A construction progress collaborative prediction method based on federated learning

Abstract

The application relates to the field of industrial big data, and discloses a building construction progress collaborative prediction method based on federal learning, which comprises the following steps: calculating the influence range of each node delay on subsequent nodes, and establishing a project local progress feature library; using a process influence propagation algorithm to calculate delay propagation features generated by state changes of construction nodes, and generating local prediction parameters; obtaining each project local prediction parameter update result, using a difference self-adaptive aggregation algorithm to screen model parameters with cross-project migration ability for federal update, and forming a federal collaborative prediction model; determining the corresponding relationship between the current construction state and the historical construction state of a target project through a stage state alignment network, and outputting completion time prediction results of the current construction node and the subsequent construction node; and according to the prediction error improvement degree of different projects after updating, adjusting the proportion of shared parameters participating in federal aggregation, and updating the federal collaborative prediction model. The application has the advantages of improving prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of industrial big data, specifically to a collaborative prediction method for building construction progress based on federated learning. Background Technology

[0002] Existing construction progress prediction methods typically rely on data collected from individual projects, including construction plans, work completion status, personnel and equipment input, and material supply, to build predictive models. However, in large-scale construction groups with multiple collaborative projects, data from different projects is usually managed independently by each construction unit. Limited by data confidentiality, management permissions, and platform differences, effective data sharing is difficult, leading to insufficient training samples for predictive models. For example, when constructing multiple residential projects across regions, different projects are affected by construction environments, labor allocation, and material supply conditions, resulting in different progress patterns. A single-project model cannot promptly learn the reasons for delays in other projects. When a project experiences abnormal situations such as work stoppages due to extreme weather, delayed arrival of key materials, or partial rework, traditional prediction methods cannot utilize historical experience from other projects to adjust the model, easily causing deviations in construction milestone predictions.

[0003] Therefore, it is necessary to design a federated learning-based collaborative prediction method for building construction schedules to improve prediction accuracy. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a collaborative prediction method for building construction progress based on federated learning, which has the advantage of improving prediction accuracy and solves the problems mentioned in the background technology.

[0005] To achieve the aforementioned goal of improving prediction accuracy, this invention provides the following technical solution: a collaborative prediction method for building construction progress based on federated learning, comprising the following steps:

[0006] Collect planned nodes, actual completion times, process transition records, resource input records, and abnormal work stoppage records from multiple construction projects. Generate a node association table based on the sequential constraints between construction nodes, calculate the impact range of each node's delay on subsequent nodes, extract construction stages with the same constraints from different projects, and establish a local project progress feature library.

[0007] For each project's local schedule feature library, the process influence propagation algorithm is used to calculate the delay propagation characteristics caused by changes in the state of construction nodes. The state of construction nodes, process constraint features, and resource input features are input into the local schedule prediction network to learn the parameter relationship between changes in the state of construction nodes and the delay propagation amount and node completion cycle, and generate local prediction parameters.

[0008] Obtain the local prediction parameter update results for each project, generate parameter matching coefficients based on the differences in feature representation of construction status characteristics of different projects in the local progress prediction network, changes in model prediction error, and consistency of parameter update direction, use the difference adaptive aggregation algorithm to select model parameters with cross-project transfer capabilities for federated updates, retain project adaptive parameters, and form a federated collaborative prediction model.

[0009] By accessing real-time construction data of the target project, the current construction stage, process constraints, and resource input are input into the federated collaborative prediction model. The correspondence between the current construction status and the historical construction status of the target project is determined through the stage status alignment network. The prediction parameters corresponding to the matching construction status are called, and the prediction results of the completion time of the current construction node and subsequent construction nodes are output.

[0010] Collect deviation data between actual construction results and predicted results, input node completion time deviation, delay propagation deviation and resource response deviation into error diagnosis network, determine the parameter update direction of the corresponding prediction branch, perform incremental gradient update locally in the project, adjust the proportion of shared parameters participating in federated aggregation according to the degree of improvement of prediction error after the update of different projects, and update the federated collaborative prediction model.

[0011] Compared with existing technologies, this invention provides a collaborative prediction method for building construction progress based on federated learning, which has the following beneficial effects:

[0012] This invention establishes a local progress feature library that reflects the progress of construction stages by constructing a construction node constraint relationship analysis mechanism, which correlates construction process data such as plan execution, process conversion, resource input, and abnormal work stoppages in different construction projects. It analyzes the delay propagation process of construction node changes on subsequent processes through a process influence propagation algorithm, and dynamically corrects the node completion cycle based on resource input changes. This enables the model to identify progress changes caused by process dependencies, resource adjustments, and abnormal factors during construction, improving the quantitative analysis capability of delay impact. A federated learning mechanism enables experience collaboration among multiple projects without sharing original construction data. It adjusts the model parameter fusion process based on differences in construction stages, process constraints, and resource allocation, reducing interference from projects with significantly different construction conditions and improving the accuracy and stability of cross-project progress prediction. A stage offset correction algorithm eliminates data correspondence bias caused by inconsistent construction speeds across different projects, allowing historical project experience to match the current construction status of the target project, improving the model's adaptability to complex construction environments. Finally, it analyzes prediction errors through feedback from actual construction results, adjusting construction cycle prediction parameters, delay propagation parameters, and resource impact parameters for different error sources, achieving continuous optimization of the federated collaborative prediction model. In construction scenarios involving multiple projects in parallel management, frequent changes in construction phases, and dynamic adjustments to resource allocation, it is essential to improve the accuracy of construction progress prediction and reduce the risk of delays. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the method of the present invention;

[0014] Figure 2 This is a flowchart illustrating the steps involved in forming the federal collaborative prediction model of this invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Example 1: Please refer to Figure 1 An embodiment of the present invention provides a collaborative prediction method for building construction progress based on federated learning, comprising the following steps:

[0017] S1: Collect planned nodes, actual completion times, process transition records, resource input records, and abnormal work stoppage records from multiple construction projects. Generate a node association table based on the sequential constraints between construction nodes, calculate the impact range of each node's delay on subsequent nodes, extract construction stages with the same constraints in different projects, and establish a local project progress feature library.

[0018] The process of generating a node association table in S1 based on the sequential constraints between construction nodes is as follows:

[0019] Acquire planned execution data, on-site feedback data, and resource scheduling data from the construction management system. Extract the planned cycle, actual completion cycle, sequential process connections, and resource occupancy for each construction stage. Retrieve construction task arrangement data from the construction project management platform, schedule planning software, and on-site data acquisition terminals. Planned execution data includes the construction task name, planned start time, planned end time, preset duration, and task stage. On-site feedback data includes the actual start time, actual completion time, current completion percentage, and acceptance status. Resource scheduling data includes the number of construction personnel, duration of machinery and equipment deployment, material supply time, and resource adjustment records. Merge the above data according to the execution stage corresponding to the construction task. Match the planned cycle and actual execution cycle for the same construction stage, calculate the time offset of the actual completion cycle relative to the planned cycle, and extract the preconditions, completion conditions, and resource occupancy between construction tasks to form basic data records containing construction status information.

[0020] Based on the sequential execution conditions between construction tasks, analyze the impact of the completion status of the current construction stage on the start time of subsequent construction stages, and record the constraint relationships between different construction stages. Read the process connection conditions between construction tasks and determine the sequential execution relationships between each construction stage according to construction technology requirements. For example, the main structure construction can only proceed after the foundation pouring is completed, and the decoration construction can only begin after the main structure acceptance is completed. For construction stages with sequential dependencies, obtain the actual completion time of the previous construction stage and the actual start time of the subsequent construction stage, and calculate the waiting interval between them. When the start time of the subsequent construction stage is affected by the completion status of the previous construction stage, the previous construction stage is taken as the influencing node, and the subsequent construction stage is taken as the response node. Record the direction of influence, waiting time, and constraint strength. The constraint strength is determined based on the proportion of subsequent stages delayed due to the delay of the preceding stage in historical construction processes, used to represent the degree of influence between different construction stages.

[0021] By combining changes in personnel input, machinery operating status, material supply process, and the duration of abnormal downtime, the actual execution deviation during the construction process is calculated. Personnel attendance records, equipment operation logs, material arrival records, and downtime registration information are read during the construction process to obtain the difference between the actual and planned personnel input, the deviation between the actual and planned equipment operating time, the delay between the material arrival time and the planned supply time, and the duration of abnormal downtime. Resource impact weights are set according to the degree of influence of different resource factors on the construction cycle. Personnel input deviation, equipment operation deviation, material supply delay, and downtime are weighted and calculated to obtain the time deviation caused by resource changes. The actual completion time is compared with the planned completion time to calculate the overall execution deviation value of the construction process. The resource impact deviation value is correlated with the time execution deviation value to distinguish between delays caused by insufficient resources and delays caused by process adjustments.

[0022] The constraints, time variations, and resource influencing factors between construction stages are linked and stored to form a node association table for progress analysis. The process constraints, preceding and following influence directions, waiting times, planned cycles, actual completion cycles, execution offsets, and resource impact parameters corresponding to each construction stage are also linked. Using a single construction stage as the data association basis, the preconditions affecting the construction stage, related subsequent tasks, and resource changes are stored together. Different association records are divided according to construction stages, enabling construction stages with the same process constraints within the same stage to quickly access their corresponding historical status information. The final node association table includes at least construction stage status information, process constraints, time offset parameters, resource impact parameters, and subsequent influence relationships, used for subsequent delay propagation calculations and training of the construction progress prediction model.

[0023] The process of establishing a local project schedule feature library in S1 is as follows:

[0024] The system reads the construction constraints from the node association table to identify the construction stage with the time deviation and its associated subsequent stages. It then calls the established node association table to read the planned completion time, actual completion time, preceding and following process constraints, and resource impact parameters for each construction stage, calculating the time difference between the actual completion time and the planned completion time. When a construction stage is detected to have an extended completion period, the system obtains information on the affected subsequent stages based on the associated subsequent construction tasks, and extracts the changes in construction status before and after the delay. Construction constraints include prerequisite construction conditions, subsequent start conditions, and process waiting requirements. By analyzing the relationship between the completion status of prerequisite stages and the start time of subsequent stages, the system identifies the construction stage with the time deviation and its associated affected stages.

[0025] Based on the changes between the actual completion cycle and the planned completion cycle of the construction nodes, the corresponding delay duration, waiting time of subsequent nodes, resource compensation input and abnormal shutdown duration are read, and the time changes are correlated with the constraints in the node association table to determine the delay impact range and resource response situation corresponding to the status changes of different construction nodes.

[0026] The completion rate, process transition time, personnel input, equipment input and material input of each construction node are uniformly processed, and the processed construction node status, the corresponding delay impact range and resource response status are organized according to the construction stage.

[0027] The organized construction node status, process constraint characteristics, resource input characteristics, scope of delay impact, and changes in construction cycle are stored separately for each project to form a local project progress characteristic library.

[0028] S2: For each project's local schedule feature library, the process influence propagation algorithm is used to calculate the delay propagation characteristics caused by changes in the state of construction nodes. The construction node state, process constraint features, and resource input features are input into the local schedule prediction network to learn the parameter relationship between changes in the state of construction nodes and the delay propagation amount and node completion cycle, and generate local prediction parameters.

[0029] The process of calculating the delay propagation characteristics caused by changes in the state of construction nodes in S2 using the process influence propagation algorithm is as follows:

[0030] The system reads the actual completion time, process transition records, and subsequent node waiting records corresponding to construction nodes in the project's local progress feature library to identify construction nodes with status changes and directly affected subsequent construction nodes. A local progress feature library is established on the local server of each construction project, using the construction node number as an index for different construction records. Construction nodes, planned start times, planned completion times, actual start times, actual completion times, preceding construction nodes, subsequent construction nodes, and personnel, equipment, and material input records from the construction plan are written into this local progress feature library. The process transition records must include at least the actual completion time of the preceding construction node, the actual start time of the following construction node, and the interval between them. Subsequent node waiting records must include at least the waiting start time, waiting end time, and waiting reason. When processing the current construction progress, the planned completion time and actual completion time of each construction node are read one by one. Construction nodes whose actual completion time is later than their planned completion time are identified as construction nodes with status changes, and the corresponding delay duration is read. The construction plan identifies subsequent construction nodes based on the constraints between preceding and following processes. Nodes that must wait for the current node to complete before starting are considered directly affected. Nodes that started before the current node's completion and do not incur waiting times are not considered directly affected. The waiting records of directly affected nodes are then reviewed, comparing the start time of the wait with the actual completion time of the preceding node to determine if the wait is caused by a delay in the current node. This excludes waiting records caused by personnel scheduling, equipment failure, or material supply issues, ultimately identifying the construction node whose status has changed and its directly affected subsequent nodes.

[0031] The delay propagation path is determined based on the sequential constraints between construction nodes. Along this path, the delay duration of the current node, the waiting time of subsequent nodes, and the compensation input of personnel, equipment, and materials are read. The ratio of the portion of the delay duration not offset by the waiting time of subsequent nodes to the current delay duration is calculated to obtain the initial delay propagation ratio between nodes. Using the construction node where the state change occurs as the starting point for delay propagation, construction nodes are located sequentially based on the pre-constraints, post-constraints, and process transition relationships in the project construction plan. Construction nodes that can be continuously connected according to their construction sequence are arranged in order, forming a delay propagation path from the delay initiation node to the subsequently affected nodes. For example, in the main structure construction, a delay propagation path can be formed according to the sequential constraints of "concrete pouring—curing—formwork removal—masonry construction." Along this path, the delay duration of the current construction node, the waiting time of subsequent construction nodes due to waiting for the pre-constraint node, and the increased input of personnel, equipment, and materials after the delay occur are read sequentially. Determine how much of the delay at the current node can be offset by the waiting time at subsequent construction nodes. If a subsequent node has a waiting time corresponding to the current delay, the portion of the delay that can be absorbed by the waiting process is deducted from the current delay duration. The remaining delay after deduction is used as the amount of delay to be passed on to subsequent nodes. This remaining delay is then compared with the original delay duration of the current node, and the initial delay transmission ratio from the current node to subsequent nodes is determined based on their ratio. The increase in personnel, equipment usage time, and material supply at subsequent construction nodes relative to normal construction conditions is read and compared with the corresponding resource inputs in the historical normal construction records for that node to determine the resource compensation input implemented for the current delay.

[0032] The degree of delay offsetting is determined based on the resource compensation input corresponding to subsequent construction nodes, and the delay attenuation ratio is calculated by combining the process interval time between adjacent construction nodes. The initial delay transmission ratio and the corresponding delay attenuation ratio are calculated step by step to obtain the delay impact ratio transmitted by each construction node along the delay propagation path. For each pair of adjacent construction nodes in the delay propagation path, the increased personnel, equipment, and material inputs of the subsequent construction node after the delay occurs are read and compared with the resource input under the normal construction state of that node. When the subsequent construction node shortens the actual construction cycle after waiting by increasing the number of workers, increasing the equipment operating time, or replenishing construction materials in advance, the above-mentioned changes in resource input are identified as delay offsetting factors. Based on the changes in the actual completion cycle of the same construction node under different resource input conditions in historical construction records, the degree of delay that the current resource compensation input can absorb is determined. At the same time, the process interval time originally planned to be retained between adjacent construction nodes is read to determine whether the process interval can absorb part of the delay generated by the current node. When the process interval time is greater than the remaining delay, it means that part of the delay will not be transmitted to the next construction node; when the process interval time is insufficient to absorb the delay, the unabsorbed part will continue to be transmitted to the next construction node. Based on this, the delay attenuation ratio for the current node is determined jointly according to the delay ratio absorbed by resource compensation and the delay ratio absorbed by work intervals. The initial delay propagation ratio obtained in the second step is then adjusted using this attenuation ratio to obtain the actual delay impact ratio propagated from the current node to the next node. For propagation paths involving multiple construction nodes, the delay ratio of the starting node is not directly applied to all subsequent nodes. Instead, it is processed level by level according to the actual construction sequence. At each construction node, the resource compensation and work interval information are reread, and the delay impact propagated from the previous node is attenuated once, thus obtaining the delay impact ratio corresponding to each level of construction node.

[0033] The path delay impact is calculated based on the delay impact ratio corresponding to each construction node and the actual completion time offset of the node. This is then accumulated according to subsequent construction nodes along different delay propagation paths to obtain the delay propagation amount corresponding to the change in the construction node's state. The delay propagation amount, delay propagation path, and delay impact ratio are used as delay propagation features. After calculating the delay impact ratio for each construction node, the planned completion time and actual completion time of each construction node are read. The delay duration of the actual completion time relative to the planned completion time is determined as the actual completion time offset of that construction node. For each construction node in the delay propagation path, its actual completion time offset is correlated with the delay impact ratio corresponding to that node to determine the delay impact generated by that construction node along the current propagation path. If multiple delay propagation paths exist for the same construction node, the corresponding delay impact is calculated for each path, and the impact amounts belonging to the same subsequent node are accumulated according to the subsequent construction node number. For example, when a main structural construction node is simultaneously affected by delays from two preceding processes, the impact of each propagation path is calculated separately, and then the delay contributions of the two paths to the main structural node are merged. For cases where the same delay originates from a node and continues to propagate through different subsequent processes, the node order and corresponding impact ratio of each propagation path are retained to avoid overlap between different paths. Finally, the delay impact received by each subsequent construction node is used as the delay propagation quantity corresponding to the change in the construction node's state. Simultaneously, the starting node that generated this delay propagation quantity, the construction nodes it passed through, the propagation direction between each node, and the corresponding delay impact ratio at each level are saved, thus forming the delay propagation characteristic. Here, the delay propagation quantity describes the actual impact of the current construction node's state change on the completion time of subsequent construction nodes, the delay propagation path describes the specific transmission process of the delay impact between construction processes, and the delay impact ratio describes the degree to which the delay is retained after passing through different construction nodes.

[0034] The process of generating local prediction parameters in S2 is as follows:

[0035] The system reads the construction node status, process constraint characteristics, resource input characteristics, and corresponding delay propagation characteristics from the local progress feature library. Construction nodes with the same process constraint relationship are divided according to changes in completion cycle. Using the local progress feature library of a single construction project as the data source, the system reads the construction stage, node completion ratio, actual start time, actual completion time, planned completion time, predecessor nodes, and successor nodes corresponding to each construction node. Simultaneously, it reads the process constraint characteristics, personnel input quantity, equipment input quantity, material input quantity, and delay propagation amount, delay propagation path, and delay impact ratio obtained from the aforementioned process influence propagation algorithm for that construction node. The system determines the process constraint type of different construction nodes based on their predecessor and successor relationships. For example, nodes where "formwork installation can only proceed after rebar tying is completed" are classified as construction nodes with direct predecessor constraints, while nodes where "formwork can only be removed after concrete pouring and curing" are classified as construction nodes with waiting constraints. For construction nodes with the same process constraint relationship, the system reads their actual completion cycle and divides them according to construction records within different completion cycle intervals to distinguish between different construction states such as normal completion, slight delay, and longer cycle. The process of partitioning retains the resource input and delay propagation characteristics corresponding to construction nodes, so that construction nodes in the same partitioning result have similar process constraints, and can reflect the changes in completion cycle under different resource input and delay states, providing the same comparison conditions for subsequent determination of the main influencing construction states;

[0036] Based on the delay propagation amount corresponding to the changes in the construction node status and the number of subsequent construction nodes affected, the construction status that has the main impact on the completion cycle of the current node is determined, and the changes in personnel, equipment, and material input under the construction status are read. For each type of construction node after completion of the division, the corresponding delay propagation amount and the number of subsequent construction nodes affected by the delay of the construction node are read, and compared according to the construction stage and process constraint relationship. When the delay propagation amount of a certain construction node is large and the number of subsequent construction nodes affected along the delay propagation path is large, the construction status is determined to be the construction status with the main impact on the completion cycle of the current node; when the delay propagation amount is small and only affects one adjacent construction node, it is recorded as a general impact status. In order to avoid determining the main impact status based solely on the delay duration, the number of personnel input, equipment usage time, and material input before and after the delay of the construction node are further read, the changes in resource input before and after the delay are compared, and the construction stage in which the changes in resource input occur and the corresponding node completion cycle are recorded. For cases where there are multiple construction records under the same construction status, compare the records under the same process constraints, similar completion ratios, and similar resource input conditions to determine whether the increase in personnel, equipment allocation, or material replenishment is accompanied by changes in the node completion cycle, thereby forming a corresponding record between construction status, delay amount, number of affected nodes, and changes in resource input;

[0037] The changes in resource input are compared with the actual completion cycles of the corresponding construction nodes to determine the cycle correction amount generated by different resource input states on the node completion cycle. The time impact of the current construction node on subsequent nodes is determined based on the delay propagation amount. For each major influencing construction state, the personnel, equipment, and material input amounts corresponding to different construction records under that state are read, and the actual completion cycles before and after the changes in resource input are compared. For example, for the rebar tying node, the actual completion times after increasing the number of rebar workers, increasing the operating time of rebar processing equipment, and supplying rebar materials in advance are read and compared with the completion time without resource compensation to determine the degree of cycle shortening corresponding to different resource input changes. For equipment resources, the actual equipment input time and parallel operation of equipment are used to determine whether the increase in equipment truly changes the construction cycle. For material resources, the actual arrival time of materials is used to determine whether material replenishment eliminates construction waiting. The stable changes in the actual completion cycle under different resource input states are recorded as the cycle correction amount for the corresponding resource input state. The system reads the amount of delay propagated from the current construction node to subsequent construction nodes through the delay propagation path, and determines the time impact of the current node's delay on subsequent nodes based on the actual start and finish times of the affected nodes. When subsequent nodes absorb part of the delay through resource compensation or existing work intervals, the absorbed portion is deducted from the time impact of continued propagation, retaining only the portion that actually affects the construction cycle of subsequent nodes. This simultaneously yields the cycle correction amount corresponding to resource input and the time impact of the delay propagated to subsequent construction nodes.

[0038] The completion cycle of construction nodes, the amount of delay propagation, and the cycle correction corresponding to resource input are input into the local schedule prediction network for training to obtain the local prediction parameters for the corresponding project. The completion cycle of construction nodes obtained from the above processing is used as the construction result data, the status of construction nodes, the process constraint characteristics, and the resource input characteristics are used as the construction status data, and the amount of delay propagation, the time impact of delay propagation paths, and the cycle correction corresponding to resource input are used as the data affecting the change in completion cycle and input into the local schedule prediction network. The network is trained separately on the local server of each project, and the original construction node data is not transmitted to other projects. During training, the actual completion cycle of construction nodes is used as the prediction target, and the network is trained using construction records with the same process constraint relationship, so that the network can identify the correspondence between different construction node statuses and completion cycles. The amount of delay propagation and its corresponding time impact of subsequent nodes are added so that the network can learn the impact of the current construction node delay on the completion cycle of subsequent construction nodes. The cycle correction corresponding to personnel, equipment, and material resource input is then added so that the network can distinguish the change in completion cycle caused by changes in resource input and the change in completion cycle caused by delay propagation. After each round of local training, the predicted node completion time is recalculated using newly added construction records for the project. The predicted result is then compared with the actual completion time, and the parameters in the network corresponding to construction status, delay propagation, and resource input are adjusted based on the deviation. After several consecutive rounds of training, when the predicted node completion time deviation decreases to the training stop condition set by the project, or when a preset number of training rounds is reached, local training stops, and the parameters in the current network used to determine the node completion cycle, the impact of delay propagation, and resource cycle correction are saved as the local prediction parameters for this construction project.

[0039] S3: Obtain the local prediction parameter update results for each project. Based on the differences in feature representation of construction status characteristics of different projects in the local progress prediction network, changes in model prediction error, and consistency of parameter update direction, generate parameter matching coefficients. Use the difference adaptive aggregation algorithm to select model parameters with cross-project transfer capabilities for federated updates, retain project adaptive parameters, and form a federated collaborative prediction model.

[0040] The process of generating parameter matching coefficients in S3 based on the differences in feature representations, model prediction error changes, and consistency of parameter update directions in the local progress prediction network for different project construction status characteristics is as follows:

[0041] The system acquires the local prediction parameter update results of the local progress prediction network for each project under the same construction stage, and reads the construction node status, process constraint characteristics, and resource input characteristics corresponding to the parameter update results. Local progress prediction networks are trained on the local servers of each construction project, and the changes in local prediction parameters before and after each round of local training are recorded. To ensure the comparability of parameters between different projects, the local prediction network uses the same network level and parameter number to process construction status, process constraints, and resource input, ensuring that parameters playing the same prediction role in different projects have the same parameter number. The system reads the node completion percentage, construction node status, completion status of preceding nodes, number of subsequent nodes, number of personnel, equipment input duration, and material input for each project according to the construction stage, and associates and saves the above construction status data with the corresponding local prediction parameter update results for the current round. For example, the construction record of "concrete pouring completion percentage between 60% and 80%" in the main structure construction stage is associated with the corresponding local prediction parameter update results for that stage, while also recording the corresponding process constraints and resource input status. For projects with different construction phase names but consistent process constraints, phase correspondence is established based on the preceding and subsequent processes and the completion ratio of nodes, enabling local prediction parameters for different projects to be compared under the same construction conditions;

[0042] Based on the differences in the completion rates of nodes, process constraints, and resource input rates at different construction stages for different projects, the state difference values ​​between project construction states are calculated. Using two construction projects undergoing parameter matching as comparison objects, the completion rates of corresponding construction nodes are read under the same construction stage conditions, and the difference between the completion rates of corresponding nodes in the two projects is calculated to reflect the differences in the progress of construction between the two projects. The number of preceding nodes, the number of subsequent nodes, waiting constraints between nodes, resource occupancy constraints, and process transition conditions for the corresponding construction nodes of the two projects are read. For those with the same process constraints, they are matched, and for process constraints that exist only in one project, the differences are recorded to obtain the process constraint differences. For resource input characteristics, the number of personnel, equipment input time, and material input are read and converted according to the normal resource requirements of the corresponding construction nodes before comparison to obtain the differences in the resource input rates of personnel, equipment, and materials. The differences in node completion ratio, process constraint, and resource input ratio are converted to a unified comparison scale and then comprehensively processed according to the pre-set impact level. The differences in node completion ratio are mainly used to determine whether two projects are in similar construction progress states, the differences in process constraint are used to determine whether two projects have the same delay propagation conditions, and the differences in resource input ratio are used to determine whether two projects have similar construction capabilities. Thus, a state difference value is obtained to represent the degree of similarity in the construction states of two projects.

[0043] The system reads the node completion time prediction errors before and after local parameter updates for each project, calculates the error improvement corresponding to the local parameter updates, and calculates the parameter update consistency based on the adjustment direction and magnitude of the same parameters during the prediction parameter update process for different projects. After each project completes one round of local prediction parameter training, the node completion time prediction results before and after the local parameter updates are read and compared with the actual completion times corresponding to the same construction nodes. The time deviation between the prediction results and the actual results is calculated according to the construction stage. The prediction deviation before the parameter update is compared with the prediction deviation after the parameter update. When the prediction deviation decreases after the update, the amount of the decrease is determined as the error improvement generated by this round of local parameter updates. When the prediction deviation increases after the update, the change is recorded as a negative improvement, and its corresponding construction stage and parameter number are retained. For local prediction parameters with the same parameter number in two different projects, the direction and magnitude of parameter value changes before and after this round of training are read. When the corresponding parameters for both projects are adjusted in the increasing direction or both are adjusted in the decreasing direction, their parameter update directions are determined to be consistent. When one project increases and the other project decreases, their parameter update directions are determined to be inconsistent. For parameters with consistent directions, the adjustment magnitudes of the two items are compared. When the adjustment magnitudes are within a preset allowable range, the parameter update consistency is increased; when the adjustment magnitudes differ significantly, the parameter update consistency is decreased. For multiple consecutive rounds of local training, the adjustment directions of the same parameters in these rounds are further checked to ensure consistency. This avoids incorrect item matching judgments caused by accidental parameter changes in a single round of training. Ultimately, the error improvement and parameter update consistency for each item's corresponding parameters are obtained.

[0044] The state difference value, error improvement amount, and parameter update consistency are normalized. The parameter matching coefficient between projects is determined based on the impact of state differences on the applicability of parameter migration, the degree of error improvement, and the parameter update consistency. The aforementioned project construction state difference value, error improvement amount corresponding to parameter updates, and parameter update consistency are dimensionless to ensure the three types of data are within a unified comparison range. Specifically, a smaller state difference value indicates that the current construction states of the two projects are closer, and the less the parameter migration across projects is affected by differences in construction conditions. A larger error improvement amount indicates that the corresponding parameter has a more significant corrective effect on the current project's construction progress prediction. A higher parameter update consistency indicates that different projects have similar adjustment needs for the same parameter under similar construction states. During processing, a reverse transformation method is used for state difference values, so that the closer the construction states are, the higher the corresponding evaluation value. Error improvement amount and parameter update consistency are directly standardized according to their degree of improvement and consistency. A comprehensive judgment is made based on the pre-set weights for the impact of state differences, error improvement, and parameter update consistency to obtain the parameter matching coefficient corresponding to the same parameter number in each pair of projects. For parameters with small state differences, significantly reduced prediction errors after parameter updates, and consistent parameter adjustment directions, their parameter matching coefficients are increased, and these parameters are marked as having strong cross-project applicability. For parameters with large construction state differences, insignificant improvement in prediction errors after parameter updates, or opposite parameter adjustment directions across different projects, their parameter matching coefficients are decreased, and these parameters are marked as project-specific parameters. Finally, the project number, parameter number, parameter matching coefficient, and corresponding construction stage are saved together for use in subsequent differential adaptive aggregation to determine the participation level of local prediction parameters for different projects.

[0045] The process of forming a federated collaborative prediction model in S3 is as follows:

[0046] The process involves obtaining local prediction parameters and their corresponding parameter matching coefficients for each project, and determining the project applicability of the parameters based on the construction node status, process constraint characteristics, and resource input characteristics corresponding to the prediction parameters. After each construction project completes the training of its local progress prediction network, it only submits the updated results of the local prediction parameters and the parameter matching coefficients obtained in the aforementioned steps to the federated coordination end, without uploading the original data of construction nodes, actual completion time, and details of personnel, equipment, and material input. The federated coordination end reads the local prediction parameters submitted by each project according to a unified parameter number, and determines the construction node status, process constraint characteristics, and resource input characteristics affected by the parameter based on the data processing position of each parameter in the local progress prediction network. For example, for a parameter used to adjust the completion cycle of concrete pouring nodes, it reads the corresponding construction stage, preceding process constraints, number of subsequent nodes, and equipment input status, and uses this information as the project applicability of the parameter. The process also compares the construction status and process constraints corresponding to the same parameter number in each project to determine whether the parameter is only applicable to a specific construction stage, specific process constraint, or specific resource configuration condition, and saves the comparison results in association with the parameter matching coefficients, thus forming a record of the project applicability of each prediction parameter.

[0047] The update magnitudes of the same prediction parameter across different projects are compared, and the parameter matching coefficients between corresponding projects are used to determine the applicability of the prediction parameter across different projects. The federated coordination end matches the local prediction parameters submitted by different projects according to the parameter number, and reads the parameter values ​​before and after this round of training for each project to determine the actual update magnitude of the parameter in each project. The update direction of the same parameter is compared among projects. When multiple projects adjust in the same direction, it indicates that the parameter may reflect a common pattern of construction progress changes across different projects; when different projects adjust in opposite directions, it indicates that the parameter may be affected by differences in project construction conditions. For parameters with consistent update directions, the update magnitudes between different projects are further compared. When the update magnitudes are within the allowable range, different projects are considered to have similar adjustment needs for the parameter; when the update magnitudes differ significantly, the applicability of the parameter across projects is reduced. Simultaneously, the parameter matching coefficients between the corresponding projects are read. When the project construction status, process constraints, and resource input status are similar, the applicability of the parameter migration is improved; when the construction status or resource allocation differs significantly between projects, even if the parameter update direction is consistent, its applicability is reduced. Through the above processing, the applicability of inter-project migration corresponding to each prediction parameter is obtained, and based on this, it is determined whether the parameter is suitable for participating in cross-project federated updates;

[0048] When the update direction of the same prediction parameter is consistent across multiple projects and the corresponding parameter matching coefficient meets a preset condition, the prediction parameter is identified as a shared parameter. The aggregation weight of the shared parameter is calculated based on the parameter matching coefficients of each project. For each prediction parameter, its local update results across multiple construction projects are read to determine whether the parameter generates effective updates simultaneously in two or more projects, and to check whether the parameter update directions for each project are consistent. When at least two projects have the same adjustment direction for the parameter, and the parameter matching coefficients for these projects all meet a preset matching condition, and the parameter, after local training, can reduce the prediction error of the corresponding project's node completion time, the parameter is identified as a shared parameter. After identifying the shared parameter, the weight of the project participating in the parameter aggregation is determined based on the parameter matching coefficients for each project, the degree of error improvement after parameter updates, and the degree of conformity between the project's construction status and the parameter's applicable range. For projects with high parameter matching coefficients, significant error improvement, and construction statuses that are close to the parameter's applicable range, their aggregation weight is increased; for projects with low parameter matching coefficients or where local parameter updates do not significantly improve the prediction error, their aggregation weight is decreased. To prevent a particular item from completely dominating the shared parameter update results due to a large number of local samples, the weights of each item are uniformly normalized after the aggregation weights are determined, so that the contribution ratio of each item to the same shared parameter is within a preset range.

[0049] When a prediction parameter has a significant error improvement effect only in a single project or when different projects have different update directions for the prediction parameter, the prediction parameter is identified as a project adaptive parameter and does not participate in cross-project parameter aggregation. For each prediction parameter, its error improvement, parameter update direction, and project applicability are checked separately across different projects. When a prediction parameter produces a significant improvement in prediction error only in a single project, but the prediction error does not decrease or even increases after updating the parameter in other projects, the parameter is identified as a project adaptive parameter. When different projects produce opposite update directions for the same parameter, the parameter is also identified as a project adaptive parameter. For example, if one project uses a tower crane for vertical material transportation, while another project uses a construction hoist, the impact of equipment resource input on the construction cycle may differ between the two projects. In this case, the corresponding resource correction parameter is not directly aggregated across projects. For parameters identified as project adaptive parameters, the federated coordinator does not send its local update value to the shared parameter aggregation process. Instead, it saves the corresponding parameter update value according to the project number, so that the parameter remains in the original construction project's local schedule prediction network. When the same parameter is suitable for sharing in some projects but manifests as a project adaptive parameter in others, only the project parameters that meet the sharing conditions are included in the shared parameter aggregation, while the remaining projects continue to use their respective project adaptive parameters, thereby avoiding interference from project-specific construction conditions on the prediction models of other projects;

[0050] The local update values ​​of the identified shared parameters are accumulated according to their corresponding aggregation weights to obtain the updated value of the shared parameter. This updated value, along with the project adaptive parameters retained by each project, is then written into the federated collaborative prediction model. The federated coordination end aggregates each identified shared parameter, reads the local update values ​​of each project participating in the update of that shared parameter, and calculates the contribution of each project's local update value according to its corresponding weight. The contributions of each project are then accumulated to obtain the federated update result of the shared parameter. After processing all shared parameters, the updated shared parameters are written into the corresponding parameter positions in the federated collaborative prediction model. Furthermore, the project adaptive parameters retained in the preceding steps are written into the model according to the project number, ensuring that the same federated collaborative prediction model simultaneously possesses both cross-project shared components and project-specific adaptive components. For shared parameters, multiple construction projects jointly update them to reflect consistent construction progress changes across different projects. For project adaptive parameters, only the local construction data of the corresponding project is updated to retain the project's unique construction organization methods, resource allocation conditions, and process constraints. After the model is written, the participating projects, aggregate weights and update times corresponding to the shared parameters are recorded on the federated coordination end. The construction stage and process constraints corresponding to the project adaptive parameters are recorded locally on each project. This enables the formed federated collaborative prediction model to simultaneously call the shared parameters and the project's corresponding adaptive parameters during the subsequent prediction of the target project.

[0051] S4: Access the real-time construction data of the target project, input the current construction stage, process constraints and resource input into the federated collaborative prediction model, determine the correspondence between the current construction status and the historical construction status of the target project through the stage state alignment network, call the prediction parameters corresponding to the matching construction status, and output the prediction results of the completion time of the current construction node and subsequent construction nodes.

[0052] In S4, the process of determining the correspondence between the current construction status and historical construction status of the target project through the stage state alignment network is as follows:

[0053] The system obtains the completion percentage, process transition status, and personnel, equipment, and material input status of the current construction stage of the target project. It retrieves construction stages with the same process constraints from the historical construction status data of the federated collaborative prediction model. At the data acquisition terminal on the target project construction site, it reads the status of the currently under-construction construction node and its predecessor and successor nodes, and obtains the completion percentage between the completed work and planned work at the current construction node. It also reads the transition status between the current node and its preceding and succeeding processes. The process transition status includes at least the end time of the previous node, the start time of the current node, the duration of construction at the current node, and the waiting status between the current node and the next node. Resource input status is obtained from the construction project's personnel attendance records, equipment operation records, and material requisition records, recording the actual number of personnel, equipment usage time, and material consumption at the current construction node. The stage status alignment network is built on the historical construction status records stored in the federated collaborative prediction model. It filters historical construction records according to the predecessor, successor, and waiting constraints of the current construction node of the target project, retaining only historical construction stages with the same process constraints. For example, if the target project is currently in the stage of "concrete pouring after formwork installation", then the historical construction status is used to select construction stages that also have the same constraints of the preceding and following processes, and historical records with different process connection methods are excluded, thereby avoiding incorrect matching based solely on the construction stage name;

[0054] Based on the difference between the current completion rate of the target project and the completion rates of historical construction stages, combined with the transition time difference between process transition states and the difference in resource input ratio, the stage offset of the current construction state of the target project relative to the historical construction state is calculated. For each selected historical construction stage, the node completion rate, preceding and following process states, process transition time, number of personnel, equipment runtime, and material input in the current construction state of the target project and the corresponding historical construction state are read. The current construction state of the target project is then mapped to each historical construction state according to the chronological position of the construction nodes. The completion rate difference, process transition time difference, and resource input ratio difference are used as input features of the stage state alignment network. The stage state alignment network extracts the current construction state and historical construction state of the target project respectively. The system learns the correlation between the progress of construction nodes, process transition status, and resource input changes based on the stage state characteristics of the project. It outputs the stage offset of historical construction stages relative to the current construction state of the target project based on the positional correspondence between the two stage state characteristics in the construction progress direction. The stage offset is determined by the stage state alignment network based on the stage state parameters obtained during training. It represents the construction position in the historical construction stage corresponding to the current construction state of the target project. When there are differences between the completion rate of the current construction node, process transition status, and resource input status of the target project and the historical construction state, the network comprehensively determines the degree of forward or backward movement of the historical construction position, rather than directly calculating based on any single difference. The stage offset and the corresponding historical construction node position are recorded together.

[0055] The corresponding positions of each construction node in the historical construction stage are adjusted according to the stage offset. The historical construction state after stage offset adjustment and the current construction state of the target project are input into the stage state alignment network. The corresponding historical construction state is determined according to the process constraints and resource input status of the two, and the parameters of the federated collaborative prediction model corresponding to the historical construction state are called. The stage offset output by the stage state alignment network for each historical construction stage is read. Using the construction node position of the corresponding historical construction stage as a reference, the completion ratio position, process transition position and resource input status of the historical construction node are adjusted along the actual progress direction of the construction node to obtain the adjusted historical construction state that is at the same stage position as the current construction state of the target project. The current construction state of the target project and the adjusted historical construction state are input into the stage state alignment network again. The stage state alignment network extracts the stage state features of the two after adjustment and determines them according to the preceding and following processes. The degree of correspondence between constraints, process transition status, and personnel, equipment, and material input status is determined. When the adjusted historical construction status maintains the original process constraints and the degree of correspondence meets the preset conditions, the historical construction status is determined as the historical construction status corresponding to the current construction status of the target project. When there are multiple historical construction statuses that meet the conditions, they are sorted according to the degree of correspondence output by the stage status alignment network, and the historical construction status with the highest degree of correspondence is selected. After determining the historical construction status, the shared prediction parameters in the federated collaborative prediction model are called according to the construction stage, process constraints, and resource input status corresponding to the historical construction status, and the project adaptive parameters corresponding to the target project are read. The shared prediction parameters are corrected under the current project resource configuration conditions. The called shared prediction parameters, project adaptive parameters, and stage status alignment results are written into the current prediction task for use in predicting the completion time of subsequent construction nodes.

[0056] The process of outputting the predicted completion time of the current construction node and subsequent construction nodes in S4 is as follows:

[0057] The system reads the process constraint features and resource input features corresponding to the current construction status of the target project. Based on the historical construction status determined by the stage status alignment network, it extracts the completion records of construction nodes with the same process constraint relationship from the local progress feature library. At the target project's construction data acquisition terminal, it reads the currently under construction node number, node completion percentage, current construction stage, preceding construction nodes, and subsequent construction nodes, and reads the process constraint features corresponding to the current node, as well as the personnel, equipment, and material input status. The process constraint features include information such as whether the preceding process must be completed, whether there is a waiting time between adjacent processes, and whether subsequent processes are allowed to be constructed simultaneously. The resource input features include the current actual number of personnel, the number of equipment used and the equipment running time, the actual amount of materials input, and the current material supply status. It then calls upon the historical construction status already determined by the stage status alignment network, searches the project's local progress feature library by construction node number, preceding process relationship, and subsequent process relationship, extracts the completion records of historical construction nodes with the same process constraint relationship as the current construction node of the target project, and further retains the actual completion cycle, node completion percentage, process transition time, and corresponding personnel, equipment, and material input status from the historical records. When multiple historical construction records exist under the same process constraint, priority is given to retaining records whose completion rate is close to the current completion rate of the target project. Further filtering is then performed according to the resource input status corresponding to the historical records, thereby obtaining historical construction node completion records that have a high degree of correspondence with the current construction status of the target project.

[0058] Compare the current resource input status of the target project with the resource input status corresponding to the historical construction status to determine the time correction amount caused by the difference in personnel, equipment, and material input on the completion cycle of construction nodes; compare the current personnel, equipment, and material input status of the target project with the historical construction status obtained after aligning with the stage status. For personnel input, read the current actual number of workers in the target project and the corresponding number of workers in the historical construction status, and combine the changes in the actual completion cycle of the same construction node under the two personnel configuration conditions to determine the impact of increasing or decreasing personnel on the construction cycle; for equipment input, compare the current number of equipment, the actual operating time of equipment, and the equipment input situation corresponding to the historical construction status, and determine the time impact of equipment input based on the actual completion cycle of the same construction node under different equipment input conditions; for material input, read the current actual material arrival quantity, material supply interval, and material input situation corresponding to the historical construction status to determine whether the difference in material supply causes construction waiting or shortens the material waiting time. The changes in construction cycles corresponding to the three types of resources—personnel, equipment, and materials—are assessed separately. When resource input increases and shortens the completion cycle, the corresponding cycle shortening is used as a positive time correction; conversely, when resource input decreases and lengthens the completion cycle, the corresponding cycle increase is used as a negative time correction. For cases where multiple changes occur simultaneously for the same resource, they are distinguished according to the chronological order of occurrence in the actual construction records to avoid repeatedly attributing the same construction cycle change to different resources. The final result is the time correction amount corresponding to the current resource input status of the target project relative to its historical construction status.

[0059] The time correction is input into the prediction parameters of the federated collaborative prediction model to calculate the estimated completion period of the current construction node. The time impact of subsequent construction nodes is calculated based on the delay propagation amount and path of the current construction node. The parameters of the federated collaborative prediction model corresponding to the historical construction states determined by the stage state alignment network are read into the current prediction task. The completion ratio, process constraint characteristics, delay propagation characteristics, and resource input characteristics of the current construction node of the target project are input into the corresponding prediction parameters for processing. The basic estimated completion period of the current construction node under the current construction state is determined based on the completion records of historical construction nodes with the same process constraint relationship. The basic estimated completion period is then adjusted using the resource input time correction obtained in the second step to obtain the estimated completion period of the current node considering the current resource input state. The delay propagation amount, delay propagation path, and delay impact ratio of each subsequent construction node formed by the current construction node during the aforementioned process impact propagation process are read. The time impact of the current node's delay on subsequent construction nodes is determined step by step according to the delay propagation path. For the delayed portion that has been mitigated through resource compensation or process intervals at subsequent nodes, it will not be passed on to the next stage; for the delayed impact that still remains, it will be recorded according to its corresponding subsequent construction node and propagation level, so as to obtain the time impact amount corresponding to each subsequent construction node.

[0060] The estimated completion period of the current construction node is adjusted against the time correction amount, and the estimated completion period of subsequent construction nodes is adjusted step by step according to the time impact of each subsequent construction node in the delay propagation path, resulting in the predicted completion time of the current construction node and subsequent construction nodes. The estimated completion period of the current construction node obtained in the third step is matched with the current construction time of the target project to determine the estimated completion time of the current construction node, and a final adjustment is made according to the resource input time correction amount determined in the second step. When there is an unresolved delay impact of the current node, the actual unresolved delay time is included in the estimated completion time of the current node; for the delay portion that has been resolved through resource compensation or process intervals, it is not included again. After completing the prediction of the current node, subsequent nodes are processed sequentially according to the actual order of construction nodes along the determined delay propagation path. The time impact of each subsequent construction node is added to the original estimated completion period of that node, and the personnel, equipment, and material input status of that subsequent node is read to determine whether there is resource compensation that can further offset the delay. If resource compensation exists, the time impact of subsequent propagation is reduced according to the periodic correction results corresponding to the same resource input changes in historical construction records. If resource compensation does not exist, the delay impact received by the current node is retained and propagated to the next construction node. In cases where multiple delay propagation paths point to the same subsequent construction node, the actual retained time impact of different paths is calculated separately, then merged to determine the estimated completion period of that node. Finally, the estimated completion time of the current construction node and each subsequent construction node is output according to the construction node number and planned sequence. Simultaneously, the resource correction amount, delay propagation amount, and delay propagation path corresponding to each prediction result are saved. This allows for node-by-node comparison between the predicted and actual completion times after the actual construction results are generated, facilitating error diagnosis and local prediction parameter updates.

[0061] S5: Collect deviation data between actual construction results and prediction results, input node completion time deviation, delay propagation deviation and resource response deviation into the error diagnosis network, determine the parameter update direction of the corresponding prediction branch, perform incremental gradient update locally in the project, adjust the proportion of shared parameters participating in federated aggregation according to the degree of improvement of prediction error after different projects, and update the federated collaborative prediction model.

[0062] The process of updating the federated collaborative prediction model in S5 is as follows:

[0063] The actual completion time, actual delay propagation range, and actual resource input changes of the target project after construction are obtained. These are compared with the predicted completion time, predicted delay propagation range, and predicted resource input changes output by the federated collaborative prediction model to obtain the node completion time deviation, delay propagation deviation, and resource response deviation. The estimated completion period of the current construction node is read, along with the actual construction start time and the current data collection time of the target project's current construction node. For the current construction node that has already started construction, the current data collection time is used as the prediction start time, and the estimated completion period is mapped to the construction timeline to obtain the estimated completion time of the current construction node. For nodes that have not yet started but have met the prerequisite construction conditions, the estimated start time of construction is used as the prediction start time, and the estimated completion period is mapped to the estimated start time of that construction node. The time correction amounts corresponding to personnel, equipment, and material inputs are read, and it is determined whether these correction amounts correspond to a shortening or extension of the construction period. The estimated completion time of the current construction node is adjusted accordingly. For example, if adding workers to the target project can shorten the construction cycle based on historical records of similar work processes, the estimated completion time will be moved forward accordingly. If insufficient equipment leads to an extended construction cycle, the extended time will be included in the estimated completion time. Through this process, the estimated completion time of the current construction node is obtained, taking into account the current resource allocation status of the target project.

[0064] The node completion time deviation, delay propagation deviation, and resource response deviation are input into the error diagnosis network. The prediction branch that generates the prediction error is determined based on the construction stage, process constraints, and resource input status corresponding to each deviation. After the actual completion of a construction node in the target project, the actual completion time and predicted completion time of that node are read to obtain the node completion time deviation. Simultaneously, the actual delay duration, the subsequent construction nodes affected by the actual delay, and their corresponding delay times are read and correlated with the delay propagation amount, delay propagation path, and expected completion times of subsequent nodes saved during the prediction phase to obtain the delay propagation deviation. Furthermore, the personnel input, equipment operation, and material input status recorded during the prediction process are read and correlated with the corresponding personnel, equipment, and material input status during the actual construction process, according to the construction node and construction stage, to obtain the resource response deviation. The node completion time deviation, delay propagation deviation, and resource response deviation are organized according to the corresponding construction stage, construction node, constraints between preceding and subsequent processes, and resource input status to form the input data of the error diagnosis network. Among them, the node completion time deviation is used to characterize the degree of deviation between the predicted completion cycle and the actual completion cycle; the delay propagation deviation is used to characterize the difference between the predicted delay propagation range and propagation time between construction nodes and the actual delay propagation; and the resource response deviation is used to characterize the difference between the impact of changes in resource input on the construction cycle during the prediction process and the actual resource input response.The error diagnosis network uses the prediction branch corresponding to the local schedule prediction network as the output category. The construction cycle prediction branch, delay propagation prediction branch, and resource correction prediction branch involved in the completion time prediction of the current prediction task are respectively used as branches to be diagnosed. Training data is constructed using prediction tasks with actual construction results already obtained from the project's historical construction records. During training, the node completion time deviation, delay propagation deviation, and resource response deviation corresponding to each prediction task are used as inputs, and the prediction branches corresponding to the main sources of actual deviations are used as training labels. The error diagnosis network learns the correspondence between different deviation combinations, construction stages, process constraints, resource input status, and prediction branches, enabling the network to identify the main prediction branches that generate the current prediction error based on the input deviations. After a new construction result is generated, the three types of deviations corresponding to the current node are input into the trained error diagnosis network, which then extracts different... The correlation characteristics between deviations are analyzed, and combined with the current construction stage, constraints between preceding and subsequent processes, and resource input status, the error diagnosis results corresponding to each prediction branch are output. Based on the error diagnosis results, the prediction branches that need parameter updates are determined. When the error diagnosis network determines that the node completion time prediction branch produces a major error, the construction cycle prediction parameters corresponding to that branch are read as local update objects. When the delay propagation prediction branch produces a major error, the prediction parameters corresponding to the delay propagation amount and delay impact ratio are read as local update objects. When the resource response prediction branch produces a major error, the resource correction parameters corresponding to personnel, equipment, and materials are read as local update objects. The determined prediction branches and corresponding error deviations are input into the local progress prediction network to perform incremental training, so that the error diagnosis network is responsible for determining the parameter update objects, and the local progress prediction network is responsible for updating the corresponding prediction parameters according to the actual construction results.

[0065] Based on the error change direction corresponding to the predicted branch, determine the local update direction and magnitude of the node completion cycle parameters, delay propagation parameters, and resource correction parameters, and perform incremental gradient updates in the project's local schedule prediction network. Using the current construction node as the starting point for delay propagation, read its directly subsequent construction nodes according to the established delay propagation path, and process them sequentially according to the priority constraints in the construction plan. For each subsequent construction node, read its basic estimated completion cycle and the time impact transmitted from the current node to that subsequent node, and map the determined actual retained delay time to the basic estimated completion cycle of that subsequent node to obtain the estimated completion cycle considering the delay impact of preceding nodes. Read the current personnel input, equipment input status, and material supply status of that subsequent construction node, and compare them with the resource input under the node's historical construction status. When subsequent nodes are detected to involve increased personnel, equipment reallocation, or earlier material supply in response to delays, the amount of time impact that resource compensation can absorb is determined based on the shortened construction cycle corresponding to the same resource changes in historical construction records. The already absorbed portion is then deducted from the received delay time impact at that node. If resource compensation is insufficient to fully absorb the received delay impact, the remaining time impact continues to be passed on from that node to the next construction node. If no effective resource compensation exists, all time impact received at that node is retained and passed on to the next node. By processing each node individually, the delay impact can be absorbed to varying degrees at different construction nodes based on the actual resource status, rather than being directly passed on to all subsequent nodes according to the delay time of the starting node.

[0066] The process involves obtaining the degree of prediction error improvement after each project completes its local update, jointly calculating the degree of improvement with the parameter matching coefficient of the corresponding project, and adjusting the proportion of shared parameters participating in federated aggregation after the update of each project. During the processing of each delay propagation path, a corresponding delay impact record is established for each subsequent construction node, recording the preceding node that caused the impact and its propagation path. When two or more preceding construction nodes affect the same subsequent construction node through different propagation paths, it is first determined whether the delay impact in each path has been absorbed by the corresponding resource compensation, work intervals, or other construction conditions, and then the actual retained time impact in each path is merged. For example, when a masonry construction node is simultaneously affected by both the main structure delay and the material supply delay, the actual retained time impact of each path is calculated separately. The portion already absorbed by the early arrival of materials is no longer included, while the portion still retained due to the main structure delay continues to be included in the masonry node. After merging the path impacts, the merged time impact is adjusted with the estimated completion period of the construction node's foundation to determine the estimated completion period of the node after considering the preceding delays. If there are still delays that can be propagated to the next node, the remaining time impact after deducting the portion already digested by this node will be written into the delay input record of the next construction node and processed along the corresponding propagation path, thus forming a continuous delay impact propagation process.

[0067] The shared parameters of each project are aggregated according to the adjusted participation ratio, and the adaptive parameters of the projects are retained in the local prediction network of the corresponding projects. The updated shared parameters and the adaptive parameters of the projects form the next round of federated collaborative prediction model. After calculating the expected completion period of the current construction node and each subsequent construction node, the prediction results are sorted according to the node number and the relationship between the preceding and following procedures in the construction plan. The current construction node, the directly following construction node, and the subsequent construction nodes that continue to be affected by the delay are arranged in the actual construction sequence. For each construction node, the expected start time, expected completion time, expected completion period of the foundation, resource input time correction, actual included delay time impact, and the propagation path that caused the delay impact are saved. The parameters of the federated collaborative prediction model called by the prediction result and the historical construction state number obtained by stage state alignment are also recorded. Finally, the expected completion time of the current construction node and the subsequent construction nodes are output. The prediction result of the current construction node is used to determine whether the current construction task can be completed as planned, and the prediction result of the subsequent construction nodes is used to determine whether the delay will continue to affect the subsequent construction stages. After the forecast record is written into the project's local progress forecast record database, the actual completion time is read when the corresponding construction node is actually completed, and compared with the forecast completion time node by node to form the node completion time deviation; at the same time, the differences between the actual delay propagation range and the forecast delay propagation range, as well as the differences between the actual resource input and the forecast resource input are read.

[0068] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0069] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A collaborative prediction method for building construction progress based on federated learning, characterized in that, Includes the following steps: Collect planned nodes, actual completion times, process transition records, resource input records, and abnormal work stoppage records from multiple construction projects. Generate a node association table based on the sequential constraints between construction nodes, calculate the impact range of each node's delay on subsequent nodes, extract construction stages with the same constraints from different projects, and establish a local project progress feature library. For each project's local schedule feature library, the process influence propagation algorithm is used to calculate the delay propagation characteristics caused by changes in the state of construction nodes. The state of construction nodes, process constraint features, and resource input features are input into the local schedule prediction network to learn the parameter relationship between changes in the state of construction nodes and the delay propagation amount and node completion cycle, and generate local prediction parameters. Obtain the local prediction parameter update results for each project, generate parameter matching coefficients based on the differences in feature representation of construction status characteristics of different projects in the local progress prediction network, changes in model prediction error, and consistency of parameter update direction, use the difference adaptive aggregation algorithm to select model parameters with cross-project transfer capabilities for federated updates, retain project adaptive parameters, and form a federated collaborative prediction model. By accessing real-time construction data of the target project, the current construction stage, process constraints, and resource input are input into the federated collaborative prediction model. The correspondence between the current construction status and the historical construction status of the target project is determined through the stage status alignment network. The prediction parameters corresponding to the matching construction status are called, and the prediction results of the completion time of the current construction node and subsequent construction nodes are output. Collect deviation data between actual construction results and predicted results, input node completion time deviation, delay propagation deviation and resource response deviation into error diagnosis network, determine the parameter update direction of the corresponding prediction branch, perform incremental gradient update locally in the project, adjust the proportion of shared parameters participating in federated aggregation according to the degree of improvement of prediction error after the update of different projects, and update the federated collaborative prediction model.

2. The collaborative prediction method for building construction progress based on federated learning according to claim 1, characterized in that, The process of generating a node association table based on the sequential constraints between construction nodes is as follows: Acquire plan execution data, on-site feedback data, and resource scheduling data from the construction management system, and extract the planned cycle, actual completion cycle, connection relationship between preceding and subsequent processes, and resource usage for each construction stage; Based on the sequential execution conditions of construction tasks, analyze the impact of the current completion status of the construction stage on the start time of the subsequent construction stage, and record the constraint relationships between different construction stages. Calculate the actual execution deviation during the construction process by taking into account changes in personnel input, machinery operating status, material supply process, and duration of abnormal shutdowns; The constraints, time changes, and resource influencing factors between construction stages are linked and stored to form a node association table for progress analysis.

3. The collaborative prediction method for building construction progress based on federated learning according to claim 1, characterized in that, The process of establishing a local project schedule feature library is as follows: Read the construction constraint relationships in the node association table to obtain the construction stage where the time offset occurred and the associated subsequent stages; Based on the changes between the actual completion cycle and the planned completion cycle of the construction nodes, the corresponding delay duration, waiting time of subsequent nodes, resource compensation input and abnormal shutdown duration are read, and the time changes are correlated with the constraints in the node association table to determine the delay impact range and resource response situation corresponding to the status changes of different construction nodes. The completion rate, process transition time, personnel input, equipment input and material input of each construction node are uniformly processed, and the processed construction node status, the corresponding delay impact range and resource response status are organized according to the construction stage. The organized construction node status, process constraint characteristics, resource input characteristics, scope of delay impact, and changes in construction cycle are stored separately for each project to form a local project progress characteristic library.

4. The collaborative prediction method for building construction progress based on federated learning according to claim 1, characterized in that, The process of calculating the delay propagation characteristics caused by changes in the state of construction nodes using the process influence propagation algorithm is as follows: Read the actual completion time, process conversion record and subsequent node waiting record corresponding to the construction node in the project's local progress feature library to determine the construction node whose status has changed and the subsequent construction node directly affected. Determine the delay propagation path based on the sequential constraints between construction nodes. Read the current node's delay duration, the waiting time of subsequent nodes, and the compensation input of personnel, equipment, and materials along the delay propagation path. Calculate the ratio of the portion of the delay duration that is not offset by the waiting time of subsequent nodes to the current delay duration to obtain the initial delay propagation ratio between nodes. The degree of delay offset is determined based on the resource compensation input corresponding to the subsequent construction nodes, and the delay attenuation ratio is calculated in combination with the process interval time between adjacent construction nodes. The initial delay transmission ratio and the corresponding delay attenuation ratio are calculated step by step to obtain the delay impact ratio transmitted by each construction node along the delay propagation path. The path delay impact is calculated based on the delay impact ratio corresponding to each construction node and the actual completion time offset of the node. The impact is then accumulated according to the subsequent construction nodes corresponding to different delay propagation paths to obtain the delay propagation amount corresponding to the change in the state of the construction node. The delay propagation amount, delay propagation path, and delay impact ratio are used as delay propagation characteristics.

5. The collaborative prediction method for building construction progress based on federated learning according to claim 1, characterized in that, The process of generating local prediction parameters is as follows: Read the construction node status, process constraint characteristics, resource input characteristics and corresponding delay propagation characteristics from the local progress feature library, and divide the construction nodes with the same process constraint relationship according to the change of completion cycle. Based on the delay amount corresponding to the change in the construction node status and the number of subsequent construction nodes affected, determine the construction status that has the main impact on the completion cycle of the current node, and read the changes in personnel, equipment and material input under the construction status. Compare the changes in resource input with the actual completion cycle of the corresponding construction node to determine the cycle correction amount of different resource input states on the node completion cycle, and determine the time impact amount of the current construction node on subsequent nodes based on the delay transmission amount. The local progress prediction network is trained by inputting the completion cycle of construction nodes, the amount of delay, and the periodic correction amount corresponding to resource input, so as to obtain the local prediction parameters for the corresponding project.

6. The collaborative prediction method for building construction progress based on federated learning according to claim 1, characterized in that, The process of generating parameter matching coefficients based on the differences in feature representations of different project construction statuses in the local progress prediction network, changes in model prediction errors, and consistency of parameter update directions is as follows: Obtain the local prediction parameter update results of the local schedule prediction network for each project under the same construction stage, and read the construction node status, process constraint characteristics and resource input characteristics corresponding to the parameter update results. Based on the differences in the completion rate of nodes, the differences in process constraints, and the differences in the proportion of resource input for different construction stages, calculate the state difference value between the construction states of the projects. Read the node completion time prediction error before and after local parameter update for each project, calculate the error improvement corresponding to local parameter update, and calculate the parameter update consistency based on the adjustment direction and adjustment magnitude of the same parameter during the prediction parameter update process for different projects. The state difference value, error improvement amount and parameter update consistency are normalized, and the parameter matching coefficient between projects is determined based on the degree of influence of state difference on the applicability of parameter migration, the degree of error improvement and the parameter update consistency.

7. The collaborative prediction method for building construction progress based on federated learning according to claim 1, characterized in that, The process of forming a federal collaborative forecasting model is as follows: Obtain the local prediction parameters for each project and their corresponding parameter matching coefficients, and determine the project applicability of the parameters according to the construction node status, process constraint characteristics and resource input characteristics corresponding to the prediction parameters; By comparing the parameter update magnitudes of the same prediction parameter in different projects and combining the parameter matching coefficients between corresponding projects, the applicability of the prediction parameter to different projects can be determined. When the same prediction parameter is updated in the same direction in multiple projects and the corresponding parameter matching coefficient reaches the preset condition, the prediction parameter is determined as a shared parameter, and the aggregate weight of the shared parameter is calculated according to the parameter matching coefficient of each project. When the prediction parameter has a significant error improvement effect only in a single project or when different projects have different update directions for the prediction parameter, the prediction parameter is determined as a project adaptive parameter and does not participate in cross-project parameter aggregation. The local update values ​​identified as shared parameters are accumulated according to the corresponding aggregation weights to obtain the shared parameter update values. The shared parameter update values ​​and the project adaptive parameters retained by each project are then written into the federated collaborative prediction model.

8. The collaborative prediction method for building construction progress based on federated learning according to claim 1, characterized in that, The process of determining the correspondence between the current construction status and historical construction status of a target project using a phase state alignment network is as follows: Obtain the node completion rate, process transition status, and personnel, equipment, and material input status corresponding to the current construction stage of the target project; and read the construction stages with the same process constraint relationship from the historical construction status corresponding to the federated collaborative prediction model. Based on the difference between the current completion rate of the target project and the completion rate of the historical construction stage nodes, combined with the transition time difference between the process transition states and the difference in resource input ratio, the stage offset of the current construction state of the target project relative to the historical construction state is calculated. Adjust the corresponding positions of each construction node in the historical construction phase according to the phase offset, input the historical construction state after the phase offset adjustment and the current construction state of the target project into the phase state alignment network, determine the corresponding historical construction state according to the process constraint relationship and resource input status of the two, and call the parameters of the federated collaborative prediction model corresponding to the historical construction state.

9. The collaborative prediction method for building construction progress based on federated learning according to claim 1, characterized in that, The process of outputting the predicted completion times for the current construction node and subsequent construction nodes is as follows: Read the process constraint features and resource input features corresponding to the current construction status of the target project, and extract the construction node completion records with the same process constraint relationship from the local progress feature library based on the historical construction status determined by the stage status alignment network. Compare the current resource input status of the target project with the resource input status corresponding to the historical construction status, and determine the time correction amount caused by the difference in personnel, equipment and material input on the construction node completion cycle. Input the time correction amount into the prediction parameters corresponding to the federated collaborative prediction model, calculate the expected completion period of the current construction node, and calculate the time impact of subsequent construction nodes based on the delay propagation amount generated by the current construction node and its delay propagation path. The estimated completion period of the current construction node is corrected with the time correction amount, and the estimated completion period of the subsequent construction nodes is adjusted step by step according to the time impact amount corresponding to each subsequent construction node in the delay propagation path, so as to obtain the completion time prediction results of the current construction node and the subsequent construction nodes.

10. The collaborative prediction method for building construction progress based on federated learning according to claim 1, characterized in that, The process of updating the federated collaborative prediction model is as follows: The actual completion time, actual delay propagation range, and actual resource input changes of the target project after construction are obtained, and compared with the predicted completion time, predicted delay propagation range, and predicted resource input changes output by the federated collaborative prediction model to obtain the node completion time deviation, delay propagation deviation, and resource response deviation. Input the node completion time deviation, delay propagation deviation, and resource response deviation into the error diagnosis network, and determine the prediction branch that generates the prediction error based on the construction stage, process constraint relationship, and resource input status corresponding to each deviation. Based on the error change direction corresponding to the prediction branch, determine the local update direction and update magnitude of the node completion cycle parameter, delay propagation parameter and resource correction parameter, and perform incremental gradient update in the project local schedule prediction network. Obtain the degree of improvement in prediction error after each project completes local updates, jointly calculate the degree of improvement in prediction error with the parameter matching coefficient of the corresponding project, and adjust the proportion of shared parameters of each project participating in federated aggregation after updates. The shared parameters of each project are aggregated according to the adjusted participation ratio, and the adaptive parameters of the projects are retained in the local prediction network of the corresponding projects. The updated shared parameters and the adaptive parameters of the projects are used to form the next round of federated collaborative prediction model.

Images