time of all task nodes in the project critical path weight graph is extracted in turn, a fixed delay disturbance is applied to each task node, the task number of the disturbed node, the disturbance injection time and the corresponding disturbance duration are recorded, and a task node delay disturbance list is generated;

[0021] According to the task node delay disturbance list, the propagation path of each disturbance to the downstream task node is simulated, all affected task nodes are traversed in topological order, and whether all previous task nodes are completed and meet the resource ready condition is judged when each task node is encountered in the simulation. If any previous task is not completed or the resource is not released, the earliest start time of the current task is adjusted to the maximum value of the latest completion time of the previous task and the earliest available time of the current resource. The difference between the latest completion time of the previous task and the earliest available time of the current resource and the original planned start time is the cumulative delay of the task. The disturbance propagation delay of all task nodes is accumulated to generate the set of full-node disturbance propagation effects.

[0022] Preferably, the step of obtaining the project structure disturbance sensitivity index is:

[0023] Based on the set of full-node disturbance propagation effects, the project structure disturbance sensitivity index is calculated.

[0024] Preferably, the step of obtaining the candidate execution sequence cost list is:

[0025] Based on the project structure disturbance sensitivity index, the task decomposition nodes with the disturbance sensitivity index value greater than the median value of the overall task node disturbance sensitivity are screened item by item, the task number, execution dependency path and resource allocation record of the corresponding task decomposition node are extracted, the expert resource information bound to each task node is retrieved by calling the task knowledge coupling network, and a high-sensitivity task and expert resource pairing list is obtained;

[0026] Based on the high-sensitivity task and expert resource pairing list, the combination cost is calculated.

[0027] Based on the combination cost, the cost of all task execution permutation and combination schemes is compared item by item, the task permutation scheme with the minimum combination cost is selected and the task order and expert arrangement are extracted, and a candidate execution sequence cost list is generated.

[0028] Preferably, the knowledge resource optimization scheduling sequence acquisition step is:

[0029] Based on the candidate execution sequence cost list, the combination cost corresponding to all candidate execution sequences in the candidate execution sequence cost list is read item by item, the candidate execution sequence combination cost is taken as the comparison object, the combination cost values of all candidate execution sequences are compared item by item, the minimum value in the combination cost values of all candidate execution sequences is extracted, and the optimal combination cost value is formed.

[0030] Based on the optimal combination cost value, the candidate execution sequence cost list is traversed item by item, the absolute value of the difference between each candidate execution sequence combination cost value and the optimal combination cost value is compared item by item, it is judged whether the absolute value of the difference is zero, if the absolute value of the difference is zero, the corresponding candidate execution sequence is marked as the optimal execution sequence, the task order and expert resource arrangement information of the optimal execution sequence are extracted, and the optimal execution sequence information is formed.

[0031] Based on the optimal execution sequence information, the task order and expert resource arrangement information in the optimal execution sequence information are read item by item, the execution order of each task node and expert resource is determined in turn according to the task order, and the connection relationship of the corresponding task node and expert resource in the task knowledge coupling network is updated, and the knowledge resource optimization scheduling sequence is formed.

[0032] The application also provides an intelligent task decomposition and optimization system, comprising:

[0033] A task dependency recognition module recognizes the dependency relationship of each task based on the project task decomposition list and the collaboration record, establishes node connection, and then extracts the flow direction of problem seeking and solving according to the human resource collaboration communication record, and establishes a task knowledge coupling network.

[0034] A critical path modeling module traverses the dependent connections of each execution path in the network based on the task knowledge coupling network, generates a dependent influence strength set, and assigns a transmission influence value as a weight to the corresponding dependent edge of the task knowledge coupling network based on the dependent influence strength set, to construct a project critical path weight graph;

[0035] A disturbance sensitivity analysis module applies a unit delay disturbance to each task node in turn based on the project critical path weight graph and a set time window, obtains a full node disturbance propagation effect set, and calculates a statistical expectation based on the full node disturbance propagation effect set to obtain a project structure disturbance sensitivity index;

[0036] A scheduling optimization module identifies an abnormally sensitive task decomposition node according to the project structure disturbance sensitivity index, calls the task knowledge coupling network to determine the required expert resources, calculates the combined cost of task execution time, expert waiting time and path sensitivity under different scheduling schemes, generates a candidate execution sequence cost list, and selects the sequence with the lowest combined cost as the knowledge resource optimized scheduling sequence based on the candidate execution sequence cost list.

[0037] Compared with the prior art, the advantages and positive effects of the present application are that:

[0038] The present application precisely depicts the knowledge dependency degree between tasks in a task and expert resource coupling manner by establishing a close association between task node dependency relationships and human resource cooperation problem help flow, improves the matching accuracy between task decomposition and resource allocation, and reduces the risk of idle or overload of expert resources during task implementation. The present application quantifies the transmission influence of task load fluctuation on downstream task execution time through deep traversal of network path dependent connections, reveals the real propagation law of load disturbance between task nodes, and makes resource allocation more in line with actual task needs. The present application identifies sensitive nodes that have a greater impact on project structure stability by introducing a unit delay disturbance into task nodes and statistically analyzing the cumulative influence of disturbance transmission, significantly improves the pertinence and reliability of resource scheduling. The present application integrates task execution time, expert waiting time and path sensitivity into a unified cost index for comprehensive evaluation, making the comparison of different schemes more comprehensive and reasonable, achieving high synergy of resource optimization and risk control, and improving overall resource utilization efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 The present application is a step schematic diagram. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0041] Referring to Figure 1 The present application provides a technical solution, an intelligent task decomposition and optimization method based on deep learning, comprising the following steps:

[0042] Based on the project task decomposition list and collaboration records, the dependency relationship of each task is identified and the node connection is established, and then according to the human resource collaboration communication records, the problem seeking and solving flow is extracted, and a task knowledge coupling network is established;

[0043] Based on the task knowledge coupling network, the dependency connection of each execution path in the network is traversed to generate a dependency influence strength set, and based on the dependency influence strength set, the transmission influence value is taken as the weight of the corresponding dependency edge of the task knowledge coupling network to construct a project critical path weight graph;

[0044] Based on the project critical path weight graph and the set time window, a unit delay disturbance is applied to each task node in turn to obtain a full node disturbance propagation effect set, and based on the full node disturbance propagation effect set, a statistical expectation is calculated to obtain a project structure disturbance sensitivity index;

[0045] According to the project structure disturbance sensitivity index, the abnormal sensitivity task decomposition node is identified, the required expert resources are determined by calling the task knowledge coupling network, the combined cost of task execution time, expert waiting time and path sensitivity under different scheduling schemes is calculated, a candidate execution sequence cost list is generated, and the sequence with the lowest combined cost is selected as the knowledge resource optimization scheduling sequence based on the candidate execution sequence cost list.

[0046] The acquisition steps of the task knowledge coupling network are:

[0047] Based on the project task decomposition list and collaboration records, the task identification of all tasks in the project task decomposition list is extracted item by item, the direct dependency order between each task identification and other task identifications is matched and determined item by item according to the dependency relationship between the task identifications in the collaboration records, and the task node dependency relationship is obtained.

[0048] Based on the task node dependency relationship, the problem seeking and problem solving records in the human resource collaboration communication records are called, the identity information of the problem initiator and the problem solver is parsed from each communication record, the direction of the problem from the initiator to the solver is identified according to the communication time sequence in the communication record, the problem flow in the communication record is determined item by item, and the problem seeking and solving flow is obtained.

[0049] Based on the problem seeking solution flow, the expert identity information of the expert resource as the problem solving party in the communication record is extracted, the task nodes associated in the problem seeking solution flow are knowledge dependency associated with the corresponding expert identity information item by item, and the connection relationship between the task nodes and the expert resource nodes is gradually established, forming a task knowledge coupling network.

[0050] Specifically, based on the project task breakdown list and collaboration records, first, the structured data of the project task breakdown list is parsed, which usually contains task identification, task name, person in charge, planned start and end date, etc. Fields, create a separate task node for each task (for example, T001: database architecture design, T002: user authentication module development), and all extracted task identifications (such as T001, T002, …, T105) are collected into an initial task node set. Then, the system starts a parsing engine configured with specific pattern matching rules to scan the text information in the collaboration records, which come from the comment area of the project management tool (such as Jira or Confluence), the submission information of the code version control system (such as Git), or the channel history of the instant messaging tool (such as Slack). The parsing engine locates and captures the explicitly declared dependency keywords and the task identifications that follow them by applying regular expressions, such as (?i)(depends on|blocked by|waits for)[\s:]*([A-Z]+\d+). For each successfully matched record, such as the collaboration record of task T002 showing "blocked by T001", the system immediately generates a directed dependency pair (T001, T002), indicating that T001 is the prerequisite task of T002. During the matching process, if dependency conflicts occur, such as record A showing (T001, T002) and record B showing (T002, T001), the system will adopt a timestamp-based conflict resolution mechanism, that is, the dependency relationship indicated by the record with the latest timestamp is preferred, and the conflicting dependency pairs are recorded in a log for manual review by the project manager. After all collaboration records are scanned and a preliminary set of dependency pairs is generated, the system constructs an initial directed graph data structure, where task identifications are graph vertices and dependency pairs are directed edges. Subsequently, the system performs a depth-first search-based loop detection algorithm to traverse all possible paths in the graph to identify logically invalid circular dependencies (for example, T001 -> T003 -> T005 -> T001). Once a loop is detected, the system automatically marks all dependency edges forming the loop as "pending" and temporarily removes them from the current valid dependency graph, while generating a warning report containing loop information. Finally, after the above extraction, matching, conflict resolution, and loop detection processes, the system integrates all valid, acyclic dependency pairs to obtain the task node dependency relationship.

[0051] Based on the task node dependency relationship, the system calls human resource collaboration communication records, which contain timestamps, sender identities, receiver identities, and message contents, to accurately identify the problem seeking and solution flow. The system first deploys a pre-trained language model of BERT, which is fine-tuned on question and answer corpus in the target field (such as software development, IT operation), and is specially used for intent classification and entity recognition. For each communication record, the model first analyzes its content and classifies it into one of the four intents: "problem seeking", "problem solving", "status update", or "other". The confidence threshold for intent classification is set to 0.75, which is determined by testing on a validation set containing 5000 manually labeled communication records and selecting the point with the best macro-average F1 score. Records below this threshold are labeled as "ambiguous intent" and skipped. For records classified as "problem seeking", the system identifies the sender as the "problem initiator" and uses named entity recognition to extract "potential solvers" from the message content and receiver list. For example, in the message "@Zhang San, I encountered a database connection timeout error when executing task T045, and the log is attached, can you help me check it?", "Zhang San" is identified as a potential solver and task T045 is associated. Then, the system tracks subsequent communications related to the problem, looking for response messages classified as "problem solving" by the model and published by "potential solvers" or other participants. For example, Zhang San replied "Check your configuration file, the database address should be db.prod.internal, not db.dev.internal". This message will be marked as a solution. The system constructs a dialogue chain in chronological order of communication timestamps, starting from the initial help from the problem initiator to the first response confirmed as "problem solving", to determine the direction of the problem from the initiator to the solver. If a help message is replied to by multiple experts, the system will identify the expert who provides the final adopted solution (e.g., the initiator replies "@Zhang San's method solves the problem, thank you!") as the final "problem solver". When it is not clear whether the solution is adopted, the first respondent who provides an effective solution is identified as the solver. The system records these relationships in a structured manner and finally aggregates all identified high-confidence interactions to obtain the problem seeking and solving flow.

[0052] Based on the problem solving solution flow, the system first traverses each successful interaction instance recorded in the flow data, such as (task T045, initiator: Li Si, solver: Zhang San), and extracts the expert identity information as the "problem solver", that is, "Zhang San". Here, the expert resource not only refers to a senior engineer, but also refers to any project member who provides key knowledge or solutions on a specific task. Subsequently, the system begins to construct a heterogeneous network graph containing two types of nodes: one is the existing task node (such as T045), and the other is the newly added expert resource node (such as Zhang San). In order to quantify the knowledge dependency strength of the task node on the expert resource, the system introduces a knowledge contribution weight calculation method. This weight is not simply based on the number of interactions, but is a comprehensive consideration of problem complexity and solution efficiency. The specific calculation method is as follows: for each problem i solved by expert E for task T, the contribution value is calculated as where the problem complexity has a value range of [1, 5], and is scored based on the description length, technical term density, and dialogue interaction rounds of the problem, for example, a simple configuration inquiry is scored as 1, while a complex defect that requires code debugging is scored as 5. The solution duration is the time interval from the problem being raised to the appearance of the first effective solution, measured in hours. In order to prevent the weight from being too large due to a short duration, a smoothing term of 0.1 is added to the denominator. For example, if Zhang San solves a problem with a complexity of 4 for T045 in 2 hours, the contribution value for this instance is The total knowledge dependency connection weight between task node T and expert node E is the sum of the contribution values of all problems solved by the expert for the task, that is, If Zhang San also solves another problem with a complexity of 2 for T045 in 0.5 hours, the total weight is The system processes all problem solving solution flow records item by item, assigns the calculated weight to the directed edge from the corresponding task node to the expert resource node, and these newly added knowledge dependency edges together with the task interdependence edges obtained in the first step form a more complete network, finally forming a task knowledge coupling network.

[0053] The acquisition steps of the dependency influence strength set are:

[0054] Based on the task knowledge coupling network, a deep learning process is used to call each complete execution path from the starting task node to the terminating task node in the task knowledge coupling network, traverse all task nodes on the execution path, and extract the work load state data of each task node item by item. The absolute value of the difference between the work load state data of each task node and the work load state data of the previous task node in the execution path is calculated to form a set of absolute values of the load state difference between nodes.

[0055] Based on the set of absolute values of the load state difference between nodes, the influence of each absolute value of the load state difference on the execution time of the directly adjacent downstream task node is analyzed in terms of the direction of the dependent connection between the task nodes in the task knowledge coupled network, and the numerical relationship between each absolute value of the load state difference and the change in the execution time of the corresponding downstream task node is quantified, and the transfer influence value between nodes is calculated to form a set of dependence influence strengths.

[0056] Specifically, based on the task knowledge coupled network, the system first identifies and enumerates all complete execution paths in the network from the starting task node without preceding dependencies to the terminating task node without subsequent dependencies using a depth-first search algorithm, forming a path list, and then starts a graph attention network (GAT)-based deep learning model for reasoning. This model is specifically designed to predict the dynamic workload of each task node. The model structure includes an input layer, two graph attention layers, and an output layer. The input layer receives the initial feature vector of each task node, which is composed of the static attributes (such as planned working hours, required skill level, priority) and dynamic attributes (such as allocated manpower, code submission frequency, number of associated issues) of the task. The two graph attention layers update the representation of each node by weighted aggregation of neighbor node information, where the attention weights are automatically learned by the model based on the relationship between nodes. The output layer is a fully connected layer that finally outputs a "workload state data" vector with a dimension of 4, which contains four key indicators: computational resource consumption index, manpower saturation, communication interaction frequency, and task backlog. The model is trained on historical data from more than 50 completed projects. The training data is a snapshot of the project at each time point (e.g., daily), the input is the network state at time point T, and the target is the actual observed workload state data at time point T+1. The mean squared error is used as the loss function, and the Adam optimizer is used for training with a learning rate of 0.001. For the current project, the system traverses each complete execution path in the path list, such as path P1: T01 -> T03 -> T08. For each task node on the path (such as T03), the system calls the pre-trained GAT model, inputs T03 and its neighbor nodes in the task knowledge coupled network (including its preceding task T01 and associated expert resource E05) current features, and infers the "workload state data" of T03, such as [85.2, 0.9, 5.0, 2.0]. Similarly, the system also obtains the workload state data of its direct preceding task node (T01), such as [40.5, 0.5, 2.0, 1.0]. Then, the system calculates the absolute value of the difference between the two vectors. The specific calculation method is the weighted sum of the absolute values of the differences in each corresponding dimension, and the calculation formula is where is the workload vector of the current node, is the workload vector of the predecessor node, the weight vector (e.g., [0.4, 0.3, 0.2, 0.1]) is derived from historical data analysis and reflects the relative importance of each load indicator on the subsequent task, for example, the change in computing resource consumption is considered to have the greatest impact, so the weight is the highest. Apply the above calculation process to all adjacent node pairs on the path, and collect all the absolute values of the calculated differences (for example, the difference between T01 and T03 is 21.5), and finally form a set of absolute values of the load state differences between nodes.

[0057] Based on the set of absolute values of inter-node load state difference, the system matches each absolute value of load state difference to its corresponding downstream task node according to the defined task node dependency connection direction in the task knowledge coupled network, and starts a feedforward neural network (FFN) model to quantify the impact, which contains an input layer, two hidden layers and an output layer. The input layer receives a feature vector composed of five parts: the absolute value of inter-node load state difference, the complexity score of the upstream task node (a value from 1 to 10 based on the number of its subtasks and the evaluation of technical novelty), the resource sensitivity of the downstream task node (a value from 0 to 1 indicating the degree of its performance dependence on a specific resource such as a senior expert), the planned slack time of the downstream task node (in days), and the average response time of the expert resource connecting the upstream and downstream tasks. The two hidden layers have 16 and 8 neurons respectively, both using ReLU activation function. The output layer is a single neuron using linear activation function, which directly outputs a predicted value, which is the expected change in the execution time of the downstream task node (in days). This FFN model is trained through supervised learning, and the training data set comes from historical projects. Each data sample contains the above input features of a completed dependent task pair (upstream task, downstream task) and a label, which is the difference between the actual execution time and the planned execution time of the downstream task. The loss function is set as mean absolute error (MAE), and the RMSprop optimizer with a learning rate of 0.005 is used for iterative training. In the quantification process, the system processes each value in the set of absolute values of inter-node load state difference one by one. For example, for the absolute value of load state difference 21.5 between T01 and T03, the system collects the complexity of T01 (e.g. 7), the resource sensitivity (e.g. 0.8) and the planned slack time (e.g. 2 days) of T03, and the average response time (e.g. 0.5 days) of the expert E05 supporting T03. These data (21.5, 7, 0.8, 2, 0.5) are combined into a feature vector and input into the trained FFN model. The model inference outputs a predicted value, for example 1.2, which indicates that the workload fluctuation of the upstream task T01 is expected to cause the execution time of the downstream task T03 to increase by 1.2 days. This 1.2 is the inter-node transmission impact value from T01 to T03. The system repeats this process for all absolute values of load state difference in the set, and finally calculates all the values to form the dependency impact strength set.

[0058] The steps for obtaining the project critical path weight graph are:

[0059] Based on the dependency influence strength set, the inter-node transmission influence value in each item of the dependency influence strength set is matched with the dependency connection relationship between the corresponding task nodes in the task knowledge coupling network, the inter-node transmission influence value is assigned to the weight of the dependency connection between the corresponding task nodes, the dependency connection weight of the task knowledge coupling network is updated item by item, and the project critical path weight graph is formed.

[0060] Specifically, based on the dependency influence strength set, the system first collects all the inter-node transmission influence values and calculates the maximum and minimum values thereof. For example, after analyzing all the dependency relationships, the minimum influence value obtained is 0.1 day and the maximum influence value is 3.5 days. Then, the system performs minimum-maximum normalization processing on each inter-node transmission influence value, and maps it to the interval [0, 1]. The normalization formula is wherein is the original inter-node transmission influence value, and are the minimum value and the maximum value in the set, respectively. For example, the normalized weight of the inter-node transmission influence value 1.2 from T01 to T03 calculated in the previous step is This normalized value will be used as the weight of the corresponding dependency edge. Subsequently, the system iterates through each record in the dependency influence strength set. Each record contains the source task node, the target task node, and the calculated inter-node transmission influence value. The system locates the exact connection edge representing the direct dependency relationship in the task knowledge coupling network according to the identification of the source task and the target task. Then, the system updates the original weight (if any) of the connection edge to the newly calculated and normalized weight value. If the edge has no weight before, it is directly assigned the new weight. For example, the system finds the directed edge from node T01 to node T03 in the task knowledge coupling network and sets its weight attribute to 0.3235. The system performs this update in an atomic operation to ensure data consistency and prevent read-write conflicts during the update process. This process covers all dependency connections between tasks in the complete execution path in the network. For those dependency connections that do not exist in any complete execution path from the start node to the end node, their weights can remain at the default value (e.g., 0) or be assigned a lower initial value according to other rules (such as historical average influence). After matching and assigning all the calculated inter-node transmission influence values, the topology of the original task knowledge coupling network remains unchanged, but the dependency connection edges between tasks are assigned with quantified weights that reflect the disturbance propagation capability, thereby forming the project critical path weight graph.

[0061] The steps for obtaining the full-node disturbance propagation effect set are as follows:

[0062] Based on the project critical path weight graph and the set time window, the original execution time of all task nodes in the project critical path weight graph is extracted in turn, and a fixed delay disturbance is applied to each task node. The task number, disturbance injection time and corresponding disturbance duration of the disturbed node are recorded to generate a task node delay disturbance list;

[0063] According to the task node delay disturbance list, the propagation path of each disturbance to the downstream task node is simulated, and all affected task nodes are traversed in topological order. In the simulation, it is judged whether all previous task nodes are completed and meet the resource ready condition every time a task node is encountered. If any previous task is not completed or the resource is not released, the earliest start time of the current task is adjusted to the maximum value of the latest completion time of the previous task and the earliest available time of the current resource. The difference between the latest completion time of the previous task and the earliest available time of the current resource and the original planned start time is the cumulative delay of the task. The disturbance propagation delay of all task nodes is accumulated to generate a full node disturbance propagation effect set.

[0064] Specifically, based on the project critical path weight graph and the set time window, the system first determines the time window for disturbance simulation. The length of the time window is not fixed but dynamically set according to the current stage of the project. For example, in the early stage of the project, the time window can be set to 30 working days consistent with the first milestone period, while in the middle of the project execution, it is shortened to 10 working days matching the single iteration period in agile development. Here, the time window is set to 10 working days. Then, the system automatically traverses all task nodes in the project critical path weight graph, accurately extracts the "original execution time" field from the attribute data of each node, and records the time data in hours. For example, the system extracts the original execution time of task T015 as 40 hours and the original execution time of task T016 as 24 hours. After extracting the execution time of all nodes, the system independently applies a "fixed delay disturbance" to each task node. The "fixed delay" is a standardized disturbance unit, and its value is set to 8 hours, which is equivalent to a standard working day. The value is set based on statistical analysis of historical project data. The analysis results show that single effective delay caused by non-technical problems (such as meetings, temporary redeployment, and requirement clarification delay) usually concentrates between 4 to 12 hours, and taking the average value of 8 hours as the standard disturbance amount is representative. The disturbance application process is sequential, that is, the system first applies disturbance to T001 and completes the subsequent propagation simulation, then applies disturbance to T002, and so on, until all task nodes are disturbed once. When applying each disturbance, the system generates a detailed disturbance record containing three core information: the unique task number of the disturbed node (e.g., "T015"), the precise time when the disturbance is injected (usually set to the planned start time of the task, e.g., 9 am on the 5th day of the project), and the injected disturbance duration (i.e., 8 hours). All these independent disturbance records are collected and summarized one by one, finally forming a structured task node delay disturbance list.

[0065] According to the task node delay disturbance list, the system starts a discrete event simulation engine, which performs an independent simulation deduction for each record in the list, for example, when processing the record about the delay of task T015 for 8 hours in the list, the simulation engine resets the project state to the initial plan, only delays the planned completion time of T015 by 8 hours, then the simulation starts to advance forward according to the topological order of the project critical path weight graph, traversing all downstream task nodes directly or indirectly affected by T015, when the simulator evaluates T015's direct downstream task T023, it first performs precondition checks, the first is task dependency check, the simulator checks the simulation completion time of all predecessor task nodes of T023, and takes the latest time as the task readiness time of T023, since the completion time of T015 is delayed by 8 hours, this latest time will also be delayed accordingly, the second is resource readiness check, the simulator queries a global resource scheduling table, which records the time allocation of all expert resources (such as senior engineer E07, database administrator E09) in the project in real time, if T023 needs E07, and E07 is still occupied by another unrelated task T099 for 4 hours after T015 is completed, then the earliest available time of the resource will be the time when E07 completes T099, the simulator sets the actual earliest start time of T023 as the maximum value of the task readiness time and the earliest available time of the resource, for example, if T015 is completed at 17:00 on the 6th day after the delay, and resource E07 is available at 11:00 on the 7th day, then the earliest start time of T023 is adjusted to 11:00 on the 7th day, then the system calculates the cumulative delay of T023, which is the time difference between its new earliest start time (11:00 on the 7th day) and the original planned start time (for example, 9:00 on the 6th day), which is 26 hours, this calculation process is recursively applied to all affected downstream nodes until the end of the path, after completing a complete simulation of the disturbance of T015, the system records the cumulative delay time of each task node in the network (including unaffected nodes, whose delay is 0) caused by the disturbance, stores these (task number, cumulative delay) data pairs as a set, after traversing all items in the task node delay disturbance list, the system aggregates the result sets generated by all simulation runs to generate the full node disturbance propagation effect set.

[0066] The steps for obtaining the project structure disturbance sensitivity index are:

[0067] Based on the full node disturbance propagation effect set, the project structure disturbance sensitivity index is calculated, and the calculation formula is:

[0068] ;

[0069] Wherein, is the project structure disturbance sensitivity index, The total number of task nodes where the perturbation propagation effect is concentrated. For task nodes The cumulative delay time generated in the disturbance simulation Represents task node Connect to all successor task nodes The sum of edge weights, For task nodes The set of all directly downstream task nodes, This is the sensitivity adjustment coefficient. For task nodes The amount of concurrent resources required for execution, For task nodes The total number of concurrent resources that the system can schedule during the execution period. To prevent extremely small positive constants with a denominator of zero.

[0070] Specifically, the formula: The advantage of the formula is that... This directly quantifies the "dynamic" ripple effect caused by a single-point disturbance in the entire project network, that is, the total scale of delayed propagation. Secondly, The term characterizes the "static" importance of task nodes in the project structure, that is, the breadth and intensity of their direct influence. Finally, the exponential term... The introduction of a non-linear penalty based on "resource constraints" identifies tasks that rely on scarce resources and dramatically amplifies their sensitivity through an exponential function. Since resource bottlenecks are often the key factors that cause project delays to increase exponentially, this multi-dimensional design enables the sensitivity index to not only identify nodes on the critical path but also to discover hidden critical nodes that, although not on the main path, are highly susceptible to systemic risks due to resource competition.

[0071] The total number of task nodes in the overall perturbation propagation effect set is obtained directly from the project critical path weight graph generated in the previous step. It is calculated by counting the total number of nodes in the graph and represents the overall project scale and complexity. For example, for a medium-sized software development project, if its tasks are decomposed into 200 independent task nodes, then... .

[0072] For task nodes The cumulative delay time generated in the perturbation simulation is extracted from the set of full-node perturbation propagation effects generated in the previous step. Specifically, it refers to the time delay generated when perturbation propagation effects on task nodes occur. The sum of delay times generated by all other task nodes in the network after applying a unit delay perturbation; it reflects the node's... the disturbance propagation ability of the task node, for example, when an 8-hour delay is imposed on task node T034, the cumulative delay sum of all its downstream tasks (T035, T041,...) is calculated to be 72.5 hours, then for node T034, hours.

[0073] denotes the task node connected to all the directly downstream task nodes , where, is the weight of the dependency edge from node to its direct successor node , which is directly read from the project critical path weight graph, and it quantifies the direct influence strength between tasks, denotes the set of all directly downstream task nodes of node , which is also determined from the adjacency relationship of the project critical path weight graph, for example, task node T034 points to two directly downstream tasks T035 and T041 in the graph, with their corresponding edge weights being and , then the calculation result of this term is .

[0074] is the sensitivity adjustment coefficient, which is used to adjust the contribution of resource tightness to the sensitivity index, and its value is determined based on regression analysis of historical project data within the organization, the analysis process is as follows: collect a database of at least 20 completed projects, extract the critical tasks that caused significant delays (delays exceeding 5% of the total project duration) in each project, for these critical tasks, calculate their resource occupation ratio , and record the actual project delay percentage they caused, use the resource occupation ratio as the independent variable and the delay percentage as the dependent variable to perform exponential regression fitting, and the optimal regression coefficient is , for example, through analysis of historical project data of a company, it is found that when takes the value of 2.5, the task sensitivity predicted by the model has the highest correlation with the actual project delay risk (the coefficient of determination reaches 0.82), therefore .

[0075] is the number of concurrent resources required for task node execution, this data comes from the resource allocation table in the project plan, which clearly defines the number of different types of resources required to execute each task, for example, the execution plan of task T034 clearly specifies that it needs "2 senior Java engineers" and "1 database architect", so the number of concurrent resources it needs is 3, that is .

[0076] For task node Total concurrent resources schedulable by the system within the execution period, this data is obtained from the enterprise resource management system, the system statistics in the task The total number of people with corresponding skills (such as "senior Java engineer" and "database architect") in the project team within the time window of the planned execution (for example, next Monday to Wednesday), which represents the size of the resource pool available for competing for the task, for example, during the planned execution of task T034, there are 8 senior Java engineers and 2 database architects in the entire project team, so the total number of available related resources is 10, that is .

[0077] To prevent the denominator from being a very small positive constant, set .

[0078] Calculation process:

[0079] To calculate the sensitivity score of a single task node T034 in the project For example, the score is the project structure disturbance sensitivity index Part of the sum term,

[0080] First, calculate the sensitivity score of T034, the formula is:

[0081] ;

[0082] According to the above parameter acquisition steps, the specific numerical value is:

[0083] ;

[0084] ;

[0085] ;

[0086] ;

[0087] ;

[0088] ;

[0089] Calculation process:

[0090] ;

[0091] ;

[0092] ;

[0093] ;

[0094] ;

[0095] ;

[0096] The system will repeat this calculation for all task nodes to get the sensitivity score of each node, and finally, the structural disturbance sensitivity index of the entire project is calculated:

[0097] ;

[0098] The results show that the sensitivity score of task node T034 is 176.505, which has no absolute meaning, but by comparing it with the sensitivity scores of all other task nodes in the project, its relative risk level can be judged. For example, after calculating the scores of all 200 nodes, if the average score is found to be 95 and the standard deviation is 30, then the score of 176.505 is much higher than the average and is at the extreme right end of the distribution, which clearly indicates that T034 is an abnormally sensitive task decomposition node. Any slight disturbance to this node is likely to be amplified and spread rapidly throughout the project through its strong propagation ability (high value), tight structural coupling (high value), and dependence on scarce resources (high resource occupancy ratio), causing serious schedule risk. The project structural disturbance sensitivity index (the average of all node scores) macroscopically reflects the robustness of the current project plan as a whole. The higher the index value, the more fragile the overall project structure, and the weaker the ability to resist risks.

[0099] The candidate execution sequence cost list acquisition step is:

[0100] Based on the project structural disturbance sensitivity index, the task decomposition nodes with a disturbance sensitivity index value greater than the median value of the overall task node disturbance sensitivity are screened one by one, the task numbers, execution dependency paths, and resource allocation records of the corresponding task decomposition nodes are extracted, the expert resource information bound to each task node is retrieved by calling the task knowledge coupling network, and a high-sensitivity task and expert resource pairing list is obtained;

[0101] Based on the high-sensitivity task and expert resource pairing list, the combination cost is calculated, and the calculation formula is:

[0102] ;

[0103] Among them, The combined cost of candidate execution sequences, This represents the total number of task nodes in the candidate execution sequence. For the first The original planned execution time for each task node. For the first The normalized sensitivity scores calculated for each task node in the previous perturbation propagation simulation. For the first The waiting time caused by scheduling delays for expert resources in each task node. This is a risk amplification factor (dimensionless). The waiting penalty factor (dimensionless). A very small positive constant set to prevent division by zero errors;

[0104] Based on the combined cost, the cost of each combination of task execution schemes is compared one by one. The task arrangement scheme with the lowest combined cost is selected, and the task order and expert arrangement are extracted to generate a list of candidate execution sequence costs.

[0105] Specifically, based on the project structure perturbation sensitivity index, the system first sorts the sensitivity scores of all task nodes and calculates the median of these scores. For example, for a project containing 200 task nodes, the system arranges the 200 sensitivity scores from low to high, takes the average of the 100th and 101st scores as the "overall task node perturbation sensitivity median", and the calculated median is 105.4. Then, the system starts the screening program and checks whether the sensitivity score of each task node is greater than 105.4. All task decomposition nodes that meet this condition are automatically marked as "high sensitivity tasks". For example, the score of task T034 is 176.505, which is greater than the median, so it is selected, while the score of task T088 is 92.1, which is less than the median, so it is ignored. After the screening is completed, the system generates a high sensitivity task list. For each task in the list, such as T034, the system performs a deep information extraction operation. First, it extracts the complete execution dependency path of the task from the project plan database, which includes all its predecessor and successor tasks, forming a subgraph structure. Second, it extracts detailed resource allocation records, clearly recording the specific human resources (such as "expert E07"), equipment resources (such as "high-performance computing server HPC02"), and budget information required to execute the task. Then, the system uses the established task knowledge coupling network to retrieve all expert resource nodes directly or indirectly connected to the high sensitivity task using the task node number (such as "T034") as the query key. This retrieval is not limited to experts explicitly assigned in the project plan, but also includes experts who have provided key knowledge support in the past, such as expert E11 who provided technical guidance to T034. The system integrates these task and expert associations to form a structured pairing record (for example, {task: T034, dependency path: [T011->T034->T056], allocated resources: [E07, HPC02], knowledge resources: [E11]}). All high sensitivity task pairing records are aggregated to ultimately obtain a high sensitivity task and expert resource pairing list.

[0106] Formula: The formula has the advantage of comprehensively evaluating the comprehensive cost of different scheduling schemes, not just the single time cost. The first term is the risk-adjusted time cost, which combines the original execution time of the task with its potential risk (reflected by the sensitivity score ). High sensitivity tasks are given higher cost weights, so they are prioritized or allocated higher quality resources in the optimization process. The second term is the waiting cost, which penalizes the waiting time of experts in the form of a quadratic function This means that even a short waiting time is significantly magnified, which is critical to preserve key expert resources, avoid talent loss and efficiency reduction due to improper scheduling. By integrating these two seemingly contradictory goals (fast completion of high-risk tasks and minimization of expert idle time) into a unified cost function, balanced by adjustable factors and , the formula can guide the scheduling algorithm to find an execution sequence that achieves the best balance between risk, time, and resource utilization.

[0107] is the total number of task nodes in the candidate execution sequence, the value of this parameter is equal to the number of tasks contained in the "high sensitivity task and expert resource pairing list" filtered in the previous step, which defines the scale of this optimization problem, for example, after screening, a total of 3 high sensitivity tasks are identified for optimization scheduling, then .

[0108] is the original planned execution time of the th task node, this data is directly extracted from the project plan database, corresponding to the estimated man-hours for each task in the Work Breakdown Structure (WBS), the unit is hour, this value is determined in the project planning stage based on factors such as the complexity of the task, historical experience, and available resources, for example, for the first task T034 in the high sensitivity task list, its original execution time in the project plan is 40 hours, then .

[0109] is the normalized sensitivity score of the th task node calculated in the previous disturbance propagation simulation, this value is derived from the calculation results of the project structure disturbance sensitivity index in the previous step, in order to make the sensitivity scores of different tasks comparable and limit their value range to [0, 1], it is necessary to perform min-max normalization on the original scores, the normalization formula is: where is the original sensitivity score of task , and are the minimum and maximum values of all original sensitivity scores of tasks, for example, the original sensitivity score of task T034 is 176.505, for example, the score range of all tasks is [50.2, 210.5], then its normalized sensitivity score is , i.e. .

[0110] is the normalized sensitivity score of the the waiting time of a task node caused by scheduling delay, which is dynamically calculated when simulating a specific candidate execution sequence, for a given sequence, the system simulates its execution process, when a specialist finishes a previous task, the time difference between the moment it is assigned to the next task (i.e. task ) and the moment it actually can start executing, is the waiting time of the specialist, this delay can be caused by task dependency, other resources not ready, etc., for example, in a candidate sequence, specialist E07 finishes task T011 at 3pm on Tuesday, but its next assigned task T034 cannot be executed until 11am on Wednesday because its predecessor task T029 cannot be finished until then, so E07 has to wait until 11am on Wednesday, its waiting time is 1 working day (8 hours) plus 2 hours from 3pm to 5pm, total 10 hours, then for task T034 in this sequence, .

[0111] is the risk amplification factor, which is a dimensionless weight coefficient, used to adjust the importance of task sensitivity in total cost calculation, the setting of this factor is based on the risk preference of the organization and the stage of the project, it is determined by the project management office (PMO) and key project stakeholders together through AHP (analytic hierarchy process), first, a judgment matrix is constructed to compare the relative importance of "minimizing project delay risk" and "minimizing resource waiting cost", for example, the specialist group unanimously believes that the former is 3 times more important than the latter, after consistency test, the corresponding weight can be obtained, in the critical stage of the project or for high-risk projects, the value of this factor is set higher to prioritize high-sensitivity tasks, for example, for a technology research project, the PMO sets to emphasize risk control.

[0112] is the waiting penalty factor, which is a dimensionless weight coefficient, used to adjust the importance of specialist waiting time in total cost calculation, its setting is similar to , also determined by AHP method, but its value is usually inversely proportional to , it reflects the sensitivity of the organization to the idle cost of key human resources, in an environment where human cost is high or key specialists are scarce, the value of this factor is set higher to maximize the utilization of specialists, for example, in the above example, if the weight of "minimizing project delay risk" is 0.75 and the weight of "minimizing resource waiting cost" is 0.25, and the basic dimension of the two items in the cost function is roughly the same, then can be set as the penalty term balanced with .

[0113] A very small positive constant set to prevent division by zero errors, same as in the above formula, set to . .

[0114] Calculation process:

[0115] As an example, calculate the combined cost of a specific candidate execution sequence (e.g., sequence A: T034 -> T081 -> T056) that contains three highly sensitive tasks (T034, T056, T081),

[0116] First, list the parameters for these three tasks:

[0117] Task T034 ( ): , , ;

[0118] Task T081 ( ): , , (analogous derived);

[0119] Task T056 ( ): , , (analogous derived, no waiting);

[0120] Plug in , , ;

[0121] The calculation process for the combined cost is as follows:

[0122] ;

[0123] Calculate the cost term for each task separately:

[0124] For task T034 ( ):

[0125] ;

[0126] ;

[0127] ;

[0128] ;

[0129] For task T081 ( ):

[0130] ;

[0131] ;

[0132] ;

[0133] ;

[0134] For task T056 ( ):

[0135] ;

[0136] ;

[0137] ;

[0138] The total combination cost is:

[0139] ;

[0140] The results show that the combined cost of execution sequence A (T034 -> T081 -> T056) is 222.036. This value is a cost metric that combines risk and efficiency. It does not represent any actual monetary value, but is a relative indicator used to compare the merits of different scheduling schemes. The system will repeat this calculation process for all possible execution sequences that satisfy the dependency constraints (e.g., sequence B: T081 -> T034 -> T056) to obtain their respective combined costs. By comparing the combined costs of all sequences, the sequence with the lowest cost can be selected as the optimal solution. A lower combined cost value means that the scheduling scheme has achieved a better balance between controlling the potential delay of high-risk tasks and reducing the waiting time of key experts.

[0141] Based on combinatorial cost, the system initiates an optimization algorithm to explore all possible combinations of task execution permutations. Since the number of highly sensitive tasks (M) is usually small (e.g., less than 10), the system uses a complete enumeration method to generate all possible topological sorts, i.e., all execution sequences that satisfy the inter-task dependency constraints. For each generated candidate execution sequence, the system calls the combinatorial cost calculation function defined in the previous step, inputs the sequence, and simulates its execution process to determine the expert waiting time corresponding to each task. And in combination with other known parameters ( , The total combination cost of the sequence is calculated, for example, for the set containing tasks T034, T056, T081, if T034 must be performed before T056, there are three possible sequences (T034, T056, T081), (T034, T081, T056), (T081, T034, T056), the system calculates the combination cost of each of the three sequences, for example, the cost list is (245.8, 222.0, 231.5), after calculating the combination cost of all candidate sequences, the system compares these cost values one by one, and determines and records the minimum value through a simple iterative search process, for example, in (245.8, 222.0, 231.5), the minimum value is 222.0, then the system reverses the cost list to find the task arrangement scheme corresponding to the minimum value, that is, (T034, T081, T056), and extracts the task order of the sequence and the optimal expert arrangement determined when simulating the sequence (for example, E07 performs T034, E11 performs T081, and E07 performs T056) in full. All this information, including each candidate sequence and its corresponding combination cost, is structured and recorded to generate a candidate execution sequence cost list.

[0142] The acquisition step of the knowledge resource optimization scheduling sequence is:

[0143] Based on the candidate execution sequence cost list, read the combination cost corresponding to each candidate execution sequence in the candidate execution sequence cost list one by one, compare the combination cost of each candidate execution sequence, extract the minimum value of the combination cost of all candidate execution sequences, and form the optimal combination cost value.

[0144] Based on the optimal combination cost value, traverse the candidate execution sequence cost list one by one, compare the absolute value of the difference between each candidate execution sequence combination cost value and the optimal combination cost value, and determine whether the absolute value of the difference is zero. If the absolute value of the difference is zero, mark the corresponding candidate execution sequence as the optimal execution sequence, extract the task order and expert resource arrangement information of the optimal execution sequence, and form the optimal execution sequence information.

[0145] Based on the optimal execution sequence information, read the task order and expert resource arrangement information in the optimal execution sequence information one by one, determine the execution order of each task node and expert resource according to the task order, and update the connection relationship between the corresponding task node and expert resource in the task knowledge coupling network, and form the knowledge resource optimization scheduling sequence.

[0146] Specifically, based on the candidate execution sequence cost list, the system first initializes a variable for storing the current minimum combination cost, which is initially set to a very large positive number, for example, 999999.9. Then, the system traverses each record in the candidate execution sequence cost list in a sequential reading manner, each record in the list containing a candidate execution sequence and its corresponding combination cost value, for example, the first record is "Sequence A: T034->T081->T056, Cost: 222.036", the system extracts the cost value 222.036 and compares it with the currently stored minimum cost value, since 222.036 is less than 999999.9, the system updates the minimum cost value to 222.036. Then, the system reads the second record, for example, "Sequence B: T081->T034->T056, Cost: 231.500", extracts the cost value 231.500 and compares it with the current minimum cost 222.036, since 231.500 is not less than 222.036, the minimum cost value remains unchanged. This process continues until all candidate execution sequences in the list are processed. During the entire traversal process, the system only retains the lowest cost value encountered, and when the traversal ends, the final value stored in the variable is the global minimum value among all candidate execution sequence combination costs. This finally determined minimum value is officially recorded by the system to form the optimal combination cost value.

[0147] Based on the optimal combination cost value, e.g. 222.036, the system initiates another traversal procedure to access each record in the candidate execution sequence cost list one by one. For each record, e.g. "Sequence A: T034->T081->T056, Cost: 222.036", the system extracts its combination cost value 222.036 and calculates the absolute value of the difference between this value and the determined optimal combination cost value 222.036, which is |222.036 - 222.036| = 0. Next, the system performs a judgment operation to check whether this absolute value of the difference is exactly equal to zero. In this case, the judgment uses a floating-point safe comparison method, i.e. whether the absolute value of the difference is less than a very small tolerance threshold, rather than directly using the "==" operator, in order to avoid false judgments caused by floating-point precision problems. In this example, the absolute value of the difference is 0, which meets the condition. Therefore, the system marks the currently processed "Sequence A" as the "optimal execution sequence" and immediately extracts the complete task ordering information (i.e. the order of T034, T081, and T056) and the detailed expert resource arrangement determined when simulating this sequence (e.g. task T034 is assigned to expert E07, task T081 is assigned to expert E11, and task T056 is assigned to expert E07) from this record. If there are multiple optimal sequences with the same combination cost, the system will mark and extract all of them to provide multiple alternative solutions for the project manager. After traversing the entire list, the task ordering and expert resource arrangement information contained in all records marked as "optimal execution sequence" are integrated into a separate data structure to form the optimal execution sequence information.

[0148] Based on the optimal execution sequence information, the system first parses the information structure, reads the unique or multiple optimal execution sequences contained therein, if there are multiple optimal sequences, the first one in the list is selected as the final execution scheme by default, or an interface is provided for the project manager to manually select, here, taking the selected unique optimal execution sequence (task order: T034, T081, T056; expert arrangement: T034-E07, T081-E11, T056-E07) as an example, the system determines the execution order of the tasks one by one according to the extracted task order, that is, T034 is the first execution task, T081 is the second, and T056 is the third, then the system updates the connection relationship in the task knowledge coupling network according to the expert resource arrangement information, this updating process is not to modify the original dependency or knowledge coupling relationship, but to add a new edge type in the network, called "dispatch execution edge", which has time stamp and priority attributes, the specific operation is to create a new directed edge between task node T034 and expert resource node E07, the edge attribute is marked as {type: dispatch execution, order: 1, start time: predicted value, end time: predicted value}, similarly, similar "dispatch execution edges" are established between T081 and E11, and T056 and E07, and the corresponding order (2 and 3) is assigned, these newly added edges clearly define the specific execution process and resource binding relationship of high sensitivity tasks after optimization, they are superimposed on the original task knowledge coupling network, so that the network can not only show the logical dependency and knowledge association between tasks, but also intuitively present the actual operation sequence after optimization, which aims to reduce risk and improve efficiency, finally, the enhanced network graph containing the original network structure and the newly added dispatch execution edges is defined and output as a whole, forming the knowledge resource optimization scheduling sequence.

[0149] The above is only a preferred embodiment of the present application, and does not limit the form of the present application in other ways. Any skilled person in the art can use the disclosed technical content to make changes or modifications to equivalent embodiments applied to other fields, but any simple modification, equivalent change and modification made according to the technical essence of the present application to the above embodiments without departing from the technical solution content of the present application still belongs to the protection scope of the present application.

Claims

1. A deep learning-based intelligent task decomposition and optimization method, characterized in that, Includes the following steps: Based on the project task breakdown list and collaboration records, the dependencies between tasks are identified and node connections are established. Then, based on the human resources collaboration communication records, the flow of problem requests and solutions is extracted to establish a task knowledge coupling network. Based on the task knowledge coupling network, the dependency connections of each execution path in the network are traversed to generate a dependency influence strength set. Based on the dependency influence strength set, the transmission influence value is assigned as a weight to the corresponding dependency edge of the task knowledge coupling network to construct the project critical path weight graph. Based on the project critical path weight graph and the set time window, a unit delay perturbation is applied to each task node in sequence to obtain the full node perturbation propagation effect set. Based on the full node perturbation propagation effect set, the statistical expectation is calculated to obtain the project structure perturbation sensitivity index. Based on the project structure disturbance sensitivity index, abnormal sensitivity task decomposition nodes are identified. The task knowledge coupling network is invoked to determine the required expert resources. The combined cost of task execution time, expert waiting time and path sensitivity under different scheduling schemes is calculated to generate a candidate execution sequence cost list. Based on the candidate execution sequence cost list, the sequence with the lowest combined cost is selected as the knowledge resource optimization scheduling sequence.

2. The intelligent task decomposition and optimization method based on deep learning according to claim 1, characterized in that, The steps for obtaining the task knowledge coupling network are as follows: Based on the project task breakdown list and collaboration records, the task identifiers of all tasks in the project task breakdown list are extracted one by one. According to the explicit dependency relationships between the task identifiers in the collaboration records, the direct dependency order between each task identifier and other task identifiers is matched and determined one by one to obtain the task node dependency relationships. Based on the task node dependency relationship, the problem request and problem resolution records in the human resources collaboration communication records are called. The identity information of the problem initiator and the problem solver is parsed from each communication record. According to the communication time order in the communication record, the direction of the problem from the initiator to the solver is identified. The problem flow in the communication record is determined one by one to obtain the problem request and resolution flow. Based on the problem-solving flow, the expert identity information of the expert resources who are problem solvers is extracted from the communication records. The task nodes associated with the problem-solving flow are linked to the corresponding expert identity information for knowledge dependency association, and the connection relationship between task nodes and expert resource nodes is gradually established to form a task knowledge coupling network.

3. The intelligent task decomposition and optimization method based on deep learning according to claim 1, characterized in that, The steps for obtaining the dependency influence intensity set are as follows: Based on the task knowledge coupling network, a deep learning process is adopted to call each complete execution path from the starting task node to the ending task node in the task knowledge coupling network one by one, traverse all task nodes on the execution path, extract the workload status data of each task node one by one, calculate the absolute value of the difference between the workload status data of each task node in the execution path and the workload status data of the preceding task node, and form a set of absolute values ​​of load status difference between nodes. Based on the set of absolute values ​​of load state differences between nodes, according to the direction of the dependent connections of each task node in the task knowledge coupling network, the influence of each absolute value of load state difference on the execution time of directly adjacent downstream task nodes is analyzed item by item. The numerical relationship between each absolute value of load state difference and the change in the execution time of the corresponding downstream task node is quantified, and the inter-node transmission influence value is calculated to form a set of dependency influence intensity.

4. The intelligent task decomposition and optimization method based on deep learning according to claim 1, characterized in that, The steps for obtaining the critical path weight map of the project are as follows: Based on the dependency influence strength set, the inter-node transmission influence value of each item in the dependency influence strength set is matched with the dependency connection relationship between the corresponding task nodes in the task knowledge coupling network. The inter-node transmission influence value is assigned a weight to the dependency connection between the corresponding task nodes. The dependency connection weight of the task knowledge coupling network is updated item by item to form the project critical path weight graph.

5. The intelligent task decomposition and optimization method based on deep learning according to claim 1, characterized in that, The steps for obtaining the set of propagation effects of the full node perturbation are as follows: Based on the project critical path weight graph and the set time window, the original execution time of all task nodes in the project critical path weight graph is extracted in sequence, and a fixed delay perturbation is applied to each task node. The task number, perturbation injection time and corresponding perturbation duration of the perturbation application node are recorded to generate a task node delay perturbation list. Based on the task node delay disturbance list, the propagation path of each disturbance to downstream task nodes is simulated. All affected task nodes are traversed in topological order. In the simulation, when a task node is encountered, it is determined whether all preceding task nodes have been completed and the resource ready conditions are met. If any preceding task has not been completed or the resource has not been released, the earliest start time of the current task is adjusted to the maximum value between the latest completion time of the preceding task and the earliest available time of the current resource. The difference between the latest completion time of the preceding task and the earliest available time of the current resource and the original planned start time is the cumulative delay of the task. The disturbance propagation delay of all task nodes is accumulated to generate a full node disturbance propagation effect set.

6. The intelligent task decomposition and optimization method based on deep learning according to claim 1, characterized in that, The steps for obtaining the project structure disturbance sensitivity index are as follows: Based on the set of all-node disturbance propagation effects, the project structure disturbance sensitivity index is calculated.

7. The intelligent task decomposition and optimization method based on deep learning according to claim 1, characterized in that, The steps for obtaining the candidate execution sequence cost list are as follows: Based on the project structure disturbance sensitivity index, task decomposition nodes with disturbance sensitivity index values ​​greater than the median of overall task node disturbance sensitivity are screened one by one. The task number, execution dependency path and resource allocation record of the corresponding task decomposition node are extracted. The task knowledge coupling network is called to retrieve the expert resource information bound to each task node to obtain a list of high-sensitivity tasks and expert resources. Based on the list of high-sensitivity tasks and expert resources, calculate the combined cost; Based on the combined cost, the cost of each task execution permutation and combination scheme is compared, the task permutation scheme with the lowest combined cost is selected, and the task order and expert arrangement are extracted to generate a candidate execution sequence cost list.

8. The intelligent task decomposition and optimization method based on deep learning according to claim 1, characterized in that, The steps for obtaining the knowledge resource optimization scheduling sequence are as follows: Based on the candidate execution sequence cost list, the combined cost corresponding to all candidate execution sequences in the candidate execution sequence cost list is read item by item. Taking the combined cost of candidate execution sequences as the comparison object, the combined cost values ​​of all candidate execution sequences are compared item by item. The minimum value among the combined cost values ​​of all candidate execution sequences is extracted to form the optimal combined cost value. Based on the optimal combination cost value, the candidate execution sequence cost list is traversed item by item, and the absolute value of the difference between the combination cost value of each candidate execution sequence and the optimal combination cost value is compared item by item. It is determined whether the absolute value of the difference is zero. If the absolute value of the difference is zero, the corresponding candidate execution sequence is marked as the optimal execution sequence. The task ordering and expert resource arrangement information of the optimal execution sequence are extracted to form the optimal execution sequence information. Based on the optimal execution sequence information, the task ordering and expert resource arrangement information in the optimal execution sequence information are read item by item. The execution order of each task node and expert resource is determined according to the task ordering order, and the connection relationship between the corresponding task node and expert resource in the task knowledge coupling network is updated to form a knowledge resource optimization scheduling sequence.

9. The intelligent task decomposition and optimization system based on the deep learning-based intelligent task decomposition and optimization method according to any one of claims 1-8, characterized in that, include: The task dependency identification module identifies the dependencies between tasks and establishes node connections based on the project task decomposition list and collaboration records. Then, based on human resources collaboration communication records, it extracts the flow of problem requests and solutions and establishes a task knowledge coupling network. The critical path modeling module, based on the task knowledge coupling network, traverses the dependency connections of each execution path in the network to generate a dependency influence strength set. Based on the dependency influence strength set, the transmitted influence value is assigned as a weight to the corresponding dependency edge of the task knowledge coupling network to construct the project critical path weight graph. The disturbance sensitivity analysis module, based on the project critical path weight graph and the set time window, sequentially applies a unit delay disturbance to each task node to obtain the full node disturbance propagation effect set. Based on the full node disturbance propagation effect set, it calculates the statistical expectation and obtains the project structure disturbance sensitivity index. The scheduling optimization module identifies abnormal sensitivity task decomposition nodes based on the project structure disturbance sensitivity index, calls the task knowledge coupling network to determine the required expert resources, calculates the combined cost of task execution time, expert waiting time and path sensitivity under different scheduling schemes, generates a candidate execution sequence cost list, and selects the sequence with the lowest combined cost as the knowledge resource optimization scheduling sequence based on the candidate execution sequence cost list.

Images