Reinforcement learning-based engineering cost project cost control strategy optimization method and system

Abstract

The application provides a reinforcement learning-based engineering cost project cost control strategy optimization method and system, and relates to the technical field of engineering cost, which comprises the following steps: in step 2, according to an initial cost control strategy, resource allocation and cost control are performed at a current construction stage to obtain actual cost, progress and quality data; in step 3, the actual cost, progress and quality data are integrated to construct a multidimensional data distribution model, form an abstract decision space, fit discrete data into a continuous decision surface, calculate the curvature of a local area, and analyze the structural change characteristics of the decision surface; based on the structural change characteristics, the abstract decision space is divided into regions, the original data is mapped to the corresponding partitions, the intrinsic characteristics of each partition are extracted, and strategy correction parameters are generated. The application realizes dynamic self-adaptive optimization of the cost control strategy of the engineering cost project, and improves the accuracy of the cost control of the engineering project.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of engineering cost, in particular to an engineering cost project cost control strategy optimization method and system based on reinforcement learning. BACKGROUND

[0002] In the field of engineering project cost control, traditional management methods mostly rely on static budgeting and phased cost accounting, which are difficult to respond to the dynamically changing environment during construction. Especially in the face of uncertain factors such as fluctuations in building material market prices, frequent design changes, and delays in construction period, existing methods mostly have some limitations, resulting in deviations between cost control and actual construction status.

[0003] For example, in a construction project of a certain urban rail transit station, a detailed cost control plan was initially developed based on historical data; however, when the construction entered the main structure stage, the prices of major building materials continued to rise, and due to the more complex geological conditions than expected, some engineering designs needed to be changed, resulting in delays in local construction period. At this time, traditional cost control methods mostly rely on manual experience to adjust the budget, and can only make post-corrections for a single cost indicator, making it difficult to synchronize the dynamic balance between cost, progress, and quality. When increasing human resources to catch up with the progress, it may ignore the quality risks and resource allocation efficiency, and instead cause higher rectification costs in the subsequent stage. This adjustment method based on discrete events and manual judgment lacks continuous analysis and overall optimization of multi-dimensional construction data, and is prone to one-sidedness and lag in control strategy, thereby affecting the realization of the overall cost target of the project. SUMMARY

[0004] The technical problem to be solved by the present application is to provide an engineering cost project cost control strategy optimization method and system based on reinforcement learning, which realizes dynamic self-adaptive optimization of the engineering cost project cost control strategy and improves the accuracy of engineering project cost control.

[0005] To solve the above technical problems, the technical solutions of the present application are as follows:

[0006] In a first aspect, the engineering cost project cost control strategy optimization method based on reinforcement learning comprises:

[0007] Step 1: Obtain the initial budget of the engineering project, the real-time construction environment state, and the past cost control records, and construct an initial cost control strategy;

[0008] Step 2: According to the initial cost control strategy, perform resource allocation and cost control in the current construction stage, and obtain actual cost, progress, and quality data;

[0009] Step 3, integrate actual cost, progress and quality data, build multi-dimensional data distribution model, form abstract decision space; fit discrete data into continuous decision surface, calculate curvature of local area, and analyze structural change characteristics of decision surface; based on structural change characteristics, divide regions in abstract decision space, map original data to corresponding partitions, extract internal characteristics of each partition, and generate strategy correction parameters;

[0010] Step 4, based on strategy correction parameters, combine actual cost, progress and quality data, normalize and map multi-dimensional data points to unit hypersphere; calculate the shortest arc length to quantify difference measure between data points, and perform difference analysis and multi-objective collaborative evaluation to obtain cost control comprehensive evaluation results;

[0011] Step 5, based on cost control comprehensive evaluation results and strategy correction parameters, use reinforcement learning model to optimize current strategy, and generate optimized cost control strategy;

[0012] Step 6, apply the optimized cost control strategy to the next construction stage, and real-time monitor building material price fluctuation, design change and delay information, and realize dynamic optimization control.

[0013] In the second aspect, the project cost control strategy optimization system based on reinforcement learning comprises:

[0014] The construction module is used to obtain the initial budget of the project, the real-time construction environment state and the past cost control records, and to construct the initial cost control strategy;

[0015] The execution module is used to execute resource allocation and cost control according to the initial cost control strategy in the current construction stage, and to obtain actual cost, progress and quality data;

[0016] The analysis module is used to integrate actual cost, progress and quality data, build multi-dimensional data distribution model, form abstract decision space; fit discrete data into continuous decision surface, calculate curvature of local area, and analyze structural change characteristics of decision surface; based on structural change characteristics, divide regions in abstract decision space, map original data to corresponding partitions, extract internal characteristics of each partition, and generate strategy correction parameters;

[0017] The evaluation module is used to normalize and map multi-dimensional data points to unit hypersphere based on strategy correction parameters, combine actual cost, progress and quality data; calculate the shortest arc length to quantify difference measure between data points, and perform difference analysis and multi-objective collaborative evaluation to obtain cost control comprehensive evaluation results;

[0018] An optimization module is configured to optimize the current strategy based on the cost control comprehensive evaluation result and the strategy correction parameter by using the reinforcement learning model, and generate an optimized cost control strategy.

[0019] A feedback module is configured to apply the optimized cost control strategy to the next construction stage, and monitor the building material price fluctuation, design change and schedule delay information in real time, so as to realize dynamic optimization control.

[0020] The above scheme of the present application at least has the following beneficial effects:

[0021] By integrating the initial budget, real-time construction environment state and past cost control records to construct an initial cost control strategy, the initial strategy is fully integrated with historical experience and current actual scene, the pertinence and operability of the strategy are improved, and the risk of cost out of control in the initial stage is reduced; secondly, resource allocation and cost control are simultaneously performed in the current construction stage, and actual cost, progress and quality data are collected in real time, so that real-time linkage of the control process and data feedback is realized, resource allocation is always matched with the current construction demand, and resource waste or insufficient supply is avoided; thirdly, a multi-dimensional data distribution structure is constructed by integrating the actual data, discrete data is fitted into a continuous decision surface, strategy correction parameters are extracted by combining curvature analysis and region division, a quantitative basis is provided for strategy adjustment, and dynamic iterative optimization of the cost control strategy is realized. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 is a flowchart of the method for optimizing the cost control strategy of the project cost of the engineering project based on reinforcement learning provided by the embodiment of the present application.

[0023] Figure 2 is a schematic diagram of the system for optimizing the cost control strategy of the project cost of the engineering project based on reinforcement learning provided by the embodiment of the present application. DETAILED DESCRIPTION

[0024] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be accurately conveyed to those skilled in the art.

[0025] As shown in Figure 1 The embodiment of the present application proposes a method for optimizing the cost control strategy of the project cost of the engineering project based on reinforcement learning, which comprises the following steps:

[0026] Step 1: obtaining the initial budget, real-time construction environment state and past cost control records of the engineering project, and constructing an initial cost control strategy;

[0027] Step 2, according to the initial cost control strategy, resource allocation and cost control are performed in the current construction stage to obtain actual cost, progress and quality data;

[0028] Step 3, the actual cost, progress and quality data are integrated to build a multi-dimensional data distribution model to form an abstract decision space; the discrete data is fitted into a continuous decision surface, the curvature of the local area is calculated, and the structural change characteristics of the decision surface are analyzed; based on the structural change characteristics, the region is divided in the abstract decision space, the original data is mapped to the corresponding partition, the intrinsic characteristics of each partition are extracted, and the strategy correction parameters are generated;

[0029] Step 4, based on the strategy correction parameters, the actual cost, progress and quality data are combined to normalize and map the multi-dimensional data points to a unit hypersphere; the shortest arc length is calculated to quantify the difference between the data points, and difference analysis and multi-objective collaborative evaluation are performed to obtain the cost control comprehensive evaluation result;

[0030] Step 5, based on the cost control comprehensive evaluation result and the strategy correction parameters, the current strategy is optimized using a reinforcement learning model to generate an optimized cost control strategy;

[0031] Step 6, the optimized cost control strategy is applied to the next construction stage, and real-time monitoring of building material price fluctuations, design changes and schedule delay information is implemented to realize dynamic optimization control.

[0032] In the embodiment of the present application, the initial budget, real-time construction environment state and past cost control records are integrated to build an initial cost control strategy, so that the initial strategy fully integrates historical experience and current actual scene, improves the pertinence and operability of the strategy, and reduces the risk of cost out of control in the initial stage; secondly, resource allocation and cost control are simultaneously performed in the current construction stage, and actual cost, progress and quality data are collected in real time to realize real-time linkage of control process and data feedback, ensure that resource allocation always matches current construction demand, and avoid resource waste or insufficient supply; thirdly, the multi-dimensional data distribution structure is built by integrating the actual data, the discrete data is fitted into a continuous decision surface, the strategy correction parameters are extracted by combining curvature analysis and region division to provide quantitative basis for strategy adjustment, and dynamic iterative optimization of the cost control strategy is realized.

[0033] In a preferred embodiment of the present application, the above step 1 of obtaining the initial budget of the engineering project, the real-time construction environment state and the past cost control records to build the initial cost control strategy can include:

[0034] In the embodiments of the present application, the budget-related core data is extracted from the pre-project filing documents, bidding information, bid notice and initial draft of construction organization design, including but not limited to bid price details, bill of quantities, sub-item engineering pricing table, fee standard explanation and provisional amount, professional engineering provisional price agreement file; the budget is split according to the engineering division and sub-item to determine the specific amount of each sub-item labor, materials, machinery, measure fee and regulation fee, form a hierarchical budget list, and mark the calculation basis of each sub-item budget; compare the deviation items of the bid control price in the bidding document and the bid budget, check whether the budget preparation is missing the key process, and for the sub-item with high provisional amount ratio and abnormal unit price, jointly review with the cost personnel and the construction director, correct the budget deviation, and form the final confirmed initial budget benchmark; organize construction, technical and cost personnel to carry out site reconnaissance, record the site basic conditions and existing facility conditions, and determine the construction limiting factors; through the channels such as supplier real-time quotation, building material market quotation platform and local housing construction department announcement, the current labor unit price, main building material market price and supply cycle, and mechanical leasing unit price are obtained; the influencing factors such as weather warning information, local policy adjustment and logistics transportation cost fluctuation are collected synchronously; the collected information is classified and recorded according to static conditions and dynamic variables, and the information update cycle is marked; from the enterprise project management database, the completed / in-construction projects similar to the type, size and process of the present project are screened, the core records are extracted, including the sub-item cost deviation data, resource consumption account, cost control measure execution record and final cost target achievement situation; the abnormal data caused by special sudden situations in the historical projects are excluded, and the conventional records with reference value are retained; the cost deviation points and corresponding response measures effect frequently appeared in similar projects are focused on; the cost control rules of historical projects, such as cost proportion in different construction stages, optimal proportion of resource allocation, common risk points, are summarized to form a historical data reference report, which provides a basis for initial strategy making.

[0035] Based on the initial budget details and historical data, establish the basic allocation ratios for labor, materials, and machinery at each construction stage, and determine the upper limit for the usage of each resource and its arrival time to avoid resource idleness or shortage. Based on the initial budget, set cost warning lines and stop-loss lines for each sub-project, such as an 8% overrun of labor costs triggering a warning and a 15% overrun triggering a stop-loss. Clarify the response process after a warning and the emergency measures corresponding to the stop-loss line. Combining the real-time construction environment and historical experience, set the constraints on cost for progress and quality, such as the foundation construction period not exceeding 10% of the budget period and the concrete strength qualification rate needing to reach over 95%, and determine the priority adjustment principles when these three factors are unbalanced. For labor, determine the shift schedule and attendance calculation standards for each trade, and limit overtime frequency to control additional costs. For materials, establish a procurement and acceptance process. The plan includes: process and inventory management solutions; for machinery, optimizing shift scheduling to improve equipment utilization and reduce idle shift costs; establishing cost expenditure approval processes and authorizing at different levels based on amount; setting regular accounting cycles and standardizing accounting forms and data reporting paths; embedding simplified risk response plans into the strategy based on real-time construction environment variables and historical risk points, such as reserving a certain percentage of price adjustment reserves for fluctuations in building material prices; for site constraints, pre-setting temporary facility optimization plans; organizing cost, construction, technical, and supervision personnel to conduct strategy reviews to verify whether the strategy matches the initial budget, construction environment, and historical experience, focusing on reviewing the rationality of resource allocation, the scientific nature of early warning thresholds, and the feasibility of risk contingency plans; collecting feedback from all parties, iteratively adjusting the strategy content, and forming a final executable initial cost control strategy.

[0036] In a preferred embodiment of the present invention, step 2 above, which involves performing resource allocation and cost control in the current construction phase according to the initial cost control strategy to obtain actual cost, schedule, and quality data, may include:

[0037] In the embodiment of the present application, step 220, based on the initial cost control strategy, the specific type, quantity and time requirement of artificial, material and mechanical resources for the current construction stage are obtained to form an initial resource allocation scheme; specifically including: the construction supervisor combines with the cost personnel and the technical personnel to determine the detailed requirements of artificial, material and mechanical resources one by one and integrate them into a scheme based on the resource allocation benchmark in the initial cost control strategy, the budget quota of each item and the construction progress constraint, combined with the specific process requirements of the current construction stage; in terms of artificial resources, first determine the specific type, which needs to cover the types of reinforcing workers, concrete workers, formwork workers, scaffolding workers, foundation pit excavation workers and other types of work that adapt to the construction of main structure and complex geological treatment; then calculate the specific quantity, according to the current stage of steel binding engineering quantity, concrete pouring engineering quantity, formwork erection engineering quantity, combined with the daily work load standard of artificial in the initial cost control strategy, the required number of people is calculated by the construction supervisor according to the daily work load of each type of work × planned construction days, while 3 to 5 mobile personnel are additionally reserved to cope with the process adjustment demand caused by complex geological conditions; finally, the time requirement is determined, and the work type nodes are arranged according to the sequence of the process + material arrival time, such as reinforcing workers need to arrive one day before the arrival of reinforcing materials, concrete workers need to arrive on the same day after the completion of formwork erection, to ensure seamless connection with subsequent processes; in terms of material resources, the specific types are determined as HRB400E grade reinforcing steel, C30 and C40 grade concrete, waterproof roll material, scaffold steel pipe and other materials that adapt to the main structure of rail transit; when calculating the specific quantity, the cost personnel calculate the quantity according to the current stage of the bill of quantities, for example, the quantity of reinforcing steel is calculated by reinforcing binding engineering quantity × (1+reinforcing steel loss rate), and the quantity of concrete is calculated by concrete pouring engineering quantity × (1+concrete loss rate), while considering the risk of building material price fluctuation, a small amount of standby materials are additionally reserved; the time requirement is determined in combination with the construction progress nodes, such as reinforcing materials need to arrive 3 days before the start of reinforcing binding process, and concrete materials need to arrive in time periods according to pouring batches, and the arrival time of each batch is 1 day in advance to interface with the supplier to avoid the impact of material lag on construction; in terms of mechanical resources, the specific types include tower crane, concrete pump truck, excavator, reinforcing steel cutting machine, electric welding machine and other equipment that adapt to the construction of main structure and foundation pit treatment; the quantity is calculated by the construction supervisor combined with the mechanical administrator, for example, the quantity of concrete pump truck is calculated by daily concrete pouring volume ÷ single-day production capacity of single concrete pump truck, and the quantity of excavator is calculated by remaining excavation volume ÷ single-day production capacity of single excavator, while the corresponding number of auxiliary machinery is matched; the time requirement is matched with the material arrival and process advancement, such as the excavator needs to arrive on the same day when the foundation pit excavation process starts, and the tower crane needs to complete installation and debugging before the arrival of reinforcing and formwork materials to ensure that resources and processes are synchronized, and the specific type, quantity and time requirement of the above artificial, material and mechanical resources are arranged into a book to form an initial resource allocation scheme.

[0038] Step 221, according to the initial resource allocation scheme, the on-site resource scheduling and cost expenditure control are performed to generate resource execution records including actual resource consumption and corresponding cost expenditure; specifically including: in terms of manual scheduling, the manual administrator organizes personnel to work according to the type, quantity and approach time determined by the scheme, checks the number of daily attendance, calculates the daily manual cost according to the actual number of attendance x daily wage standard of each type of work, and strictly prohibits unplanned personnel dispatch; for the situation that the complex geological conditions lead to partial process rework, the temporary allocation of mobile personnel can be made only after the joint confirmation of the construction responsible person and the cost specialist, and the temporary allocation reason, number and working hours are recorded, and the corresponding labor cost is calculated; in terms of material scheduling, the material administrator interfaces with the supplier according to the time and quantity agreed in the scheme, checks the specifications, quantity and quality certificate documents with the quality inspection personnel when the materials are put into the field, calculates the material procurement cost according to the actual quantity x procurement unit price after acceptance, calculates the actual usage amount of materials every day, and calculates the actual usage amount = daily quantity in the field - daily inventory - previous day inventory, and records the material loss situation, for the part exceeding the initial strategy loss rate, the reason needs to be found out and jointly signed and confirmed by the material administrator and the cost specialist, and is included in the cost expenditure record; in terms of mechanical scheduling, the mechanical administrator arranges the mechanical approach and debugging according to the scheme, records the actual number of mechanical usage shifts every day, calculates the mechanical rental cost according to the actual number of shifts x corresponding mechanical shift rental unit price, and at the same time, calculates the mechanical maintenance and repair cost by summing up the single maintenance and repair consumable cost + labor service fee; the mechanical scheduling sequence is optimized to avoid mechanical idling, and for the situation that the mechanical standby is caused by design changes, the mechanical purpose is adjusted in time, and the standby time and cost loss are recorded; in terms of cost expenditure control, the cost specialist audits each expenditure according to the initial budget and resource allocation scheme, and the single expenditure amount below 50,000 yuan is paid after being approved by the project responsible person; the single expenditure amount of 50,000 yuan and above needs to be reported to the enterprise cost management department for review and approval before payment, and all expenditures need to keep the vouchers. The cost specialist summarizes the actual consumption of labor, materials and machinery and the corresponding cost expenditure every day, and performs stage summary every week to form detailed resource execution records, and the record content needs to cover resource type, planned quantity, actual quantity, difference quantity, corresponding cost, difference reason, and is filed after being signed and confirmed by all parties.

[0039] At step 222, based on the resource execution record, real-time collection of cost flow data actually occurring in the construction process, engineering progress completion data and construction quality key indicator detection data; Specifically including: based on the resource execution record, synchronous collection of three types of core data to ensure data real-time and accuracy, cost flow data collection is responsible for the cost personnel, daily docking financial personnel and various resource administrators, collect labor wage payment documents, material purchase invoices, mechanical rental agreements and payment vouchers, maintenance and repair expense documents, etc., classified according to each sub-item cost + comprehensive expense, among which the sub-item cost is summarized according to the actual cost of labor + the actual cost of materials + the actual cost of machinery, and the comprehensive expense includes site management fee, temporary facility fee, etc., and the actual amount of the current period is recorded, while the cost expenditure data in the resource execution record is compared for consistency, and for the case of incomplete bills or amount deviation, timely verification and correction with the corresponding post personnel to ensure that each cost flow can be traced back; The engineering progress completion data collection is responsible for the construction statistician, who measures the actual completion of each sub-item engineering at the construction site every day, combines the construction progress plan corresponding to the initial resource allocation scheme, calculates the daily and cumulative progress completion ratio according to the actual completion amount ÷ the total engineering quantity of the sub-item, for example, the steel binding sub-item calculates the progress according to the actual binding steel length ÷ the total steel binding length of the stage, and the concrete pouring sub-item calculates the progress according to the actual pouring volume ÷ the total concrete pouring volume of the stage; For the progress lagging caused by complex geology and building material price fluctuations, record the lagging reasons and lagging time, and mark the influence on the subsequent process, and record the progress data into the construction log every day, and associate the resource consumption data in the resource execution record.

[0040] Through on-site multi-post collaborative scheduling, real-time patrol control and cost expenditure synchronous audit, dynamic control of resource consumption and cost expenditure is realized.

[0041] In a preferred embodiment of the present application, step 3 above integrates actual cost, progress and quality data to construct a multi-dimensional data distribution model, forming an abstract decision space; fits discrete data into a continuous decision surface, calculates the curvature of the local area, and analyzes the structural change characteristics of the decision surface; based on the structural change characteristics, divides the region in the abstract decision space, maps the original data to the corresponding partition, extracts the intrinsic characteristics in each partition, and generates strategy correction parameters, which can include:

[0042] In the embodiment of the present application, in step 330, the actual cost data, progress data and quality data are standardized, and the three-dimensional data after standardization is integrated into a multi-dimensional vector in a unified format, wherein each dimension corresponds to a quantitative indicator of cost, progress and quality respectively; specifically, taking the multi-dimensional data set of the main structure construction of the rail transit station collected in step 222 as the core object, the operation is carried out according to the principle of dimension processing and unified integration, and the whole process is combined with the actual process and data characteristics of the project for detailed processing; in the aspect of cost data standardization, the actual cost flow data is split according to the core process, the labor, material and mechanical sub-item costs of the three main processes of steel binding, formwork installation and concrete pouring are determined, and then the standardization value range of each sub-item cost is separately defined; taking the steel material cost of the steel binding process as an example, first, all steel procurement expenditure data collected in the current main structure construction stage is extracted, and the minimum actual expenditure and the maximum actual expenditure in the sub-item cost are screened out for confirmation, which are used as the upper and lower limits of the standardization of the sub-item cost; then, the steel material cost is standardized and converted batch by batch, the conversion logic is that the actual expenditure of each steel procurement is subtracted from the minimum actual expenditure of the sub-item, and the result is divided by the difference between the maximum actual expenditure and the minimum actual expenditure of the sub-item; if the actual expenditure of each steel procurement is equal to the minimum expenditure, the standardized value is 0, if it is equal to the maximum expenditure, the standardized value is 1, if it exceeds the maximum expenditure, it is still counted as 1, and if it is lower than the minimum expenditure, it is still counted as 0; similarly, the standardized conversion of the labor cost and the mechanical cost of the steel binding process, and the sub-item costs of the formwork installation and concrete pouring processes are completed, and then the average value of the standardized values of all sub-item costs of each process is calculated, the calculation method is that the standardized values of the sub-item costs of the process are added up, and the total sum is divided by the number of sub-item costs, and the average value is used as the cost dimension standardized value of the corresponding process, which ensures to cover all sub-item costs and avoid distortion of single sub-item data; in the aspect of progress data standardization, taking the progress plan of each process set in the initial resource allocation scheme as the benchmark, the daily planned completion amount and the stage planned completion amount of each process of steel binding, formwork installation and concrete pouring are split, and then the value range of the progress standardization of each process is defined; taking the concrete pouring process as an example, the minimum planned completion amount is the basic pouring amount that must be completed daily, and the maximum planned completion amount is the upper limit pouring amount after optimizing the progress; then, the progress standardized value is calculated daily, the calculation logic is that the actual pouring amount of concrete is subtracted from the minimum planned completion amount daily, and the result is divided by the difference between the maximum planned completion amount and the minimum planned completion amount daily; if the actual pouring amount of concrete exceeds the maximum planned completion amount, the standardized value is counted as 1, if it is lower than the minimum planned completion amount, the standardized value is counted as 0, and the reason for the progress lag is also marked; the formwork installation and steel binding processes are completed by the same logic, and the progress standardized value of each process is updated daily to ensure that it fits the real-time construction progress.In terms of quality data standardization, the indicators are classified into qualified range indicators and deviation indicators, and the construction specifications and design requirements are used as the benchmark. For the qualified range indicators, the tensile strength of the steel bar is taken as an example. The design requirement qualified range is 400-500 MPa, which is used as the upper and lower limits of standardization. The calculation logic is that the actual test value of a single sample is subtracted by 400 MPa, and the result is divided by the difference between 500 MPa and 400 MPa. If the test value is lower than 400 MPa (unqualified), the standardized value is 0. If it is higher than 500 MPa (better than the standard), the standardized value is 1. For the deviation indicators, the thickness of the steel bar protective layer is taken as an example. The design allows the deviation range to be ±5 mm, which is the minimum allowable deviation value of -5 mm and the maximum allowable deviation value of 5 mm. The calculation logic is 1 minus (the absolute value of the actual deviation value minus the absolute value of the minimum allowable deviation value, divided by the difference between the absolute value of the maximum allowable deviation value and the absolute value of the minimum allowable deviation value). The closer the actual deviation value is to 0, the closer the standardized value is to 1. If the actual deviation value exceeds the allowed range, the standardized value is 0. The quality standardization value of each process is the average of the standardized values of all detection indicators of that process. After the standardization of the three types of data is completed, a three-dimensional multi-dimensional vector is generated for each process and each day in the format of a vector. The three dimensions of each vector represent the cost standardization value, the progress standardization value, and the quality standardization value of that day's process, respectively. At the same time, the vector association information is labeled, including the construction date, the specific process part, the corresponding resource execution record number, and whether there are abnormal factors, ensuring that each vector can be traced back to the original construction data and the on-site situation.

[0043] Step 331, map the multi-dimensional vector into the preset abstract decision space to form a set of data points representing the project state, and analyze the distribution of the data points in the decision space to construct a multi-dimensional data distribution structure reflecting the dynamic correlation between cost, schedule, and quality; Specifically, first determine the preset abstract decision space as a three-dimensional space, and the three coordinate axes of the space correspond to the cost, schedule, and quality standardization dimensions obtained in step 330, respectively. The value range of each coordinate axis is set to 0 to 1, which is completely matched with the value range of the standardized data. Any point in the space uniquely corresponds to a standardized combination of cost, schedule, and quality, which can directly represent the project control state of a certain process at a certain time period. Then, map all multi-dimensional vectors generated in step 330 into the three-dimensional space one by one. During mapping, operate in order according to the process part and date marked by the vector. Each vector corresponds to an independent data point in the space, and each data point is supplemented with complete associated information, including corresponding construction period, core process and part, resource consumption deviation, cost deviation, schedule deviation, quality test results, and abnormal factor details. After mapping is completed, distribution rule analysis is carried out. First, identify the aggregation area and dispersed area of the data points by using the method of regional investigation and time period comparison. When investigating the aggregation area, divide it into five intervals according to the value range of the coordinate axis: 0-0.2, 0.2-0.4, 0.4-0.6, 0.6-0.8, and 0.8-1.0. Investigate the distribution of data points in each interval, such as whether there is an aggregation area with cost standardization value 0.3-0.5, schedule standardization value 0.7-0.9, and quality standardization value 0.8-1.0. Such areas represent controllable cost, advanced schedule, and excellent quality control state, and the process characteristics of the data points in this area are recorded. When investigating dispersed data points, focus on independent data points that deviate from all aggregation areas. Check their associated information one by one, analyze the reasons for dispersion, such as a data point with cost standardization value 0.85 (cost is too high), schedule standardization value 0.35 (schedule is lagging), and quality standardization value 0.9 (quality is qualified). Check the resource execution records and on-site construction logs to confirm whether the steel reinforcement procurement cost has exceeded the budget due to a significant increase in building material prices, which has affected the construction schedule. Then, mark the type and degree of influence of the abnormal factors corresponding to the dispersed point. Subsequently, for the aggregation area and dispersed data points, deeply analyze the dynamic correlation between cost, schedule, and quality. For the aggregation area, analyze the trends of the three elements of the data points in the area, such as whether the schedule standardization value increases from 0.7 to 0.9 when the cost standardization value increases from 0.3 to 0.5, i.e., increasing mechanical and labor input leads to cost increase, while accelerating the construction schedule, and whether the quality standardization value remains at 0.8 or more; for discrete data points, analyze the interference of abnormal factors on the correlation of the three elements, for example, when the design is changed due to complex geological conditions, the corresponding data points are mostly characterized by an increase in cost standardization value, a decrease in progress standardization value, and a maintenance of qualified quality standardization value, that is, the design change requires additional investment in labor and materials, resulting in cost overruns and progress delays, but the quality is strictly controlled according to the new design requirements and no problems have occurred; combined with the core needs of the main structure construction of rail transit stations, the influence law of three key abnormal factors, such as building material price fluctuation, geological condition change, and design change, on the correlation of the three elements is focused on, a multi-dimensional data distribution structure is formed, the structure is presented in the form of written accounts, the coordinate range of each cluster area, the corresponding three-element correlation mode, and the adaptation to the construction scene are determined, the coordinate position of the discrete data points, the discrete reason, the abnormal factor influence path, and the core law of the three-element deviation caused by different abnormal factors are determined, to ensure that the structure reflects the cooperation relationship between the on-site construction state and the three elements.

[0044] Step 332, according to the multi-dimensional data distribution structure, the statistical distribution characteristics of the data point set are utilized to finally generate an abstract decision space defined by the multi-dimensional data distribution structure; specifically including: first, the statistical distribution characteristics of the data point set are extracted, and the multi-dimensional data distribution structure is used as the basis throughout the process to classify and process according to the aggregation area and the discrete data point; for each aggregation area, first, the total number of data points in the area is counted, and then the cost standardized average value, the progress standardized average value, and the quality standardized average value of all data points in the area are calculated, and the calculation method is to add the standardized values of all data points in the corresponding dimension in the area, and divide the obtained sum by the total number of data points in the area; at the same time, the fluctuation range of the standardized value of each dimension is counted, that is, the maximum value minus the minimum value of the standardized value of the corresponding dimension in the area, and the fluctuation reason is recorded; for example, the cost-progress and quality balanced aggregation area, the total number of data points is 50, the cost standardized average value is 0.42, the fluctuation range is 0.08 (the maximum value is 0.47, and the minimum value is 0.39), the progress standardized average value is 0.78, the fluctuation range is 0.12 (the maximum value is 0.85, and the minimum value is 0.73), the quality standardized average value is 0.86, and the fluctuation range is 0.07 (the maximum value is 0.90, and the minimum value is 0.83), and the fluctuation reason is normal process connection difference without abnormal factor interference; for the discrete data points, the relative distance from each aggregation area is counted, such as a certain discrete point being closest to the balanced aggregation area, only the cost dimension value deviating from the upper limit of the area by 0.3, the progress and quality dimension values are within the range of the area, and the number and offset amplitude of discrete data points caused by different abnormal factors are classified and counted; subsequently, combined with the control constraint conditions of the main structure construction of the rail transit station, each aggregation area is screened and optimized, and the aggregation areas exceeding the constraint conditions are eliminated, such as the areas with a cost standardized average value exceeding 0.7 and areas with a quality standardized average value less than 0.6; the feasible aggregation areas meeting the control requirements are retained, the core adaptation scenarios of each feasible area are determined, such as the balanced aggregation area adapting to the normal construction period without abnormal factors, and the progress priority aggregation area adapting to the construction period requiring to catch up; at the same time, the abnormal factors, the offset rule, and the coping boundary corresponding to the discrete data points are included in the definition of the decision space, and the offset range allowed by the data points when different abnormal factors occur is determined, such as the cost dimension value offset not exceeding 0.2 caused by the price rise of building materials, which can still be adjusted by optimizing resource allocation, and exceeding 0.2 requires adjusting the initial cost strategy, and the control direction after offset, such as the progress dimension value offset being less than 0.3, can be through the extension of the operation length catch, without additional cost; finally, the integration of the statistical characteristics of the feasible aggregation area, abnormal factors coping rules, three-element correlation law, generate the formal abstract decision space, the decision space is still three-dimensional space, with the feasible aggregation area as the core feasible decision domain, mark the statistical characteristics and adaptation scene of each area; with the offset range of discrete data points as the abnormal early warning boundary, mark the abnormal type and coping direction corresponding to different boundaries; At the same time, determine the interpretation rule of any data point in the space, such as the data point is located in a feasible area, which represents the normal state of control; beyond the feasible area but not reach the early warning boundary, which represents a slight deviation, need to fine-tune resource allocation; beyond the early warning boundary, which represents a serious deviation, need to start strategy correction; The decision space is completely defined by multi-dimensional data distribution structure, all rules and characteristics come from the actual construction data on site, adapt to the dynamic control needs of the main structure construction of rail transit station.

[0045] Step 333, based on the discrete data points of the cost, schedule and quality indicators distributed in the abstract decision space, a continuous decision surface is formed by using a surface fitting algorithm to fit the discrete data points; Specifically including: taking the abstract decision space generated in step 332 as the carrier, locking all discrete data points, including conventional construction data points, abnormal data points caused by building material price fluctuations, geological changes and design changes, selecting a cubic spline interpolation algorithm to carry out fitting, and throughout the process adhering to the process characteristics and data characteristics of the main structure construction of the rail transit station; First, data point preprocessing is carried out, all discrete data points are first classified according to construction process and construction period to form a subdivided data group, for example, steel binding process-every morning-conventional construction group, concrete pouring process-every afternoon-building material price fluctuation group, to ensure that the fitting result can accurately match different construction scenarios; Screening extreme abnormal data points, extreme abnormal points are defined as data points deviating from the mean value of the three-dimensional standardized value of the majority of data points in the same group, such data points are directly excluded to avoid interfering with the fitting accuracy; Finally, set the weight of the remaining data points, the weight of the conventional construction data points is uniformly set to 1, and the weight of the data points affected by abnormal factors is set to 1.2, the weight setting basis is that the abnormal scene has more significant influence on cost control, so the contribution of the abnormal scene in the fitting curve should be improved to ensure that the curve can truly reflect the correlation trend of the three elements of the abnormal scene. Then, according to the logic of the cubic spline interpolation algorithm, each process construction progress time axis is taken as a clue, the discrete data points in the same subdivision data group are arranged in time sequence, an independent fitting curve is constructed between two adjacent data points, and each curve segment must pass through the two end point data points accurately, and the tangent slope of the adjacent two curve segments at the junction point is consistent, that is, the slope of the previous curve segment at the junction point is determined first, and then the starting slope of the next curve segment is adjusted to be the same as the previous one, so as to avoid discontinuities such as breaks and sharp corners in the curve. For example, the conventional group of reinforcement binding process, taking the discrete data points of No. 1 foundation reinforcement binding for two consecutive mornings, a smooth curve between the two points is constructed, the values of the points in the curve are determined according to the principle of gradual change of the values of the two end points and smooth transition of the slope, and then all the segmented curves of all time periods and all parts of the process are spliced one by one to form a process-level fitting curve segment. Then, the curve integration and verification are carried out, the fitting curve segments of reinforcement binding, formwork installation and concrete pouring processes are integrated, the data points at the process junction are fitted and adjusted again to ensure smooth transition of the curve segments of different processes at the junction without obvious protrusions or depressions. After fitting, 10% of the original data points are randomly selected, the coordinate difference between the actual coordinates of the data points and the coordinates of the corresponding curve position is compared, if the difference is too large, the weight of the corresponding abnormal data point is adjusted again, and fitting is carried out again until the deviation of all sampled data points and the curve is within a reasonable range, and finally a continuous and smooth decision curve covering the entire abstract decision space is formed. The curve can completely reproduce the dynamic correlation trend of the cost, progress and quality elements with the construction process and scene changes, and clearly show the influence of different abnormal factors on the correlation of the three elements.

[0046] At step 334, the geometric curvature of each local position of the decision surface is calculated, and according to the numerical distribution of the geometric curvature, the structural mutation region and the gentle region of the curvature are identified; specifically including: the process operation of drawing local region by point, manually calculating curvature, and partitioning statistics distribution, first, a local region is drawn, taking any data point on the decision surface as the center, selecting 10 adjacent data points around the center according to the principle of nearest space distance, and together forming a local region, each local region corresponds to a small construction scene segment in the abstract decision space, and the basis for selecting 10 adjacent points is to cover the related data around the center data point and to avoid the distortion of curvature calculation caused by too large region; then, the geometric curvature is calculated region by region, first, the normal vector direction of the center data point is determined, the direction of each connecting line is manually marked through the space coordinates of the center data point and the 10 adjacent data points, and then the normal vector (the direction perpendicular to the tangent plane of the surface) is determined according to the concave-convex trend of the decision surface at the center data point; secondly, the included angle between each adjacent point connecting line and the normal vector is calculated, and the total included angle is obtained by adding the 10 included angle values one by one, and the average included angle is obtained by dividing the total included angle by 10 (the number of adjacent data points); thirdly, the distance between the 10 adjacent data points is calculated, and the distance is calculated between each two adjacent points, and 9 groups of distance values are obtained, and the total distance is obtained by adding the 9 groups of distance values, and the average distance is obtained by dividing the total distance by 9 (the number of distance groups); fourthly, the curvature value is calculated, which is calculated according to the logic of dividing the average included angle by the average distance, the larger the average included angle and the smaller the average distance, the larger the curvature value, which represents that the local region surface fluctuates more dramatically, and the corresponding three element correlation in the construction scene changes more significantly; the smaller the average included angle and the larger the average distance, the smaller the curvature value, which represents that the local region surface is more gentle, and the corresponding three element correlation in the construction scene is more stable; after the curvature of all local regions is calculated, all curvature values are arranged in order from small to large, and the curvature distribution interval is drawn, first, the range where most of the curvature values concentrate is counted, and the range is defined as the gentle interval, the curvature value in this region fluctuates in a narrow range, and the curvature value of the adjacent region has small difference, which corresponds to the normal construction scene without abnormal factor interference and with three elements changing in a fixed proportion; then, the region where the curvature value suddenly exceeds the upper limit of the gentle interval and has significant difference with the curvature value of the adjacent region is identified as the structural mutation region, which corresponds to the scene where the three element correlation changes dramatically, such as the sharp rise of building material price leading to the sharp rise of cost while the progress is maintained stable, or the sudden change of geological conditions leading to the fluctuation of both cost and progress; at the same time, the associated information of the original data point of each structural mutation region is checked one by one, the type of abnormal factor and the influence scene corresponding to the region are marked, and the corresponding account of local region-curvature value-abnormal factor is formed.

[0047] At step 335, the abstract decision space is divided into several sub-zones with different structural characteristics according to the natural boundaries between the structural mutation zones and the flat zones; specifically including: taking the curvature distribution account formed in step 334 as the core basis, first accurately identifying the natural boundaries, tracking the transition process of the curvature value from the flat interval to the mutation interval or from the mutation interval to the flat interval point by point, the region with continuous change of curvature value and in the middle of the flat interval and the mutation interval is defined as the transition zone, the middle line of the transition zone is the natural boundary of the sub-zone, the boundary demarcation needs to ensure that it conforms to the change law of the curved surface structure, does not split the data points under the same correlation mode, and does not merge the mutation regions caused by different abnormal factors; then, according to the principle of independent partitioning of flat zones and partitioning of mutation zones according to abnormal types, the abstract decision space is divided into several exclusive sub-zones, each sub-zone is clearly labeled with structural characteristics, adaptive construction scenarios, corresponding abnormal factors and boundary ranges, expressed in three-dimensional standardized value intervals; the specific sub-zones are as follows, first, the cost-schedule-quality balance zone, corresponding to the flat region with stable curvature value, the boundary range is that the cost, schedule and quality standardized values are in the middle interval and the curvature value is in the flat interval, adaptive to the normal construction scene without abnormal factors, the structural characteristics are flat curved surface, stable correlation among the three elements, moderate cost increase, synchronous progress improvement and excellent quality; second, the cost-sensitive-schedule-stable zone, corresponding to the structural mutation region caused by fluctuation of building material prices, the boundary range is that the cost standardized value fluctuation interval is wide, the schedule and quality standardized value fluctuation interval is narrow, and the curvature value is in the mutation interval, adaptive to the building material price rise / fall scene, the structural characteristics are that the curved surface fluctuates dramatically in the cost dimension and is flat in the schedule and quality dimensions, the cost fluctuates greatly and the schedule and quality are basically stable; third, the cost-schedule double fluctuation zone, corresponding to the structural mutation region caused by changes in geological conditions, the boundary range is that the cost and schedule standardized value fluctuation intervals are wide, the quality standardized value fluctuation interval is narrow, and the curvature value is in the mutation interval, adaptive to the complex geological scene, the structural characteristics are that the curved surface fluctuates dramatically in the cost and schedule dimensions and is flat in the quality dimension, the cost is over budget, the schedule is lagging behind, and the quality is qualified; fourth, the full-dimensional adjustment zone, corresponding to the structural mutation region caused by design changes, the boundary range is that the cost, schedule and quality standardized values all have certain fluctuations, the curvature value is in the mutation interval, adaptive to the design scheme adjustment scene, the structural characteristics are that the curved surface fluctuates in all dimensions and the correlation among the three elements is re-adapted, the cost is added, the schedule is rearranged, and the quality is adjusted; after the sub-zoning is completed, a written version of the abstract decision space sub-zone diagram is drawn, the boundary coordinates, structural characteristics, adaptive scenarios and abnormal factors of each sub-zone are labeled, and the connection rules between the sub-zones are determined to ensure that there is no omission, no overlap and no mismatch in the sub-zones.

[0048] Step 336, the multi-dimensional vector form original data points of the abstract decision space basis are mapped into the corresponding partition according to the spatial position coordinates of each data point; specifically including: first, extracting the spatial coordinates of all multi-dimensional vector form original data points, i.e. the cost standardization value, the progress standardization value and the quality standardization value corresponding to each data point, registering and recording the associated information and corresponding resource execution record number of each data point; then, checking the coordinates point by point, comparing the three-dimensional standardization value of a single data point with the boundary range of each partition marked on the partition diagram one by one to determine which partition the data point falls into; for example, a certain original data point corresponds to the conventional construction of No. 1 foundation reinforcement binding process, with a cost standardization value of 0.42, a progress standardization value of 0.76 and a quality standardization value of 0.87. According to the partition diagram, the three-dimensional coordinates all fall within the cost-progress-quality balanced zone boundary range, i.e. the data point is mapped to this partition, and the partition name is marked in the account book; a certain original data point corresponds to No. 2 beam concrete pouring process, with a cost standardization value of 0.83, a progress standardization value of 0.72 and a quality standardization value of 0.85, due to the rising building material price leading to cost overrun, the coordinates fall within the cost-sensitive-progress-stable zone range, and the data point is mapped to this partition, and the abnormal factor is marked as building material price fluctuation; for the data points falling on the partition boundary line, the nearest attribution and data density priority principle is adopted, i.e. first judging which partition the boundary point is closer to, and then counting the data point density of the two adjacent partitions, attributing the boundary point to the partition with closer distance and higher data point density, and marking the boundary point attribute and attribution basis in the account book to avoid ambiguous attribution; during the mapping process, one data point is randomly reviewed after every 10 data points are mapped to ensure that the mapping is error-free; after the completion of the mapping, a comprehensive review is carried out: counting the number of data points in each partition, checking whether the associated scene of the data points in each partition is consistent with the partition adaptation scene, correcting the data points with mapping errors one by one, and supplementing the missing associated information; finally, a complete account book of partition name-data point list-associated information-resource execution record number is formed, and each data point in each partition corresponds to a specific construction scene.

[0049] Step 337, analyze all data points mapped within each partition, identify the correlation pattern between cost, schedule and quality reflected by the data points, and extract the unique data distribution statistical features and dominant change law within each partition; Specifically including: First, analyze the cost-schedule-quality balance zone, count the cost standardization value of all data points in this partition, then count the schedule standardization value and quality standardization value, calculate the average value of each dimension (add the value of all data points in the same dimension, divide by the total number of data points in the partition) and the fluctuation range (the maximum value minus the minimum value of the dimension); Through point-by-point analysis of the synchronous change trend of three-dimensional values, identify the correlation pattern as cost moderate growth, schedule synchronous improvement, and quality stable and excellent, that is, by reasonably configuring artificial and mechanical resources, the cost increases slightly according to the initial plan, the schedule advances according to the node or even slightly ahead, and the quality always maintains above the qualified level without abnormal factors interference; The statistical features extracted are that the three-dimensional standardization values are concentrated in the middle interval, the fluctuation range is narrow, the data point clustering degree is high, and there are no obvious discrete points; The dominant change law is that the cost standardization value increases by one small interval (such as 0.01-0.02), the progress standardization value is synchronized to increase in the corresponding microcell interval, the quality standardization value is basically maintained in a fixed range without obvious fluctuations, and the change trend is consistent with the expectation of the initial resource allocation scheme; then the cost-sensitive-progress stable zone is analyzed, the cost dimension fluctuations are analyzed, and the associated mode is identified as a dramatic fluctuation in the cost dimension, a basic stability in the progress and quality dimensions, i.e., the rise in the price of building materials leads to cost overruns, and the cost standardization value rises sharply, or the price drop leads to cost savings, and the cost standardization value drops sharply, but the progress is not significantly affected through optimization of artificial scheduling and improvement of mechanical operation efficiency, and the quality is strictly controlled according to the construction specification without fluctuations; the extracted statistical features are a wide fluctuation range of the cost dimension standardization value, a significant difference between the maximum and minimum values, a narrow fluctuation range of the progress and quality dimensions, and concentrated values; the dominant change law is that the larger the price increase of building materials, the more significant the increase in the cost standardization value, the progress and quality dimension values are always maintained in a fixed interval and are not affected by cost fluctuations, and the cost fluctuations can be offset by subsequent cost savings; then the cost-progress double fluctuation zone is analyzed, the linkage change of cost and progress is analyzed, the details of the geological condition change are combined, and the associated mode is identified as cost increase, progress lag, and quality compliance, i.e., complex geological conditions increase the difficulty of construction, additional labor is needed for supplementary reconnaissance and adjustment of the process, waterproof materials and reinforcing materials are additionally purchased, leading to cost overruns, and the process connection is delayed, causing progress lag, but the quality is strictly controlled according to the new geological condition requirements and is always maintained at a qualified level; the extracted statistical features are that the cost and progress dimension standardization values fluctuate synchronously and change in the opposite direction, and the quality dimension standardization value is stable in the qualified interval; the dominant change law is that the higher the complexity of the geology, the greater the increase in the cost standardization value and the decrease in the progress standardization value, and the quality dimension value is always maintained above the qualified line with a very small fluctuation range; finally, the full-dimension adjustment zone is analyzed, the linkage change of the three dimensions is analyzed, the design change details are combined, and the associated mode is identified as adaptive adjustment of cost, progress, and quality, i.e., after the design change, additional labor and material inputs are needed, leading to cost increases, the progress needs to be re-planned, and the quality is adjusted according to the new design requirements, and all three dimensions change; the extracted statistical features are that the three-dimension standardization values all have some fluctuations, the fluctuation directions are consistent, and the values quickly stabilize after adjustment; the dominant change law is that the larger the design change adjustment range, the larger the adjustment range of the three-dimension standardization values, and after the adjustment is completed, the three elements quickly form a new synergistic relationship, the fluctuation range is reduced, and the system tends to be stable; after the analysis of each zone is completed, a detailed zonal analysis account is formed, the associated mode, statistical features, dominant law, and adaptive scenarios of each zone are determined.

[0050] Step 338, according to the data distribution statistical characteristics and the dominant change law, the strategy correction parameter of targeted adjustment is generated; Specifically including: combining with the analysis results of each partition, the exclusive and landable strategy correction parameter is generated for each partition, the parameter is consistent with the demand of rail transit main structure construction, directly guides the site control adjustment, for the cost-schedule-quality balanced area, based on the law of three element collaborative stability, the resource allocation optimization parameter is generated, for artificial, the final number ratio of each type of work is determined, to avoid artificial redundancy or shortage, at the same time, the upper limit of daily effective operation time is set to prevent fatigue work affecting efficiency; For materials, based on the low loss characteristics of partition statistics, the upper limit of the loss rate of various building materials is slightly reduced on the basis of the initial scheme, and the loss control responsibility is determined; For machinery, the operation time period is optimized to improve the utilization rate of machinery, and the minimum daily operation time of machinery is set to reduce idling; The core goal of the parameter is to maintain the current balanced state, slightly improve the resource utilization efficiency, and control the cost growth; For the cost-sensitive-schedule-stable area, based on the law of cost fluctuation, schedule and quality stability, the cost control correction parameter is generated; For material procurement, the building material price fluctuation warning threshold is set, which is determined according to the partition statistical cost fluctuation range, when the building material price rise reaches the warning threshold, two adjustment measures are immediately started, one is to negotiate with the supplier for batch purchase, and the other is to replace the low-price substitute material of the same specification under the premise of meeting the quality standard, and the quality detection requirements of the substitute material are specified; For cost allocation, the proportion of over-budget cost allocation is determined to avoid excessive pressure on single item cost; In addition, the cost saving target of subsequent process is set to offset part of the over-budget cost; The core goal of the parameter is to control the cost fluctuation range and maintain the stability of schedule and quality; For the cost-schedule double fluctuation area, based on the law of two-dimensional fluctuation influenced by geology, the process and resource adjustment parameters are generated; For process, the process connection buffer time of complex geological area is determined to avoid further delay of schedule, and the process sequence is optimized to reduce interference; For resources, the upper limit of additional investment of labor and materials is set, the starting condition of standby machinery is specified, and the approval process of additional cost is set; In addition, the schedule catching up plan is set, but the quality detection frequency in the catching up process is not reduced; The core goal of the parameter is to alleviate the cost and schedule double fluctuation and ensure the quality to be qualified; For the full-dimensional adjustment area, based on the law of three-dimensional adaptive adjustment, the comprehensive adjustment parameter is generated; For cost, the upper limit of cost addition after design change is determined, the composition of additional cost is refined, and the approval process is determined; For schedule, the node time of each process is re-planned, the upper limit of resource investment for schedule catching up is determined, the buffer period of schedule adjustment is set to avoid excessive catching up affecting quality; For quality, the detection frequency and index threshold under the new design standard are set, and the rectification time limit is determined; The core goal of the parameter is to quickly adapt to design change, promote the three-dimensional cost, schedule and quality to form a new collaborative stable relationship, and avoid imbalance after adjustment;All policy correction parameters are compiled in written form, and corresponding partitions, adaptation scenarios, adjustment basis and responsible positions are marked.

[0051] By surface fitting, scattered discrete data is converted into continuous decision surface, which can completely capture the dynamic correlation trend of cost, schedule and quality, and ensure the overall cost target and schedule of the project.

[0052] In a preferred embodiment of the present application, step 4 is based on policy correction parameters, combined with actual cost, schedule and quality data, to normalize and map multi-dimensional data points to a unit hypersphere; by calculating the shortest arc length between data points to quantify the difference measure, and performing difference analysis and multi-objective collaborative evaluation, the cost control comprehensive evaluation result can include:

[0053] In the embodiment of the present application, step 440, the actual cost, progress and quality data of the engineering cost project in the preset period are obtained, aligned and recombined according to the time sequence to form new multidimensional data points in units of each monitoring time point; specifically including: first, determine the preset period and the monitoring time point, the preset period is set to 7 days, which is in line with the weekly progress review requirement; two core monitoring time points are set every day, which are 10 am and 4 pm respectively, and special time periods such as holidays and bad weather are marked at the same time to ensure that the data covers the whole construction scene; then, data acquisition is carried out synchronously by post, and the previous resource execution record and partition account are associated throughout the process, the actual cost data in the preset period is extracted from the cost flow account, which is split to specific process parts according to the monitoring time point, covering labor compensation, material procurement, mechanical leasing, temporary expenses, each data is marked with the corresponding monitoring time point, process part, resource execution record number, and the invoice and payment voucher are checked at the same time to ensure that the cost data is real and traceable; the actual progress data in the corresponding period is extracted from the progress report and the construction log, the cumulative completion quantity is calculated according to the monitoring time point, the steel binding process is calculated according to the actual binding steel weight, the template installation process is calculated according to the actual laying area, and the concrete pouring process is calculated according to the actual pouring volume, the progress data of each monitoring time point is the completed process quantity before the time point, and the progress influencing factors are marked; the actual quality data in the corresponding period is extracted from the quality detection account, which is arranged to specific detection batches according to the monitoring time point, the material detection covers the detection results of the tensile strength of each batch of steel and the slump of concrete, the process detection covers the detection results of the steel spacing, the template flatness and the concrete density, the quality data of each monitoring time point is the detection batch result of the completed process before the time point, the detection personnel and the names of the standing supervisors are marked to ensure that the quality data can be reviewed; after the data acquisition is completed, the data points are aligned according to the time sequence, the cost data, progress data and quality data of the same monitoring time point and the same process part are matched one by one, and recombined into new three-dimensional multidimensional data points, each data point corresponds to a certain year, month, day and time point-process part, contains three core dimensions, the cumulative actual cost value (the sum of all subcosts) of the time point and the part, the cumulative actual progress completion quantity (the cumulative value of the corresponding process quantity) and the cumulative quality detection qualified pass rate (the number of qualified detection batches divided by the total number of detection batches), and the associated information is marked completely; for example, the steel binding data point of No. 1 foundation pit beam at 10 am on X day in X month of 202X, the cost dimension is the total sum of the steel material cost of the steel binding of the part before the time point, the steel worker's salary and the steel processing mechanical leasing cost, the progress dimension is the total weight of the bound steel of the part, and the quality dimension is the qualified pass rate of the 2 batches of steel binding process detection completed by the part, forming a new multidimensional data point set with monitoring time point as the unit, process part as the carrier and data complete and traceable.

[0054] Step 441, based on the feature partition adjustment coefficient defined in the strategy correction parameter, the new multi-dimensional data point is normalized and all numerical values are converted to a unified scale range to generate a standard multi-dimensional data point; Specifically, first, the strategy correction parameter account generated in step 338 is retrieved, and the dimension adjustment coefficient corresponding to each feature partition is extracted; The coefficient is strictly set to fit the partition data fluctuation characteristics, and the normalized core control dimension is ensured, and the specific coefficient standard is that the cost-schedule-quality balanced three-dimensional adjustment coefficient is 1.0; The cost-sensitive-schedule stable area cost dimension adjustment coefficient is 1.2, and the schedule and quality dimension adjustment coefficients are both 1.0; The cost-schedule double fluctuation area cost and schedule dimension adjustment coefficients are both 1.1, and the quality dimension adjustment coefficient is 1.0; The full dimension adjustment area three-dimensional adjustment coefficient is 1.1; Then, the new multi-dimensional data point is normalized and converted to a scale range of 0-1; The cost dimension normalization first calculates the maximum and minimum values of the cost dimension of all new multi-dimensional data points in the preset period; The maximum value is the maximum cumulative cost of each monitoring time point and each process part in the period, and the minimum value is the minimum value of the corresponding dimension in the period, while excluding extreme abnormal cost values; Then, the preliminary normalization value is calculated by subtracting the cost minimum value from the single data point cost value, dividing the result by the difference between the cost maximum value and the cost minimum value, and multiplying the cost adjustment coefficient of the corresponding feature partition of the data point; If the preliminary normalization value is greater than 1, it is counted as 1; If it is less than 0, it is counted as 0; For example, a data point belongs to the cost-sensitive-schedule stable area, corresponding to the concrete pouring process of the 1st foundation column part, the cost value is 62,000 yuan, the maximum cost of the concrete pouring process in the preset period is 98,000 yuan, and the minimum value is 34,000 yuan; The preliminary calculation process is 62,000 yuan minus 34,000 yuan, which is 28,000 yuan; 28,000 yuan divided by the difference between 98,000 yuan and 34,000 yuan, which is 64,000 yuan, is 0.4375; 0.4375 multiplied by the cost adjustment coefficient 1.2 is 0.525, and the final cost normalization value is 0.525; The progress dimension normalization is the same, first calculate the maximum and minimum values of the progress dimension of all new multi-dimensional data points in the preset period; The maximum value is the maximum cumulative progress completion amount of each monitoring time point and each process part in the period, and the minimum value is the minimum value of the corresponding dimension in the period, while excluding extreme abnormal progress values; Then, the calculation is performed by subtracting the progress minimum value from the single data point progress value, dividing the result by the difference between the progress maximum value and the progress minimum value, and multiplying the progress adjustment coefficient of the corresponding feature partition of the data point; If it exceeds the range of 0-1, it is counted as the boundary value; For example, a data point belongs to the cost-schedule double fluctuation area, corresponding to the 1st foundation bottom reinforcement binding process, the progress value is 280 kg, the maximum progress of the process in the preset period is 520 kg, and the minimum value is 120 kg; The calculation process is 280 kg minus 120 kg, which is 160 kg; 160 kg divided by the difference between 520 kg and 120 kg, which is 400 kg, is 0.4; 0.4 multiplied by the progress adjustment coefficient 1.1 is 0.44, and the progress normalization value is 0.44; the quality dimension normalization is based on the quality detection qualified rate, calculated by multiplying the single data point quality qualified rate by the quality adjustment coefficient of the corresponding feature partition, if the result is greater than 1, take 1, less than 0, take 0; if a data point belongs to the full dimension adjustment area, corresponding to the No. 2 beam template installation process, the quality qualified rate is 92% (2 batches qualified, 1 batch qualified after rectification, qualified rate = 3 ÷ 3 × 100%), the calculation process is 92% multiplied by the quality adjustment coefficient 1.1, 1.012 is obtained, because the result is greater than 1, the final quality normalization value is 1.0; after processing all dimensions, the cost, progress, quality normalization value of each monitoring time point and each process position are combined to generate a standard multi-dimensional data point.

[0055] Step 442, mapping the standard multi-dimensional data points to the high-dimensional feature space through a preset coordinate transformation relationship, so that each standard multi-dimensional data point is projected and located on the surface of a unit hypersphere in the high-dimensional space; specifically including: determining the preset high-dimensional feature space as a three-dimensional expansion space, the unit hypersphere as a three-dimensional sphere with a fixed radius of 1, and the sphere center as the coordinate origin (0, 0, 0); the core principle of the coordinate transformation relationship is to keep the relative proportions of each dimension unchanged, and to adjust the data point coordinates so that the straight-line distance to the sphere center is exactly 1, to ensure that all data points are comparable in the same spherical space; then, the coordinate transformation mapping is performed on the standard multi-dimensional data points one by one, first, the square sum of the three-dimensional normalized value of a single standard data point is calculated, that is, the cost normalized value is multiplied by itself, plus the progress normalized value is multiplied by itself, plus the quality normalized value is multiplied by itself, and the total sum is the square sum of the three-dimensional value; second, the square root of the square sum is calculated, that is, the initial straight-line distance of the data point to the sphere center, for example, if the square sum is 1.5329, the square root is 1.238; third, the normalized value of each dimension is divided by the initial distance to obtain the mapped spherical coordinate value, that is, the mapped cost coordinate = cost normalized value ÷ initial distance, the mapped progress coordinate = progress normalized value ÷ initial distance, and the mapped quality coordinate = quality normalized value ÷ initial distance, and through the transformation, the distance of the data point to the sphere center is ensured to be 1, which falls exactly on the surface of the unit hypersphere; for example, the normalized value of a certain standard data point is (0.525, 0.44, 1.0), the first step is to calculate the square sum, 0.525 is multiplied by itself to get 0.2756, 0.44 is multiplied by itself to get 0.1936, and 1.0 is multiplied by itself to get 1.0, the sum of the three is 0.2756+0.1936+1.0=1.4692; the second step is to calculate the initial distance, that is, the square root of 1.4692 (about 1.212); the third step is to calculate the mapped coordinates: cost coordinate = 0.525 ÷ 1.212 ≈ 0.433, progress coordinate = 0.44 ÷ 1.212 ≈ 0.363, and quality coordinate = 1.0 ÷ 1.212 ≈ 0.825, finally the data point is mapped to the spherical coordinate (0.433, 0.363, 0.825); during the mapping process, after completing the mapping of 10 data points, the coordinate position is marked by a written spherical diagram, and the distance of the data point to the sphere center is measured by a ruler, if the distance deviates from 1, the calculation results of the square sum and the initial distance are rechecked, and the mapping coordinates are corrected; at the same time, in combination with the feature partition attribute in the foregoing, the standard data points in different partitions are marked and distinguished by text, the cost-progress-quality balanced zone is marked as a regular type, the cost-sensitive-progress stable zone is marked as a cost warning type, the cost-progress double fluctuation zone is marked as a double fluctuation type, and the full-dimensional adjustment zone is marked as a full adjustment type, and finally a positioning account of all standard data points on the unit hypersphere is formed.

[0056] Step 443, on the surface of the unit hypersphere, select the reference data point representing the target or reference state for the current data point representing the actual execution state of the current period, calculate the shortest arc length on the hypersphere surface, and take the geometric distance value as the quantitative difference measure of the current state and the target state in the multi-dimensional space; Specifically includes: the current data point is the standard data point at the last monitoring time point of the preset period, and the representative data point of the core process in the period is selected to represent the final state actually executed in the period, and the characteristic partition attribute and abnormal factor influence of the data point are marked; The reference data point is divided into two categories: one is the target reference point, which is generated based on the expected target set by the initial cost control strategy, that is, the spherical coordinate point obtained after the cost upper limit in the initial scheme, the progress node, and the quality standard rate in the period are normalized in step 441 and mapped in step 442, to ensure that the target reference point is consistent with the characteristic partition to which the current data point belongs; The other is the high-quality reference point, which is obtained by selecting high-quality project data under the same type of rail transit station main structure construction, the same process, and the same geological conditions after the same process, as an auxiliary reference basis; Then calculate the shortest arc length of the unit hypersphere, first, determine the coordinate position of the current data point and the reference data point on the sphere, mark the positions of the two points and the center of the sphere on the written spherical diagram, and form a triangle (two points and the center form); Second, calculate the central angle formed by the two points and the center, judge the angle size by the corresponding dimension product of the two point coordinates, multiply the cost coordinates, the progress coordinates, and the quality coordinates of the current data point and the reference data point, and then add the three products, the result is closer to 1, the angle between the two points and the center is smaller, and the result is closer to -1, the angle is larger; Third, calculate the shortest arc length, because the unit hypersphere radius is 1, the value of the shortest arc length is equal to the value of the central angle (when the radius is 1, the arc length = central angle), the arc length is the quantitative difference measure of the current state and the target state, the longer the arc length, the greater the difference between the actual execution and the target / reference state, the shorter the arc length, the smaller the difference; For example, the current data point (0.433, 0.363, 0.825) and the target reference point (0.45, 0.4, 0.8), calculate the coordinate product sum 0.433x0.45+0.363x0.4+0.825x0.8≈0.195+0.145+0.66=1.0, corresponding to the central angle is very small, the shortest arc length is very short, which indicates that the actual execution and the target state difference is small; If the coordinate product sum of the current data point and the target reference point is 0.5, The corresponding central angle is larger, the arc length is longer, and the difference is larger; at the same time, for high-quality reference points, the arc length is calculated according to the same logic, and the difference between the current data point and the target reference point and the high-quality reference point is compared. If the difference with the target reference point is large but the difference with the high-quality reference point is small, it means that the initial target setting may be too strict, and the target needs to be adjusted according to the actual construction. If the difference with both is large, the deviation reason needs to be investigated in detail; After calculation, the arc length value, the size of the central angle, the difference level, and the preliminary analysis results of the difference reason are recorded to the account book, and the selection basis of the target reference point and the high-quality reference point is marked.

[0057] Step 444, dimension analysis is performed on the quantified difference measure to determine the specific deviation direction and deviation magnitude of the current execution effect and the expected target in the cost dimension, the progress dimension, and the quality dimension, to obtain a dimension deviation analysis result; specifically including: decomposing the quantified difference by dimension one by one, combining the data point coordinates on the unit hypersphere, the original data, and the characteristic partition attributes, to determine the deviation direction, the deviation magnitude, and the causes of each dimension, and to analyze the correlation between dimensions, first extract the cost dimension hypersphere coordinates of the current data point and the target reference point, compare their sizes, if the current cost coordinate is greater than the target reference point cost coordinate, the deviation direction is cost overrun; if it is less than the target reference point cost coordinate, the deviation direction is cost saving; the deviation magnitude is calculated by the absolute value of the difference between the current cost coordinate and the target reference point cost coordinate, divided by the target reference point cost coordinate, the larger the result, the greater the deviation magnitude, and the actual overrun / saving amount is calculated by combining the original cost data, and is associated to the specific itemized cost to investigate the deviation causes; for example, the current data point cost coordinate is 0.433, the target reference point cost coordinate is 0.4, the absolute value of the difference is 0.033, the deviation magnitude = 0.033 ÷ 0.4 = 0.0825, the deviation direction is slight overrun, and the original data is verified to be caused by a small increase in the purchase price of steel materials, without other abnormal factors; the progress dimension analysis compares the progress dimension hypersphere coordinates of the current data point and the target reference point, if the current coordinate is greater than the target coordinate, the deviation direction is progress advance; if it is less than the target coordinate, the deviation direction is progress lag; the deviation magnitude is calculated by the absolute value of the difference between the current progress coordinate and the target reference point progress coordinate, divided by the target reference point progress coordinate, and the actual advance / lag engineering quantity and the corresponding time length are calculated by combining the original progress data to analyze the deviation causes; for example, the current data point progress coordinate is 0.363, the target reference point progress coordinate is 0.4, the absolute value of the difference is 0.037, the deviation magnitude = 0.037 ÷ 0.4 = 0.0925, the deviation direction is slight lag, and the original data is verified to be caused by a one-hour suspension of the steel binding process due to slight water seepage, without other interference; the quality dimension analysis compares the quality dimension hypersphere coordinates of the current data point and the target reference point, if the current coordinate is less than the target coordinate, the deviation direction is quality substandard; if it is greater than or equal to the target coordinate, the deviation direction is quality standard / exceeding the target; the deviation magnitude is calculated only when the quality is substandard, by the absolute value of the difference between the target reference point quality coordinate and the current quality coordinate, divided by the target reference point quality coordinate, and the specific indicators of the substandard are determined by combining the quality detection records to analyze the causes; for example, the current data point quality coordinate is 0.825, the target reference point quality coordinate is 0.8, the current coordinate is greater than the target coordinate, the deviation direction is quality better than the target, there is no deviation amplitude, it is verified that the steel bar binding process detection batch is all qualified, and the deviation of the steel bar spacing is controlled within the lower limit of the allowed range, and the control effect is excellent; then integrate the analysis results of each dimension to determine the correlation between dimensions, such as whether there is a correlation between the slight cost overrun and the slight lag of progress, whether the quality better than the target depends on the cost addition, and finally form the dimension deviation analysis result account book.

[0058] Step 445, according to the preset cost, progress, quality three-dimensional collaborative target adjustment coefficient, the dimension deviation amplitude is matched and calculated, the overall collaborative performance under the multi-target constraint of cost, progress and quality is evaluated, and the multi-target collaborative evaluation value is obtained; Specifically including: first, determine the preset collaborative target adjustment coefficient, the total sum of the coefficient is 1.0, and the specific allocation is quality dimension adjustment coefficient 0.4, quality directly related to structure safety, which is the core control target, the weight is the highest, cost dimension adjustment coefficient 0.35, building material price fluctuation has significant influence on cost, which needs to be controlled, progress dimension adjustment coefficient 0.25, progress can be optimized and caught up through subsequent process, the weight is relatively low, then the comprehensive calculation is carried out, first, the dimension deviation amplitude is standardized and corrected, the correction rule is set according to the deviation direction, if the deviation direction is positive, the deviation amplitude is calculated as negative value; If the deviation direction is negative, the deviation amplitude is calculated as positive value; When there is no deviation amplitude, the corrected deviation amplitude is calculated as 0; For example, the cost dimension is slightly over budget, the deviation amplitude is 0.0825, and the corrected deviation amplitude is +0.0825; The progress dimension is slightly delayed, the deviation amplitude is 0.0925, and the corrected deviation amplitude is +0.0925; The quality dimension is better than the target, and there is no deviation amplitude, and the corrected deviation amplitude is 0; Second, the weighted deviation value of each dimension is calculated, that is, the cost weighted deviation value=cost corrected deviation amplitude*0.35, the progress weighted deviation value=progress corrected deviation amplitude*0.25, and the quality weighted deviation value=quality corrected deviation amplitude*0.4, the influence weight of each dimension on the overall collaborative performance is calculated; For example, the cost weighted deviation value=0.0825*0.35≈0.0289, the progress weighted deviation value=0.0925*0.25≈0.0231, and the quality weighted deviation value=0*0.4=0; Third, the multi-target collaborative evaluation value is calculated, the weighted deviation values of the three dimensions are added, and the sum is the collaborative evaluation value; The closer the evaluation value is to 0, the better the overall collaborative performance is, and the closer the multi-target constraint is; The evaluation value is positive and larger, which means that the collaborative performance is worse, and there is a multi-dimensional control imbalance; The evaluation value is negative, which means that the overall performance is better than the target, and the collaborative effect is good; For example, the above data is added 0.0289+0.0231+0=0.052, the collaborative evaluation value is 0.052, which is close to 0, indicating that the overall collaborative performance is good; At the same time, combined with the feature partition attribute in the foregoing, the collaborative evaluation value is calibrated again, the calibration rule is that the cost-progressive-quality balanced region evaluation value*1.0 (no calibration is needed), the cost-sensitive-progressive-stable region evaluation value*1.05 (the cost dimension influence is strengthened), the cost-progressive double fluctuation region evaluation value*1.1 (the cost and progress double dimension influence is strengthened), and the full dimension adjustment region evaluation value*1.15 (the full dimension influence is strengthened, and the adjustment demand is adapted); For example, the current data point belongs to the cost-sensitive-progressive-stable region, and the calibrated collaborative evaluation value=0.052*1.05≈0.0546, After calibration, still close to 0, the synergy performance is maintained well; after calculation, the synergy evaluation value, correction process, weighted calculation details, secondary calibration basis, synergy level are recorded, and the corresponding standard of synergy level is that the evaluation value is less than 0.05, which is excellent, 0.05-0.1 is good, 0.1-0.2 is general, and greater than 0.2 is poor.

[0059] Step 446, the dimensional deviation analysis result and the multi-target synergy evaluation value are integrated and calculated, and the cost control comprehensive evaluation result is finally generated; specifically including: carrying out integrated analysis according to the principle of dimensional decomposition and overall synergy to generate a comprehensive evaluation result that can be implemented; first, the multi-target synergy evaluation value is combined with the average value of the dimensional deviation amplitude, that is, the comprehensive evaluation basis value = multi-target synergy evaluation value + (sum of absolute values of dimensional deviation amplitude ÷ 3), the closer the basis value is to 0, the better the comprehensive performance is; then, the grading judgment is carried out, the judgment standard is drawn according to the construction control requirements of the main structure of rail transit, the basis value is less than 0.05, which is judged as excellent, which means that the actual execution deviates from the target and the multi-dimensional synergy is good; the basis value is between 0.05 and 0.1, which is judged as good, which means that there is slight deviation and the overall synergy is controllable; the basis value is between 0.1 and 0.2, which is judged as general, which means that there is obvious deviation and needs to be adjusted; the basis value is greater than 0.2, which is judged as poor, which means that the deviation is serious and the synergy is unbalanced, which needs to be comprehensively rectified; at the same time, combined with the dimensional deviation analysis result, the specific rectification direction is determined, such as evaluating as good and the deviation is caused by slight cost overrun, which needs to implement the cost control correction parameter; such as evaluating as general and the deviation is caused by progress lag + cost overrun, which needs to optimize the process and resource adjustment parameter; such as evaluating as poor and there is no deviation in the quality dimension, the cost and progress are seriously unbalanced, which needs to re-comb the strategy correction parameter and adjust the synergy target, and finally generate a written cost control comprehensive evaluation result.

[0060] By aligning and reorganizing the data in time sequence, quantifying the difference, and dimensionally decomposing and analyzing, the deviation of the cost, progress, and quality dimensions and the related influence can be directly presented.

[0061] In a preferred embodiment of the present application, the above step 5, based on the cost control comprehensive evaluation result and the strategy correction parameter, the current strategy is optimized by using the reinforcement learning model to generate the optimized cost control strategy, which can include:

[0062] In the embodiment of the present application, step 550, the cost control comprehensive evaluation result is converted into a penalty signal; at the same time, the strategy correction parameter is fused with the current environmental state information of the project to form a comprehensive state feature vector; specifically including: the first part, the cost control comprehensive evaluation result is converted into a penalty signal, first, the cost control comprehensive evaluation result account generated in step 446 is called, the core evaluation index is extracted, the comprehensive judgment level, the comprehensive evaluation basic value, the dimension deviation amplitude, the feature partition attribute, the penalty signal conversion rule is set, the conversion logic is based on the evaluation level to set the benchmark penalty value, the basic value and the deviation amplitude are adjusted, the partition attribute is calibrated, the comprehensive judgment is excellent, the benchmark penalty value is set to 0 (no penalty, representing that the control meets the standard); the benchmark penalty value is set to 1 (slight penalty, representing that it needs to be fine-tuned) when it is good; the benchmark penalty value is set to 3 (moderate penalty, representing that it needs to be optimized) when it is general; the benchmark penalty value is set to 5 (heavy penalty, representing that it needs to be rectified) when it is poor; then adjust according to the comprehensive evaluation basic value, the more the basic value deviates from 0, the more the penalty value is added, the accumulation rule is penalty value = benchmark penalty value + (comprehensive evaluation basic value ÷ corresponding level upper limit value) × 2, for example, a project is judged to be general, the comprehensive evaluation basic value is 0.1129, the calculation is 3 + (0.1129 ÷ 0.2) × 2 = 3 + 1.129 = 4.129, and the preliminary penalty value is obtained; then calibrate according to the feature partition attribute, the cost-schedule-quality balanced area penalty value × 1.0, the cost-sensitive-schedule stable area × 1.05 (strengthening the cost deviation penalty), the cost-schedule double fluctuation area × 1.1 (strengthening the double dimension deviation penalty), the full dimension adjustment area × 1.15 (strengthening the full dimension deviation penalty), the final penalty signal is obtained after calibration, and is marked as a specific numerical value, the conversion process and accounting basis are recorded at the same time; the second part, the strategy correction parameter is fused with the environmental state information to form a comprehensive state feature vector, the strategy correction parameter of step 338 is extracted, classified according to the feature partition, covering resource allocation optimization parameter, cost control correction parameter, process and resource adjustment parameter, comprehensive adjustment parameter, each parameter is marked with corresponding partition and adaptive scene; collect the current environmental state information of the project, specifically including construction process state, resource real-time state, abnormal factor current situation, external environment, progress node pressure; then fuse according to the dimension alignment, numerical normalization and ordered combination principle, the strategy correction parameter and the environmental state information are corresponding to 12 core dimensions, each dimension corresponds to a specific numerical value; for example, the cost control dimension, the price warning threshold normalization value in the strategy correction parameter is taken, multiplied by the current building material price fluctuation amplitude normalization value in the environmental state, to obtain the fusion value of this dimension; the artificial resource dimension, the artificial ratio normalization value in the strategy parameter is taken, added to the normalization value of the ratio of the current on-duty number to the planned number, and then divided by 2 to obtain the fusion value; after the 12 dimension fusion values are calculated one by one, they are combined in a fixed order to form a comprehensive state feature vector.

[0063] Step 551, input the penalty signal and the comprehensive state feature vector into the pre-trained reinforcement learning model, calculate the comprehensive state feature vector through the policy network, and generate the decision probability distribution under the current state; Specifically, first construct the reinforcement learning model in detail, the reinforcement learning model is a three-layer network structure, the number of input layer nodes is set to 12, corresponding to the 12 dimensions of the comprehensive state feature vector, each node is responsible for receiving the fusion value of the corresponding dimension, the node activation rule is to only receive 0-1 range values, and values exceeding the range are corrected according to the boundary value and then transmitted; The hidden layer is set to 2 layers, and the number of nodes in each layer is 24, the node setting is based on 2 times the number of input layer nodes to ensure that the correlation of each dimension can be fully mined; The first layer of hidden layer is responsible for the preliminary integration of input layer data, each node receives the values of all 12 nodes of input layer, calculates the corresponding weight for each input value, the initial weight is set to 0.05-0.1 according to the same type of project data, and the node output value is calculated by adding the bias value to the sum of all results; The output value of the second layer of hidden layer is received by the 24 nodes of the first layer of hidden layer, and the weight and bias value are independently set according to the same logic to ensure deep mining of feature correlation; The number of output layer nodes is set to 8, corresponding to 8 types of core decision actions (adjusting artificial proportion, optimizing material loss rate, prolonging mechanical operation time, starting material negotiation, increasing process buffer time, strengthening quality detection, adding cost investment, and adjusting progress catch-up plan), each node corresponds to a type of decision; The pre-training process of the reinforcement learning model, select the historical data of 3 same type of rail transit station main structure construction projects as training samples, the samples cover different construction stages, abnormal scenes, and evaluation result corresponding strategy adjustment cases, a total of 1000 effective samples are selected; During training, input the sample feature vector and the penalty signal into the reinforcement learning model, and adjust the network weight and bias value through repeated iteration, so that the deviation between the decision output by the reinforcement learning model and the final decision of the sample gradually reduces, and the iteration is stopped when the deviation is within a reasonable range, that is, more than 80% of the sample output decisions are consistent with the final decisions, the pre-training is completed, the pre-trained reinforcement learning model is obtained, and the final initial weight, bias value and network structure parameters are recorded; After the pre-training is completed, the data processor inputs the penalty signal and the comprehensive state feature vector generated in step 550 into the model, the signal and the vector are checked by the input layer nodes and then transmitted to the hidden layer, the two layers of hidden layer are integrated and calculated according to the set logic, and the output layer nodes receive the final output value of the hidden layer; Then calculate the decision probability distribution through the policy network, the policy network is the core calculation unit of the reinforcement learning model, and the probability distribution rule is set based on the output of each node value and the penalty signal; First, calculate the sum of all node values of the output layer, and divide each node value by the sum to get the initial probability; Then adjust the probability combined with the penalty signal, the larger the penalty signal value, the higher the probability of risk-averse decision by proportion, and the lower the probability of aggressive adjustment decision by proportion, the adjustment rule is risk-averse decision probability = initial probability × (1 + penalty signal value × 0.1), the aggressive adjustment type decision probability = initial probability x (1-punishment signal value x 0.1), ensuring that the greater the punishment, the more conservative the decision; for example, the punishment signal value is 4.129, the initial probability of a certain risk-averse type decision is 0.15, and after adjustment, it is 0.15 x (1+4.129 x 0.1) = 0.15 x 1.4129 ≈ 0.2119; the initial probability of a certain aggressive adjustment type decision is 0.2, and after adjustment, it is 0.2 x (1-4.129 x 0.1) = 0.2 x 0.5871 ≈ 0.1174; after all the decision probabilities are adjusted, the total probability is ensured to be 1, forming the decision probability distribution under the current state.

[0064] At step 552, the internal parameters of the policy network are adjusted according to the punishment signal and the preset policy gradient update algorithm to generate an updated policy network parameter set. Specifically, the internal parameters of the policy network are first disassembled, including 12x24 groups of weights from the input layer to the first hidden layer, 24x24 groups of weights from the first hidden layer to the second hidden layer, and 24 bias values for each of the two hidden layers, a total of 12x24+24x24+24+24=960 parameters, each parameter is labeled with the corresponding node association; then the punishment signal is analyzed, the parameter adjustment direction is determined, the adjustment amplitude is calculated, the parameter is updated, and the verification process is operated. First, the punishment signal is analyzed, the punishment signal value and the corresponding evaluation result are analyzed, the core direction of parameter adjustment is determined, and the punishment signal value is greater than 3 (moderate and above punishment). It is indicated that the current decision deviation is large, and it is necessary to adjust to the direction of strengthening cost control, optimizing progress adaptation, and maintaining quality priority. The punishment signal value is 1-3 (slight to moderate punishment), and the direction of adjusting the fine-tuned resource configuration and strengthening the deviation correction is adjusted. The punishment signal value is 0-1 (no punishment to slight punishment), and only the weights and bias values are fine-tuned. Second, the adjustment direction of each parameter is determined. The weights from the input layer to the hidden layer are adjusted in the direction of increasing the corresponding deviation dimension parameters and stabilizing the corresponding advantage dimension parameters. The adjustment direction of the bias value is consistent with the trend of the output value of the corresponding node. The output value is low, and the bias value is increased. The third step is to calculate the adjustment amplitude. The adjustment amplitude is calculated as adjustment amplitude=base adjustment coefficient x punishment signal value x dimension weight. The base adjustment coefficient is uniformly set to 0.01, and the dimension weight is set according to the coordination target adjustment coefficient of step 445. The fourth step is to update the parameters one by one. According to the adjustment direction and amplitude, all 960 parameters are updated, and the initial value, adjustment amplitude, updated value, and calculation basis of each parameter are recorded simultaneously. The fifth step is to verify the updated parameters. The updated parameters are substituted into the reinforcement learning model, the punishment signal and the comprehensive state feature vector of step 550 are input again, the output decision probability distribution is calculated, the distribution difference before and after the update is compared, and if the probability of the corresponding final decision is improved after the update and the punishment signal simulation value is decreased, the parameter update is effective. If it is not effective, the adjustment amplitude is adjusted by 50% in the opposite direction, and the verification is updated again until the requirements are met. After all the parameters are updated and verified, the updated policy network parameter set is formed.

[0065] Step 553, based on the updated policy network, the state features are responded to generate the optimized cost control strategy; specifically including: the comprehensive state feature vector generated in step 550 is input into the updated policy network again, the network re-executes layer-by-layer operation according to the adjusted parameters, and outputs the updated decision probability distribution; then, according to the actual engineering requirements, a decision screening threshold such as 0.1 is set, and inefficient decisions with a probability lower than the threshold are removed, and core decisions with a higher probability are retained, while ensuring that the screened decisions cover the three core dimensions of resource allocation, cost threshold, and progress-quality constraints, avoiding one-sided decisions; the screened decisions are quantitatively converted to form a feasible optimized cost control strategy; for example, if the decision probability of increasing the steel reinforcement procurement reserve is the highest, the material resource adjustment coefficient in the strategy correction parameter and the building material price fluctuation range are combined to determine the procurement quantity adjustment ratio (procurement quantity = original planned procurement quantity × (1 + material resource adjustment coefficient × building material price fluctuation range)), and the procurement time node is determined; if the decision probability of adjusting the labor scheduling method is higher, the labor resource allocation adjustment coefficient and the construction period delay situation are combined to determine the scheduling adjustment scheme (from single shift to two shifts, labor quantity = original planned labor quantity × (1 + labor resource adjustment coefficient × construction period delay ratio)), and the cost warning threshold is simultaneously corrected (new threshold = original threshold × (1 + cost warning threshold correction value)), the progress-quality constraint relationship is clarified, and finally all the quantified decisions are integrated to form an optimized cost control strategy covering resource allocation, cost control, and progress-quality coordination.

[0066] By quantifying the comprehensive evaluation results as a punishment signal and combining multi-dimensional parameter fusion to form a feature vector, the quantization driving of strategy optimization is realized, and the strategy adjustment has a clear data basis.

[0067] In a preferred embodiment of the present application, the above step 6, the optimized cost control strategy is applied to the next construction stage, and the building material price fluctuation, design change and construction period delay information are monitored in real time to realize dynamic optimization control, which can include:

[0068] In the embodiment of the present application, step 660, in the next construction stage, the optimized cost control strategy is executed to guide the completion of resource allocation and cost control according to the optimized cost control strategy; specifically including: in terms of artificial resources, according to the artificial adjustment coefficient and the schedule delay ratio given by the strategy, the actual number of artificial resources to be allocated is calculated, the calculation method is to add the original planned artificial quantity to the original planned artificial quantity multiplied by the artificial adjustment coefficient and multiplied by the schedule delay ratio, at the same time, the scheduling mode is determined, the time of each type of work, the operation cycle and the assessment standard are determined, to ensure that the manpower input matches the progress catching demand and does not cause redundancy; in terms of material resources, combined with the material procurement adjustment ratio in the strategy and the building material price fluctuation range, the procurement quantity of each type of main material is calculated, that is, the original planned procurement quantity plus the original planned procurement quantity multiplied by the material adjustment ratio and multiplied by the building material price fluctuation range, the procurement time node, the supplier selection standard and the material storage loss control requirement are determined at the same time, and new materials involved in design changes are separately accounted for and included in the procurement plan; in terms of mechanical resources, according to the mechanical operation time adjustment coefficient and the engineering quantity change ratio given by the strategy, the actual operation time of the machine is calculated, that is, the original planned operation time plus the original planned operation time multiplied by the mechanical adjustment coefficient and multiplied by the engineering quantity change ratio, the equipment entering sequence, the operation shift and the maintenance cycle are reasonably arranged to avoid idle or insufficient operation capacity of the machine; taking the cost warning threshold of the optimized strategy as the core, the whole process control is carried out, first, the cost control account book is updated, the sub-item engineering cost threshold and the deviation tolerance in the optimized strategy are entered into the system, each cost expenditure needs to be corresponded to a specific construction link, a hierarchical approval process is implemented, and when a single expenditure exceeds 10% of the corresponding link threshold and above, a special audit report needs to be submitted; secondly, a real-time cost deviation monitoring mechanism is established, the deviation value between the actual cost and the optimized budget is calculated daily, the deviation value is the actual cost minus the optimized budget, if the deviation value is positive and exceeds the tolerance, unnecessary expenditure of the corresponding link is immediately suspended, the overexpenditure reason is investigated and countermeasures are taken; if the deviation value is negative, analyze whether the saving reason affects the progress and quality, to ensure that cost saving does not sacrifice the core objectives of the project, at the same time, the linkage influence of progress and quality on cost is tracked synchronously, if the progress is ahead of schedule or the quality standard exceeds the expectation, the cost budget allocation of the subsequent link is adjusted according to the strategy requirements, and the saved amount is reasonably planned to the potential overexpenditure risk link.

[0069] Step 661, in the process of executing the optimized cost control strategy, real-time monitoring of the internal and external environment of the project is synchronized to obtain dynamic information including building material market price, design changes that have occurred, and actual construction progress deviation; Specifically including: Establishing a hierarchical monitoring mechanism at the hour level + day level + week level, the hour level monitoring focuses on key building material prices, through building material market real-time pricing platforms, supplier synchronous feedback and other channels, collecting current transaction prices of various building materials, synchronously recording price changes, change amplitude and influencing factors; Daily monitoring covers construction progress and site environment, daily statistics of actual completed quantities at the current stage before work is completed, comparing the progress plan in the optimized strategy, calculating the progress deviation; At the same time, record the changes of construction site conditions, weather influence, and form daily monitoring log; Week level monitoring is aimed at design changes and policy adjustments, collecting change notification sheets issued by the design unit, sorting out the change content, the involved quantities and the influence on construction technology; Synchronously track local environmental protection and safety policy changes, evaluate the potential impact on construction process and cost; After the monitoring data is collected, it is standardized and quantified to ensure that the data can be directly used for subsequent analysis, the fluctuation of building material price is quantified as the difference between the current price and the budget price in the optimized strategy, and then divided by the optimized budget price to get the price fluctuation amplitude; The design change is quantified as the change part of the quantity divided by the total quantity of the current stage to get the change influence proportion, and the change type and the corresponding cost influence direction are marked; The progress deviation is quantified as the actual completed quantity minus the optimized plan quantity, and then divided by the optimized plan quantity to get the progress deviation proportion; Weather, site and other environmental factors are quantified as influence coefficients, the influence coefficient is the actual influence time divided by the planned operation time of the day, and the coefficient is 0 when there is no influence. All quantified data are arranged in the fixed order of building material price-design change-progress deviation-environmental influence to form a dynamic information account.

[0070] Step 662, convert the dynamic information into a feature vector representing the latest project environment state, and input it into the reinforcement learning model, combined with the newly generated penalty signal according to the latest project environment state, to drive the reinforcement learning model to optimize and update the policy network parameters in a new round; Specifically, first, the dynamic information quantized in step 661 is standardized to eliminate dimensional differences, the processing method is to divide each dynamic information quantization value by the maximum value of the same type of information in the same type of engineering project, and all data are mapped to the [0, 1] interval, then set the weight for each dynamic information item, the weight is determined according to the influence degree on cost control, among them, the building material price fluctuation range and the design change influence proportion weight are the highest, the progress deviation proportion is the second, the weather and site influence coefficient weight is the lowest, and the sum of all weights is 1, then multiply each standardized dynamic information item by the corresponding weight, and then arrange all the product results and the remaining values of the strategy correction parameters involved in step 550 in a fixed order to form a comprehensive state feature vector representing the latest environment state. Each element of the vector corresponds to a quantized environment or strategy parameter, fully reflecting the latest project dynamics; Refer to the penalty signal conversion logic in step 550, generate a new penalty signal combined with the cost control effect under the latest environment state, first, the cost, progress, and quality data in the current stage are comprehensively evaluated to determine the evaluation level and the corresponding basic penalty coefficient; Then calculate the dimension weight, that is, the single dimension deviation amplitude divided by the sum of the three dimension deviation amplitudes; Finally, multiply the basic penalty coefficient of each level by the corresponding dimension weight, and add all the dimension calculation results to get the latest penalty signal, the signal value directly reflects the adaptation degree of the strategy under the current environment state; The latest comprehensive state feature vector and the latest penalty signal are input into the reinforcement learning network at the same time, and the parameter adjustment direction is determined according to the penalty signal. When the penalty signal is positive, the adjustment direction is to reduce the parameter influence that leads to deviation; When the penalty signal is negative, the adjustment direction is to strengthen the current parameter influence, then calculate the parameter adjustment amount, which is the latest penalty signal multiplied by the preset learning rate, and then multiplied by the partial derivative of the corresponding parameter of the policy network, wherein the learning rate remains the same as step 552 to ensure the stability of parameter adjustment, adjust the connection weights and bias items of the policy network one by one according to the adjustment amount, while following the engineering scene constraints, if the adjusted parameter exceeds the reasonable range, take the boundary value as the final parameter, after all parameter adjustments are completed, organize a new set of updated policy network parameters to ensure that the network operation logic is adapted to the latest environment state.

[0071] Step 663, based on the updated policy network, a cost control strategy suitable for the new environment state is generated, and the newly generated cost control strategy is applied to the current construction stage to realize dynamic optimization control; Specifically, the latest comprehensive state feature vector obtained in step 662 is input into the updated policy network, and layer-by-layer operation is performed according to the adjusted parameters, that is, each layer node receives the input value, multiplies it with the corresponding connection weight, adds the bias term, and then passes it to the next layer after being converted by the activation function. Finally, the new decision probability distribution is output; Then set the decision screening threshold to be consistent with step 553, eliminate inefficient decisions with a probability lower than the threshold, retain the core decisions covering resource allocation, cost threshold, and progress-quality constraints, and quantitatively convert the core decisions to generate a new strategy that can be implemented, such as adjusting the material procurement strategy to increase the reserve amount + lock in long-term suppliers in response to the continuous rise in building material prices. The reserve amount is calculated as the current remaining engineering quantity multiplied by (1+latest price fluctuation amplitude); In response to the increase in engineering quantity caused by design changes, adjust the proportion of labor and machinery input, the increase in the number of workers is the original number multiplied by the design change influence ratio, and the increase in the length of time of mechanical operation is the original length of time multiplied by the design change influence ratio; The cost warning threshold is corrected simultaneously, and the new threshold is the original threshold multiplied by (1+latest deviation ratio), and the quality bottom line in the progress catching process is determined. The newly generated cost control strategy is quickly synchronized to the field management team and the cost management team, replacing the original optimization strategy to be applied to the current construction stage, guiding subsequent resource allocation, cost expenditure, progress control and other work; At the same time, the real-time monitoring mechanism of step 661 remains unchanged, and dynamic information after the execution of the new strategy is continuously collected, and the process of steps 662-663 is repeated to realize the closed-loop control of strategy execution-dynamic monitoring-parameter updating-strategy iteration. In response to major uncertain factors in the construction process, the monitoring period and parameter updating interval can be temporarily shortened to ensure that the strategy can quickly respond to extreme scenarios and avoid problems such as cost out of control and progress lag.

[0072] Through the closed-loop process of real-time monitoring-quick iteration, the cost control strategy can dynamically adapt to uncertain factors such as building material price fluctuations and design changes, ensuring that the strategy always matches the current construction environment.

[0073] As shown in Figure 2 , the embodiment of the application also provides a reinforcement learning-based cost control strategy optimization system for an engineering cost project, comprising:

[0074] The construction module is configured to obtain an initial budget of the engineering project, a real-time construction environment state, and a past cost control record, and construct an initial cost control strategy.

[0075] The execution module is configured to execute resource allocation and cost control in the current construction stage according to the initial cost control strategy, and obtain actual cost, progress, and quality data.

[0076] The analysis module is configured to integrate the actual cost, progress and quality data, build a multi-dimensional data distribution model, form an abstract decision space, fit discrete data into a continuous decision surface, calculate the curvature of a local area, analyze the structural change characteristics of the decision surface, divide the abstract decision space based on the structural change characteristics, map the original data to the corresponding partitions, extract the intrinsic characteristics of each partition, and generate strategy correction parameters.

[0077] The evaluation module is configured to normalize and map the multi-dimensional data points to a unit hypersphere based on the strategy correction parameters and in combination with the actual cost, progress and quality data, calculate the shortest arc length to quantify the difference between the data points, perform difference analysis and multi-objective collaborative evaluation, and obtain a cost control comprehensive evaluation result.

[0078] The optimization module is configured to optimize the current strategy based on the cost control comprehensive evaluation result and the strategy correction parameters, and generate an optimized cost control strategy by using a reinforcement learning model.

[0079] The feedback module is configured to apply the optimized cost control strategy to the next construction stage, and monitor the building material price fluctuations, design changes and schedule delay information in real time, and realize dynamic optimization control.

[0080] It should be noted that the system corresponds to the above method, and all implementation manners in the above method embodiment are applicable to this embodiment and can achieve the same technical effects.

[0081] The above is the preferred embodiment of the present application. It should be noted that for ordinary skilled persons in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which should also be considered within the scope of protection of the present application.

Claims

1. A method for optimizing the cost control strategy of an engineering cost project based on reinforcement learning, characterized in that, The method comprises: Step 1, obtaining the initial budget of the engineering project, the real-time construction environment state and the past cost control record, and constructing an initial cost control strategy; Step 2, according to the initial cost control strategy, performing resource allocation and cost control in the current construction stage to obtain actual cost, progress and quality data; Step 3, integrating the actual cost, progress and quality data to construct a multi-dimensional data distribution model and form an abstract decision space; fitting discrete data into a continuous decision surface, calculating the curvature of the local area, and analyzing the structural change characteristics of the decision surface; based on the structural change characteristics, the original data is mapped to the corresponding partition, the internal characteristics of each partition are extracted, and the strategy correction parameters are generated; Step 4, based on the strategy correction parameters, combining the actual cost, progress and quality data, normalizing and mapping the multi-dimensional data points to a unit hypersphere; by calculating the shortest arc length of the data points to quantify the difference measure, and performing difference analysis and multi-objective collaborative evaluation, the cost control comprehensive evaluation result is obtained; including: obtaining the actual cost, progress and quality data of the engineering cost project within the preset period, aligning and recombining according to the time sequence, forming new multi-dimensional data points with each monitoring time point as a unit; based on the feature partition adjustment coefficient defined in the strategy correction parameter, the new multi-dimensional data points are normalized and all numerical values are converted to a unified scale range to generate standard multi-dimensional data points; the standard multi-dimensional data points are mapped to a high-dimensional feature space through a predetermined coordinate transformation relationship, so that each standard multi-dimensional data point is projected and located on the surface of a unit hypersphere in the high-dimensional space; on the surface of the unit hypersphere, for the current data point representing the actual execution state of the current period, select the reference data point representing the target or reference state, calculate the shortest arc length on the surface of the hypersphere, and take the geometric distance value as the quantitative difference measure of the current state and the target state in the multi-dimensional space; dimension analysis is performed on the quantitative difference measure to determine the specific deviation direction and deviation amplitude of the current execution effect and the expected target in the cost dimension, the progress dimension and the quality dimension, to obtain the dimension deviation analysis result; according to the preset collaborative target adjustment coefficient of cost, progress and quality, the dimension deviation amplitude is proportionally calculated to evaluate the overall collaborative performance under the multi-objective constraint of cost, progress and quality, and the multi-objective collaborative evaluation value is obtained; the dimension deviation analysis result and the multi-objective collaborative evaluation value are integrated and calculated and judged to finally generate the cost control comprehensive evaluation result; Step 5, based on the cost control comprehensive evaluation result and the strategy correction parameter, using a reinforcement learning model to optimize the current strategy to generate an optimized cost control strategy; Step 6, applying the optimized cost control strategy to the next construction stage, and real-time monitoring of building material price fluctuation, design change and delay information to realize dynamic optimization control.

2. The method of claim 1, wherein, According to the initial cost control strategy, resource allocation and cost control are performed in the current construction stage to obtain actual cost, progress and quality data, including: Based on the initial cost control strategy, the specific types, quantities and time requirements of artificial, material and mechanical resources in the current construction stage are obtained, and an initial resource allocation scheme is formed; According to the initial resource allocation scheme, the on-site resource scheduling and cost expenditure control are performed, and a resource execution record including the actual resource consumption and the corresponding cost expenditure is generated; Based on the resource execution record, the actual cost flow data, engineering progress completion data and construction quality key indicator detection data occurring in the construction process are collected in real time.

3. The method of claim 2, wherein, Integrate the actual cost, progress and quality data to construct a multi-dimensional data distribution model and form an abstract decision space, including: Standardize the actual cost data, progress data and quality data, and integrate the standardized three-dimensional data into a unified format multi-dimensional vector, where each dimension corresponds to a quantitative indicator of cost, progress and quality; Map the multi-dimensional vector to the preset abstract decision space to form a data point set representing the project state, and analyze the distribution of the data point set in the decision space to construct a multi-dimensional data distribution structure reflecting the dynamic correlation between cost, progress and quality; According to the multi-dimensional data distribution structure, the statistical distribution characteristics of the data point set are used to finally generate an abstract decision space defined by the multi-dimensional data distribution structure.

4. The method of claim 3, wherein, The abstract decision space is used to represent the feasible decision domain of the project cost, progress and quality under certain correlation constraints.

5. The method of claim 4, wherein, Fit discrete data into a continuous decision surface, calculate the curvature of the local area, and analyze the structural change characteristics of the decision surface; based on the structural change characteristics, divide the abstract decision space into regions, map the original data to the corresponding partitions, extract the characteristics in each partition, and generate strategy correction parameters, including: Based on the discrete data points of cost, progress and quality indicators distributed in the abstract decision space, use surface fitting algorithm to fit the discrete data points to form a continuous decision surface; Calculate the geometric curvature of each local position on the surface of the decision surface, and identify the structural mutation area and gentle area of the curvature according to the numerical distribution of the geometric curvature; According to the natural boundary between the structural mutation area and the gentle area, the abstract decision space is divided into several partitions with different structural characteristics; Map the original data points in the form of multi-dimensional vectors of the abstract decision space to the corresponding partitions according to the spatial position coordinates of each data point; Analyze all data points mapped in each partition, identify the correlation pattern between cost, progress and quality reflected by the data points, and extract the unique data distribution statistical characteristics and dominant change law in each partition; Generate strategy correction parameters for targeted adjustment according to the data distribution statistical characteristics and dominant change law.

6. The method of claim 5, wherein the method further comprises: Based on the cost control comprehensive evaluation result and the strategy correction parameter, the current strategy is optimized using a reinforcement learning model to generate an optimized cost control strategy, including: Convert the cost control comprehensive evaluation result into a punishment signal; at the same time, fuse the strategy correction parameter and the current environmental state information of the project to form a comprehensive state feature vector; The penalty signal and the comprehensive state feature vector are input into a pre-trained reinforcement learning model, the comprehensive state feature vector is calculated through a policy network, and a decision probability distribution under the current state is generated; According to the penalty signal and the preset policy gradient update algorithm, the internal parameters of the policy network are adjusted to generate an updated policy network parameter set; Based on the updated policy network, the state features are responded to, and an optimized cost control strategy is generated.

7. The method of claim 6, wherein the method further comprises: The optimized cost control strategy is applied to the next construction stage, and real-time monitoring of building material price fluctuations, design changes and schedule delay information is realized, including: In the next construction stage, the optimized cost control strategy is executed, and the optimized cost control strategy is used to guide the completion of resource allocation and cost control; In the process of executing the optimized cost control strategy, the internal and external environment of the project is monitored in real time, and dynamic information including building material market price, design change and actual construction progress deviation is obtained; The dynamic information is converted into a feature vector representing the latest project environment state, and input into the reinforcement learning model, combined with the penalty signal regenerated according to the latest project environment state, to drive the reinforcement learning model to optimize and update the policy network parameters for a new round; Based on the updated policy network, a cost control strategy suitable for the new environment state is generated, and the newly generated cost control strategy is applied to the current construction stage to realize dynamic optimization control.

8. A reinforcement learning-based project cost control strategy optimization system, the system implements the method of any one of claims 1 to 7, characterized in that, Including: The construction module is used to obtain the initial budget of the engineering project, the real-time construction environment state and the past cost control record, and to construct an initial cost control strategy; The execution module is used to execute resource allocation and cost control in the current construction stage according to the initial cost control strategy, and to obtain actual cost, progress and quality data; The analysis module is used to integrate the actual cost, progress and quality data, construct a multi-dimensional data distribution model, form an abstract decision space, fit discrete data into continuous decision surface, calculate the curvature of local area, and analyze the structural change characteristics of decision surface; Based on the structural change characteristics, the region is divided in the abstract decision space, the original data is mapped to the corresponding partition, the characteristics in each partition are extracted, and the strategy correction parameters are generated; The evaluation module is used to normalize and map the multi-dimensional data points to the unit hypersphere based on the strategy correction parameters and the actual cost, progress and quality data. The difference measure is quantified by calculating the shortest arc length between data points, and difference analysis and multi-objective collaborative evaluation are performed to obtain the cost control comprehensive evaluation result; including: obtaining the actual cost, progress and quality data of the engineering cost project within the preset period, and aligning and recombining according to the time sequence to form new multidimensional data points with each monitoring time point as a unit; based on the feature partition adjustment coefficient defined in the strategy correction parameter, the new multidimensional data points are normalized and all numerical values are converted to a unified scale range to generate standard multidimensional data points; the standard multidimensional data points are mapped to a high-dimensional feature space through a preset coordinate transformation relationship, so that each standard multidimensional data point is projected and located on the surface of a unit hypersphere in the high-dimensional space; on the surface of the unit hypersphere, for the current data point representing the actual execution state of the current period, select the reference data point representing the target or reference state, calculate the shortest arc length on the surface of the hypersphere, and take the geometric distance value as the quantitative difference measure of the current state and the target state in the multidimensional space; the quantitative difference measure is analyzed in dimension to determine the specific deviation direction and deviation amplitude of the current execution effect and the expected target in the cost dimension, the progress dimension and the quality dimension three main dimensions, to obtain the dimension deviation analysis result; according to the preset collaborative target adjustment coefficient of cost, progress and quality, the dimension deviation amplitude is proportionally calculated to evaluate the overall collaborative performance under the multi-objective constraint of cost, progress and quality, and the multi-objective collaborative evaluation value is obtained; the dimension deviation analysis result and the multi-objective collaborative evaluation value are integrated and calculated and judged to finally generate the cost control comprehensive evaluation result; The optimization module is used to optimize the current strategy based on the cost control comprehensive evaluation result and the strategy correction parameter, and to generate an optimized cost control strategy by using a reinforcement learning model; The feedback module is used to apply the optimized cost control strategy to the next construction stage and to monitor the building material price fluctuation, design change and construction period delay information in real time, so as to realize dynamic optimization control.

Images