Engineering cost control method and system based on big data

Abstract

The application relates to the technical field of engineering cost control, and discloses an engineering cost control method and system based on big data. The method comprises the following steps: collecting whole-cycle cost data flow of an engineering project, generating a standardized cost data set through data cleaning; constructing a dynamic cost feature library, extracting multi-dimensional features such as time series fluctuation, resource allocation dispersion and supplier correlation; inputting the feature library into a cost anomaly detection model constructed by a pre-trained deep learning network, outputting a cost deviation index, and triggering a correction instruction if the cost deviation index exceeds a threshold; matching historical cases to generate an optimization strategy set containing material replacement, construction period adjustment and supplier replacement; and virtually deducing the optimization strategies through an engineering digital twin system, and selecting a strategy with a predicted cost curve closest to a target value as a final execution scheme. The method relies on big data and deep learning technology, realizes dynamic management and control of the whole cycle of engineering cost, and improves the accuracy and feasibility of cost control.

Description

Technical Field

[0001] This invention relates to the field of engineering cost control technology, specifically to a method and system for engineering cost control based on big data. Background Technology

[0002] Cost control in engineering projects spans the entire lifecycle, from design and construction to operation and maintenance, and is a core element in ensuring the economic benefits and smooth progress of projects. As engineering projects expand in scale and become more technologically complex, cost data at each stage exhibits characteristics of complexity, dispersed sources, and dynamic changes, gradually revealing many limitations of traditional cost control methods.

[0003] Traditional cost control relies heavily on manual statistics and experience-based judgment. Data collection is often one-sided, focusing only on material consumption data during the construction phase while neglecting the integration of budget parameters from the design phase and loss data from the operation and maintenance phase. This results in a lack of a full-cycle perspective in cost analysis. Furthermore, manual data processing is prone to omissions and errors, and it is difficult to effectively remove outliers, leading to insufficient accuracy and standardization of cost data, directly impacting the reliability of subsequent cost analysis.

[0004] In terms of cost feature extraction, existing technologies typically focus only on single-dimensional cost numerical changes, failing to fully explore multi-dimensional features such as time-series fluctuation patterns, resource allocation rationality, and supplier relationships. This results in an incomplete characterization of cost features, failing to accurately reflect the underlying logic and influencing factors of cost changes. Consequently, subsequent anomaly detection models struggle to obtain effective data support and cannot promptly identify potential cost deviation risks.

[0005] In the anomaly detection and strategy generation stages, traditional methods often employ simple statistical models or fixed threshold judgments, which are poorly adapted to complex and ever-changing engineering cost data. This frequently leads to delayed or false anomaly detection, resulting in increased cost overruns. Furthermore, the formulation of cost optimization strategies lacks effective linkage with historical engineering cases, resulting in weak strategy targeting and a lack of effective verification mechanisms. Blindly implementing strategies may lead to secondary cost risks, further exacerbating the difficulty of cost control. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for engineering cost control based on big data, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a big data-based engineering cost control method, the method comprising:

[0008] Collect the full lifecycle cost data stream of the project, including budget parameters in the design phase, material consumption parameters in the construction phase, and loss parameters in the operation and maintenance phase. After removing outliers through a data cleaning engine, a standardized cost dataset is generated.

[0009] A dynamic cost feature library is constructed, and multi-dimensional features are extracted from the standardized cost dataset. The extracted features include time series fluctuation features, resource allocation discrete features, and supplier correlation features.

[0010] Establish a cost anomaly detection model, input a dynamic cost feature library into a pre-trained deep learning network, output a cost deviation index, and trigger a cost correction instruction when the deviation index exceeds a preset threshold;

[0011] Generate a set of cost optimization strategies, match similar scenarios in the historical engineering case library based on cost correction instructions, and call the strategy generator to output optimization strategies including material replacement schemes, schedule adjustment schemes, and supplier replacement schemes;

[0012] The execution strategy verification process involves inputting the optimized strategy into the engineering digital twin system for virtual simulation, outputting the predicted cost curve after strategy execution, and selecting the strategy whose predicted cost curve is closest to the target value as the final execution plan.

[0013] Preferably, the specific steps for constructing the dynamic cost feature library include:

[0014] The standardized cost dataset is sliced ​​along the time dimension to divide it into time-series data blocks arranged according to construction progress nodes;

[0015] The feature decomposition algorithm is used to separate the resource consumption rate feature and the capital flow cycle feature from the time series data block;

[0016] Establish a supplier relationship graph and extract the supply overlap and price fluctuation synergy characteristics among various material suppliers;

[0017] The time-series features, resource features, and supplier features are integrated into a dynamic cost feature vector set according to their weight ratios.

[0018] Preferably, the specific steps for establishing the cost anomaly detection model include:

[0019] Configure the structural parameters of the deep neural network so that the number of nodes in the input layer is consistent with the dimension of the dynamic cost feature vector;

[0020] Historical engineering anomaly data is loaded as training samples, and the network weight parameters are adjusted through the backpropagation algorithm.

[0021] A cost deviation index calculation unit is set in the output layer. This index is composed of the product of the Euclidean distance between the actual cost and the budgeted cost and the characteristic fluctuation amplitude.

[0022] When the dynamic cost feature vector input in real time triggers the generation of the deviation index, the anomaly type classification label is output synchronously.

[0023] Preferably, the specific steps for generating the cost optimization strategy set include:

[0024] Establish a three-dimensional search index in the historical engineering case library, with index dimensions including cost deviation range, engineering type, and combination of abnormal features;

[0025] The nearest neighbor search algorithm is used to locate the top N historical cases that are most similar to the current abnormal scenario;

[0026] Analyze the corrective measures and their effects data from historical cases to generate a list of strategy options with effect evaluation weights;

[0027] The strategy combiner combines individual strategies from the dimensions of materials, schedule, and suppliers to generate composite optimization strategies.

[0028] Preferably, the specific steps of the execution strategy verification process include:

[0029] Import the current project's BIM model and resource scheduling plan into the engineering digital twin system;

[0030] The composite optimization strategy is decomposed into executable material procurement instructions, construction schedule instructions, and supplier switching instructions;

[0031] Run multiple rounds of Monte Carlo simulations to calculate the cost forecast distribution under different market volatility scenarios;

[0032] The strategy with the median cost distribution closest to the budget target and the smallest variance was selected as the validation scheme.

[0033] Preferably, the method further includes a real-time feedback adjustment step:

[0034] During the strategy execution phase, deploy IoT data acquisition terminals to capture real-time data on actual material consumption and construction progress deviations.

[0035] The real-time data is dynamically compared with the predicted cost curve to generate an execution deviation coefficient.

[0036] When the deviation coefficient exceeds the fault tolerance threshold, the strategy re-optimization process is triggered and the digital twin verification is re-executed.

[0037] Preferably, the specific steps of the trigger strategy re-optimization process include:

[0038] Analyze the main sources of execution deviations and identify the dominant factors in sudden changes in material prices, decreased construction efficiency, or abnormal supplier performance.

[0039] Add search criteria to the historical case database based on the type of dominant factor to narrow down the scope of similar cases;

[0040] Extract emergency strategies from new cases and assess their suitability for the current engineering scenario;

[0041] Generate an incremental optimization strategy package that includes risk hedging measures.

[0042] Preferably, the method further includes a cost knowledge graph update step:

[0043] The cost data stream, abnormal events, and execution strategies for the entire project lifecycle are packaged into a case data package.

[0044] Extract feature vectors and policy performance metrics from case data packages and transform them into new nodes and edge relationships in the knowledge graph;

[0045] The node vector representation is updated by graph embedding algorithm to optimize the semantic matching accuracy of subsequent case retrieval.

[0046] Preferably, the method further includes a multi-project collaborative optimization step:

[0047] Establish cost correlation analysis channels across engineering projects in the knowledge graph;

[0048] Identify the cost linkage effects caused by sharing suppliers or the same construction techniques across different projects;

[0049] When a new project triggers cost anomalies, the historical strategies of related projects are retrieved simultaneously, and collaborative optimization suggestions are generated.

[0050] Preferably, the present invention also includes an engineering cost control system based on big data, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the engineering cost control method based on big data as described above.

[0051] Compared with the prior art, the beneficial effects of the present invention are:

[0052] By collecting cost data streams throughout the entire lifecycle of engineering projects, comprehensively covering budget parameters in the design phase, material consumption parameters in the construction phase, and loss parameters in the operation and maintenance phase, and combining this with a data cleaning engine to remove outliers, a standardized cost dataset is generated. This achieves full-dimensional integration and precise standardization of cost data, breaking the limitations of traditional methods that only focus on data from a single phase. It allows cost control to be based on complete and reliable data sources, ensuring the comprehensiveness and accuracy of cost analysis from the source.

[0053] The construction of the dynamic cost feature library comprehensively captures the inherent laws and influencing factors of cost data by extracting multi-dimensional information such as time series fluctuation features, resource allocation discrete features, and supplier association features. It breaks through the shortcomings of traditional single-dimensional feature extraction, making the characterization of cost features more in line with the actual change logic of engineering costs. It provides rich and valuable feature support for subsequent cost anomaly detection, enabling anomaly detection to be carried out based on more comprehensive feature dimensions and more accurately identify potential cost risk points.

[0054] The cost anomaly detection model based on a pre-trained deep learning network leverages the powerful processing and analysis capabilities of deep learning algorithms for complex data to quickly calculate the cost deviation index and promptly detect anomalies in cost control. When the deviation index exceeds a preset threshold, a cost correction command is immediately triggered. Compared to traditional statistical methods or simple machine learning models, this significantly improves the sensitivity and accuracy of anomaly detection, allowing for timely intervention in the early stages of cost deviation to prevent further escalation of cost overruns and achieve proactive cost risk management.

[0055] In the cost optimization strategy generation stage, matching similar scenarios from a historical engineering case library ensures a high degree of adaptability between the optimization strategies and actual cost anomalies. The strategy generator outputs material replacement solutions, schedule adjustment solutions, and supplier replacement solutions, forming a diversified set of optimization strategies. This can meet the needs of resolving different types and degrees of cost anomalies, providing engineering managers with comprehensive and targeted decision-making directions and avoiding the blindness of traditional strategy formulation.

[0056] During the strategy verification process, the application of the engineering digital twin system enabled virtual simulation of the optimization strategy, accurately predicting the cost change trend after the strategy's actual implementation. By selecting the strategy whose predicted cost curve is closest to the target value as the final implementation plan, the risk of cost runaway that may occur after strategy implementation was effectively avoided, ensuring the smooth achievement of cost control objectives.

[0057] This methodology forms a complete closed-loop management system, encompassing data acquisition, feature extraction, anomaly detection, strategy generation, and verification execution. It deeply integrates big data technology with deep learning algorithms and digital twin technology, enabling dynamic and intelligent management of engineering costs. This method is adaptable to engineering projects of different scales and types, playing an effective role in cost control throughout the design, construction, and operation and maintenance phases. It helps engineering projects achieve reasonable cost allocation and precise control throughout their entire lifecycle, ensuring economic benefits and project efficiency, and driving the field of engineering cost control towards greater efficiency and precision. Attached Figure Description

[0058] Figure 1 A full-cycle cost control analysis diagram;

[0059] Figure 2 A flowchart for constructing a dynamic cost feature library;

[0060] Figure 3 A flowchart for generating a set of cost optimization strategies;

[0061] Figure 4 This is a diagram illustrating the effects of real-time monitoring and strategy optimization. Detailed Implementation

[0062] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] Please see Figure 1 This invention provides a big data-based engineering cost control method, which includes: achieving full-cycle cost management by integrating data acquisition, feature analysis, model detection, strategy generation, and virtual verification. The method uses budget parameters from the design phase, material consumption parameters from the construction phase, and loss parameters from the operation and maintenance phase of the engineering project as input sources. The data stream is processed by a data cleaning engine to remove outliers, forming a standardized cost dataset. A dynamic cost feature library extracts multi-dimensional features from the standardized cost dataset, covering time-series fluctuation features, resource allocation discrete features, and supplier correlation features. A cost anomaly detection model analyzes the dynamic cost feature library based on a pre-trained deep learning network, outputting a cost deviation index. When the index exceeds a preset threshold, a cost correction instruction is automatically triggered. A cost optimization strategy set matches similar scenarios in a historical engineering case library according to the instructions, and a strategy generator outputs material replacement schemes, schedule adjustment schemes, and supplier replacement schemes. An engineering digital twin system virtually simulates the optimization strategies, generates predicted cost curves, and selects the strategy closest to the target value as the final execution plan.

[0064] Example 1: See Figure 2Building a dynamic cost feature library is a crucial step in engineering cost control methods. This library directly processes the standardized cost dataset generated by the data cleaning engine. The standardized cost dataset contains cost records with a unified format and dimensions throughout the entire project lifecycle. These records cover budget parameters during the design phase, material consumption parameters during construction, and loss parameters during operation and maintenance. The construction process begins with time-division slicing of the standardized cost dataset. Time-division slicing is based on predefined construction progress nodes as boundaries. These nodes originate from the project's overall schedule Gantt chart, marking the completion points of key processes such as earthwork excavation, main structure capping, or electromechanical installation completion. The slicing operation divides the continuous standardized cost data stream into a series of time-series data blocks. Each time-series data block contains all cost data entries within a complete construction node cycle. The time-series data blocks maintain temporal continuity, and the data within each block has business logic integrity. A feature decomposition algorithm is used to separate resource consumption rate features and cash flow cycle features from the time-series data blocks. This algorithm employs mode decomposition technology based on signal processing principles. Resource consumption rate characteristics depict the change in the usage of a specific resource per unit time. The calculation process involves taking the first derivative of the material consumption quantity sequence in the time-series data block to obtain the instantaneous rate curve of consumption change over time. Cash flow cycle characteristics focus on the payment rhythm and cash outflow patterns of costs. The analysis objects are accounts payable and actual payment records in the time-series data block. Periodic parameters are extracted by identifying the time interval patterns and amount distribution patterns of payment events. Feature decomposition algorithms can separate the long-term trend components, seasonal cyclical components, and random noise components in the original cost data. Resource consumption rate characteristics are mainly extracted from the trend components, while cash flow cycle characteristics are strongly correlated with the seasonal cyclical components.

[0065] Establishing a supplier relationship graph is another component of building a dynamic cost feature library. The supplier relationship graph is a graph data structure where nodes represent various material suppliers participating in the project. The graph is constructed based on purchase contract data, delivery note data, and invoice data, extracting supplier entities and the business relationships between them. Supply overlap characteristics are calculated by analyzing the material category lists provided by different suppliers. Specifically, this involves performing an intersection operation on the supply material sets of any two suppliers and calculating the Jaccard similarity coefficient. A high coefficient value reflects the substitutability or competitive relationship between suppliers in material supply. Price fluctuation synergy characteristics measure the consistency of price changes among different suppliers for the same or similar materials. The calculation process requires collecting historical price sequences from each supplier and performing Pearson correlation analysis on these sequences. A high positive correlation indicates that synchronized changes in supplier pricing strategies may be influenced by common market factors. Time-series features, resource features, and supplier features are fused into a dynamic cost feature vector set according to weighted proportions. This fusion process is a multi-source information integration step. The time-series features include resource consumption rate features and capital flow cycle features extracted from each time-series data block. Resource features mainly point to the intensity distribution and structural characteristics of resource consumption, while supplier features cover supply overlap and price fluctuation synergy. The determination of weight ratios relies on statistical analysis of historical engineering project data, analyzing the importance scores of different feature types in the cost anomaly diagnosis model. The importance score is quantified by the feature's contribution to the model's prediction accuracy. Weight allocation follows the principle that features with higher importance receive higher weights; for example, resource consumption rate features may be assigned higher weights because they directly reflect construction efficiency. The fusion operation is implemented through vector concatenation and weighted averaging algorithms. Time-series features, resource features, and supplier features are quantified into numerical vectors, which are then linearly combined according to a preset weight matrix to generate a comprehensive dynamic cost feature vector. Each vector in the dynamic cost feature vector set corresponds to a time slice or business scenario, and the entire set constitutes the core input of the subsequent cost anomaly detection model.

[0066] The specific implementation of time-dimension slicing relies on the project's work breakdown structure (WBS). The WBS breaks down the entire project into several manageable work packages, and the start and end times of each work package naturally constitute the boundaries of the time slice. The data engine reads the timestamp fields from the standardized cost dataset and allocates data records to corresponding time-series data blocks based on these boundary points. The time-series data blocks not only contain cost amount information but also associate metadata such as construction activity codes and resource classification codes. This metadata helps ensure the contextual relevance of subsequent feature extraction. Slicing ensures the preservation of the temporal locality of cost data, enabling features to accurately reflect cost behavior patterns within specific construction phases. The execution of the feature decomposition algorithm requires configuring sliding window parameters. The size of the sliding window is set according to the typical duration of the construction activity; for example, the window is smaller for short-cycle activities like concrete pouring and larger for long-cycle activities like steel structure installation. The algorithm slides the window across the time-series data blocks, independently calculating resource consumption rates and cash flow cycles within each window position. The calculation of resource consumption rates may further distinguish between primary materials such as steel reinforcement and cement and auxiliary materials such as formwork supports, calculating their consumption curves and extracting features separately. Cash flow cycle analysis distinguishes between different payment types such as progress payments, settlement payments, and warranty deposits. Each payment type has its own inherent cycle pattern, and feature decomposition algorithms need to adapt to this diversity.

[0067] The construction of the supplier association graph includes two sub-steps: entity resolution and relationship establishment. Entity resolution eliminates ambiguity caused by inconsistent supplier names; for example, the same supplier may use abbreviations and full names in different documents. Entity resolution solves this problem through string matching and verification of the unified social credit code. Relationship establishment is based on actual supply records. For example, when two suppliers have simultaneously supplied steel bars to different sections of a project, an edge representing the supply overlap relationship is established between these two supplier nodes in the graph. The edge attributes can record information such as the quantity of overlapping material types and the total overlap amount. The calculation of price fluctuation synergy is usually performed by grouping by material category because the factors affecting the prices of different categories of materials vary greatly. The generation of dynamic cost feature vector sets needs to address the problem of inconsistent feature dimensions. Resource consumption rate may be measured in tons / day, cash flow cycle in days, supply overlap is a dimensionless ratio, and price synergy is a correlation coefficient. Before merging by weight ratio, feature standardization must be performed to normalize all feature values ​​to the same numerical range, such as between zero and one. The weighting ratios themselves can be designed as adjustable parameters, allowing project managers to fine-tune the emphasis on different features based on the specific characteristics of the current project. The generated dynamic cost feature vector set is stored in the form of a database table or feature file, with each vector bearing a timestamp and project stage label, facilitating retrieval and learning by the cost anomaly detection model according to time sequence or business stage. The update frequency of the vector set is consistent with the data collection frequency, ensuring that the feature library can dynamically reflect the latest status of project costs. The entire dynamic cost feature library is built on a distributed computing framework to handle the processing needs of massive cost data from large-scale engineering projects. The standardized cost dataset is stored in a distributed file system, and the time dimension slicing and feature extraction tasks are decomposed into multiple parallel subtasks executed on the computing cluster. The construction and updating of the supplier association graph utilizes the efficient graph traversal capabilities of the graph computing engine. The final generated dynamic cost feature vector set serves as the data intermediate layer of the entire cost control method, providing a structured, machine-readable cost feature representation for upstream anomaly detection and downstream strategy generation.

[0068] Example 2: Establishing a cost anomaly detection model requires configuring the structural parameters of a deep neural network. The number of nodes in the input layer of the deep neural network is consistent with the dimension of the dynamic cost feature vector. This design ensures that the feature vector from the dynamic cost feature library can be completely received without losing dimensional information. The dimension of the dynamic cost feature vector is jointly determined by the number of time-series features, resource features, and supplier features fused during the feature extraction stage. Each node in the input layer corresponds to a specific component of the feature vector. The hidden layers of the deep neural network adopt a multilayer perceptron architecture. The number of hidden layers is adjusted according to the complexity of historical engineering data. For large infrastructure projects, more hidden layers may be set to capture deep nonlinear relationships. The ReLU function is chosen as the activation function. The ReLU function has a constant gradient in the positive interval, which helps to alleviate the gradient vanishing problem. The output layer is designed as a single node to calculate the cost deviation exponent for continuous values. The weights of the deep neural network are initialized using the Xavier method to keep the variance of each layer's output stable, thereby accelerating training convergence.

[0069] Historical project anomaly data is loaded as training samples. This data comes from archived documents of completed projects, including normal cost data segments recorded during project execution and cost data segments marked as anomalies. The anomaly segments are labeled with specific anomaly types, such as sudden cost increases due to supplier defaults or budget overruns caused by design changes. Constructing training samples involves pairing the dynamic cost feature vectors of historical projects with the corresponding ground truth values ​​of cost deviation indices. These ground truth values ​​are calibrated by domain experts based on the final audit results of the projects. The training sample set is divided into training, validation, and testing subsets. The training subset is used for model parameter learning, the validation subset for hyperparameter tuning, and the testing subset for evaluating the model's generalization ability. The sample loading process includes batch reading and preprocessing. Preprocessing operations include feature scaling and shuffling to improve training effectiveness.

[0070] The backpropagation algorithm is used to adjust the network weight parameters of the deep neural network. Backpropagation works based on the principle of gradient descent, calculating the gradient of the loss function with respect to the network weights. The loss function is defined as the mean squared error between the model's predicted cost deviation index and the actual cost deviation index; the optimization objective is to minimize this error function. Training is performed iteratively. In each iteration, a mini-batch of samples is drawn from the training subset, the predicted values ​​are calculated via forward propagation, and the error is backpropagated to update the weights. The learning rate, as a key hyperparameter, controls the step size of weight updates. The value of the learning rate is selected based on the performance on the validation subset. Sometimes, a learning rate decay strategy is used to reduce the step size in the later stages of training to improve accuracy. The training process continues until the model's error on the validation subset no longer decreases significantly or the preset maximum number of iterations is reached. Early stopping is used to prevent overfitting.

[0071] A cost deviation index calculation unit is set in the output layer. This unit is the activation function part of the deep neural network's output node, converting the network's linear weighted sum into an exponential value with practical physical meaning. The cost deviation index is composed of the product of the Euclidean distance between the actual cost and the budgeted cost and the feature fluctuation amplitude. The Euclidean distance between the actual cost and the budgeted cost measures the absolute deviation between the two in the numerical space. The feature fluctuation amplitude reflects the degree of variation of each component in the dynamic cost feature vector, and is quantified by calculating the standard deviation of the feature vector within the sliding time window. The product operation allows the cost deviation index to consider both the absolute magnitude of the deviation and the stability of the data. A larger Euclidean distance combined with a higher feature fluctuation amplitude will produce a significant cost deviation index. The output range of the cost deviation index calculation unit is limited to between zero and one by the Sigmoid function, facilitating the setting of a uniform preset threshold for anomaly detection. When the real-time input dynamic cost feature vector triggers the generation of the deviation index, the deep neural network simultaneously outputs anomaly type classification labels. The generation of anomaly type classification labels relies on a parallel classification branch within the deep neural network. This branch separates at a higher level of the hidden layers and uses the Softmax function to output the probability of belonging to each predefined anomaly type. Predefined anomaly types include multiple categories such as material price anomalies, construction progress anomalies, and supplier performance anomalies. These categories are summarized based on common cost problem patterns in historical projects. Real-time input dynamic cost feature vectors flow through the network, calculating not only a continuous cost deviation index but also a discrete category label in the classification branch. The cost deviation index indicates the severity of the anomaly, while the anomaly type classification label indicates the possible causes of the anomaly. The combination of these two provides more comprehensive information input for generating subsequent cost optimization strategy sets. The triggering mechanism is real-time; each new dynamic cost feature vector entering the model immediately generates an output, ensuring the immediacy of cost monitoring.

[0072] The specific architecture of deep neural networks may employ convolutional layers to process local spatial patterns in feature vectors, such as local correlations in supplier association features. Long Short-Term Memory (LSTM) network units may also be introduced into the architecture to capture temporal dependencies in cost data; these memory units can learn the temporal development patterns of cost deviation events. The adjustment of network weight parameters can utilize stochastic gradient descent with momentum, as the momentum term helps accelerate convergence and escape local optima. Dropout techniques are often applied to hidden layer nodes during training, randomly discarding some node connections to enhance the model's robustness and prevent overfitting to the training data. The design of the cost deviation index calculation unit allows for the introduction of attention mechanisms, enabling the model to automatically focus on the feature components in the dynamic cost feature vector that have the greatest impact on the current cost state. The quality of historical project anomaly data directly affects the performance of the cost anomaly detection model; data cleaning steps need to exclude incorrectly labeled or incomplete records. Data augmentation techniques are sometimes used to increase the diversity of training samples, such as applying small random perturbations to normal dynamic cost feature vectors to generate new samples. The calculation method for the cost deviation index can be further refined, for example, by introducing weight coefficients for Euclidean distance and feature fluctuation amplitude, with these weight coefficients adjusted according to project type or stage importance. The output of anomaly type classification labels does not need to be mutually exclusive; a cost event may involve multiple anomaly types simultaneously. In this case, the model can output a multi-label classification result. The real-time input processing workflow needs to be tightly integrated with the data acquisition system. Each time the dynamic cost feature library updates a set of feature vectors, the cost anomaly detection model immediately performs a forward propagation calculation. The model's computational efficiency is optimized to meet the needs of real-time monitoring at the engineering site. Model quantization or pruning techniques are used to reduce computational load when necessary. The parameters and performance of the cost anomaly detection model need to be periodically re-evaluated and updated using new project data to adapt to concept drift phenomena in engineering practice.

[0073] Example 3: See Figure 3The generation of cost optimization strategy sets and the execution strategy verification process constitute a coherent decision support chain. The starting point for generating the cost optimization strategy set is to establish a three-dimensional retrieval index in the historical engineering case library. The historical engineering case library is a structured database that stores cost data, abnormal event records, corrective measures taken, and evaluation data of the effects of these measures on completed projects. The three-dimensional retrieval index has three axial dimensions: cost deviation magnitude, project type, and abnormal feature combination. The cost deviation magnitude dimension maps the cost deviation index calculated by the cost anomaly detection model to a preset discrete interval. For example, interval one corresponds to a deviation index of 0.05 to 0.1, and interval two corresponds to a deviation index of 0.1 to 0.2. The project type dimension adopts an industry standard classification system, classifying projects into specific categories such as building construction projects, municipal road projects, rail transit projects, and water conservancy projects. The abnormal feature combination dimension jointly encodes the anomaly type classification label output by the cost anomaly detection model and the salient features in the dynamic cost feature vector to generate a multi-dimensional feature signature. The process of building an index involves extracting values ​​from these three dimensions for each case record in the historical engineering case library and constructing an efficient multidimensional query data structure, such as a KD-tree or R-tree, to achieve fast similarity retrieval of massive cases.

[0074] The nearest neighbor search algorithm is used to locate the top N historical cases with the highest similarity to the current anomalous scenario. This algorithm operates within a vector space defined by a three-dimensional index. The current anomalous scenario is transformed into a query vector, whose elements include the interval code of the current cost deviation index, the project type code of the current project, and the combination code of anomalous features generated from the current dynamic cost feature vector and the anomalous type label. Each case in the historical project case library is also represented as a vector with the same structure. The nearest neighbor search algorithm calculates the distance metric between the query vector and all case vectors, using either Euclidean distance or Manhattan distance formulas. The algorithm retrieves the N cases with the smallest distance to the query vector. These cases are most similar to the current scenario in terms of cost deviation, project attributes, and anomalous feature patterns, forming the reference case set for strategy generation. The value of parameter N is determined through cross-validation to balance the comprehensiveness and relevance of the retrieval.

[0075] The system analyzes corrective measures and their impact data from historical cases to generate a list of strategy options with impact assessment weights. Corrective measure records contain specific operational instructions, such as switching the steel rebar supplier from Company A to Company B, adjusting the construction schedule from sequential to parallel operations, and adopting a new formwork system to shorten the construction period. Impact data quantifies the results after the measures are implemented, including the actual percentage cost savings, the number of days the construction period changes, and fluctuations in quality acceptance indicators. Impact assessment weights are assigned to each extracted corrective measure, with the weight calculation comprehensively considering the cost-effectiveness and reliability of the measure. The impact assessment weight of a measure is a linear combination of its average cost savings rate and impact stability within the reference case set. The mathematical expression for calculating the impact assessment weight is:

[0076]

[0077] in: Representing the Weighting of the effectiveness evaluation of each corrective measure; Representing the The average cost savings achieved by this measure across all reference cases; It is a measure of the first The indicator of the stability of the effectiveness of the measures is the standard deviation of the cost-saving rate in these cases; and These are pre-set weighting coefficients used to balance the relative importance of cost savings and effectiveness stability in the final weighting. and The value of is between zero and one, and satisfies The constraints are as follows. The analysis process traverses the reference case set, extracts each unique correction measure, and calculates its value according to the above formula. The values ​​ultimately form a list of strategy options, with each strategy in the list associated with a quantified performance evaluation weight.

[0078] A strategy combiner cross-combines individual strategies from the material, schedule, and supplier dimensions to generate composite optimization strategies. The strategy combiner is a rule-based system whose input is a list of individual strategies from a strategy option list. Material-level strategies include changing material specifications, enabling alternative raw materials, and adjusting procurement batches. Schedule-level strategies involve compressing the critical path, increasing construction resources, and optimizing workflows. Supplier-level strategies include retendering, signing supplementary agreements, and introducing performance guarantees. The strategy combiner internally defines a set of strategy compatibility rules, specifying allowed or prohibited combinations between strategies from different dimensions. For example, a schedule strategy requiring faster progress is incompatible with a material strategy requiring extended material inspection cycles. The strategy combiner uses graph theory algorithms, treating individual strategies as nodes and compatibility relationships as edges, to find the largest complete subgraph or the optimal strategy cluster, thereby generating a series of internally coordinated and parallel-implementable composite optimization strategy packages. Each composite optimization strategy package contains a comprehensive solution for collaboratively addressing cost issues from multiple perspectives.

[0079] The execution strategy verification process imports the current project's BIM model and resource scheduling plan into the engineering digital twin system, a high-fidelity virtual simulation environment. The BIM model provides a precise 3D geometric representation of the project and rich component attribute parameters, while the resource scheduling plan defines the input plans and logical dependencies of personnel, machinery, and engineering materials in a timeline format. The import process ensures that the digital twin system can accurately replicate the project's physical state and resource constraints. The composite optimization strategy is decomposed into a series of executable discrete instructions, including material procurement instructions, construction schedule instructions, and supplier switching instructions. Material procurement instructions specify the type, specifications, quantity, delivery location, and time window of materials; construction schedule instructions change the overall project duration by adjusting the logical relationships and durations between tasks; and supplier switching instructions handle contract changes with the original supplier and the collaboration process with the new supplier. These instructions are converted into event scripts or API calls recognizable by the digital twin engine.

[0080] Multiple rounds of Monte Carlo simulations are run to calculate the cost forecast distribution under different market fluctuation scenarios. Monte Carlo simulation is a probabilistic statistical method based on random sampling. In the digital twin system, the simulation focuses on the uncertainties of the external environment. Market fluctuation scenarios include price fluctuations of key building materials, changes in labor subcontracting unit prices, fluctuations in large equipment rental costs, and the risk of project delays due to extreme weather. At the beginning of each simulation round, the system generates a set of independent random numbers for these uncertain variables based on their historical probability distributions and assigns them to the simulation model. The digital twin system then drives the project simulation execution from start to finish under this specific random scenario, strictly following the decomposed instructions and dynamically calculating resource consumption, labor costs, machinery costs, etc., and finally sums them up to obtain the total cost forecast value for the project execution cycle. After a large number of iterative simulations, usually thousands of rounds, the system collects the results of all simulation runs to form a set of predicted total project costs. By performing statistical analysis on this set, the probability distribution function of the cost can be obtained, which intuitively shows the probability that the final cost will fall within different ranges.

[0081] The strategy with the median cost distribution closest to the budget target and the smallest variance is selected as the validated solution. The median cost distribution represents the most likely project cost value after considering various possible market fluctuations. The budget target is the cost control benchmark set by project management. Proximity is measured by the absolute value of the difference between the median cost distribution and the budget target. The variance of the cost distribution quantifies the uncertainty or risk level of the strategy's execution result; the smaller the variance, the more stable and predictable the cost outcome of the strategy. The selection criterion is a multi-objective optimization problem. The ideal solution is a composite optimization strategy package that both brings the most likely cost value close to the management target and minimizes the risk of cost overruns. The engineering digital twin system has a built-in optimization algorithm that automatically performs multi-objective comparison and ranking of the cost prediction distributions generated by all participating composite optimization strategy packages. Based on the above proximity and variance minimization criteria, it outputs one or a few Pareto optimal validated solutions, providing data support for the project manager's final decision. The integration of the entire set of generation cost optimization strategies and the execution strategy verification process enables closed-loop intelligent decision-making, from historical experience mining to future scenario projection.

[0082] Example 4: Real-time Feedback Adjustment Steps. During the strategy execution phase, IoT data acquisition terminals are deployed. These terminals include smart meters, water sensors, material weighing RFID tags, tower crane monitoring cameras, and GPS positioning modules for transport vehicles installed at the construction site. These terminals continuously collect actual material consumption data at a minute-level frequency, recording construction progress images and location information. The data is transmitted to the edge computing gateway via the site's wireless LAN. The edge computing gateway performs preliminary filtering and packaging of the raw data, generating a standard-format real-time data stream and uploading it to the cloud-based cost control platform. The platform dynamically compares the real-time data with the predicted cost curve generated by the engineering digital twin system during the strategy verification phase. The predicted cost curve is a time-cost function graph. The dynamic comparison calculates the execution deviation coefficient, a comprehensive indicator reflecting the degree of deviation between the actual execution trajectory and the predicted path. The calculation of the execution deviation coefficient integrates deviation data from multiple dimensions, including the difference between the actual and planned consumption of major materials, the difference between the actual completion time and planned duration of sub-projects, and the difference between on-site labor hours and the budget. The deviation for each dimension is normalized to a dimensionless relative percentage difference and then summed according to preset weights. The weighting is determined based on the proportion of each dimension's cost in the total cost and its impact on subsequent work; material cost fluctuations are typically assigned a higher weight. The calculated execution deviation coefficient is compared to a preset tolerance threshold, which is a dynamically changing numerical range. The threshold size depends on the risk level and the stringency of cost control at the current stage of the project. When the deviation coefficient exceeds the tolerance threshold, the system automatically triggers a strategy re-optimization process. Simultaneously, a strategy re-optimization process initiation signal is sent to the engineering digital twin system to prepare for re-execution and verification.

[0083] The strategy re-optimization process analyzes the main sources of deviations, identifying the dominant factors in material price fluctuations, decreased construction efficiency, or supplier performance anomalies. The analysis employs a decision tree classification algorithm, with inputs being feature variables from the real-time data stream. These variables include the difference between invoice prices and contract prices for each material, the number of machine shifts consumed per unit of work, and the frequency and duration of supplier delivery delays. The decision tree model learns the mapping relationship between different deviation patterns and root causes using historical training data, outputting the most likely category label for the current deviation. The dominant factor identification results categorize the root causes of deviations into specific types such as changes in the market environment, internal management issues, or supply chain risks, providing direction for subsequent targeted searches. Based on the dominant factor type, search conditions are added to the historical project case library to narrow the scope of similar cases. The search query in the historical project case library adds filtering conditions related to the root causes of deviations to the existing three-dimensional index. For example, when the dominant factor is identified as a material price fluctuation, the search conditions are further limited to searching only cases where similar sharp material price fluctuations have occurred historically. Additional search conditions may include fields such as the lower limit of supplier credit rating, material price fluctuation thresholds, and specific construction process codes. The retrieval system then runs a nearest neighbor search algorithm again within the narrowed case set to locate a subset of historical cases that are highly similar to the current scenario at the root cause level of the deviation. Narrowing the case selection scope helps improve the relevance and adaptability of the extracted emergency strategies.

[0084] The contingency strategies from newly added cases were extracted and their suitability to the current engineering scenario was evaluated. Newly added cases refer to highly relevant historical cases selected through additional search criteria. Contingency strategies are the emergency response measures taken in these cases when facing similar unforeseen deviations; these measures typically feature rapid response and risk control. The suitability assessment involved a multi-dimensional matching analysis, including remaining project timeline, available budget, on-site technical conditions, and contractual constraints. A multi-factor scoring table was used to quantitatively score each candidate contingency strategy. Table 1 illustrates the main considerations and scoring criteria for the suitability assessment.

[0085] Table 1: Evaluation Factors for the Adaptability of Emergency Strategies to Current Engineering Scenarios

[0086]

[0087] An incremental optimization strategy package is generated, incorporating risk hedging measures. This package is a supplementary set of measures added to the original composite optimization strategy. Risk hedging measures are designed to address identified dominant deviation factors, preventing further deterioration or mitigating some negative impacts. Hedging measures for sudden changes in material prices may include signing price-locking agreements, purchasing futures contracts, and utilizing local alternative material sources. Measures for declining construction efficiency involve introducing skilled worker teams, leasing high-efficiency equipment, and optimizing the division of construction workflows. Measures for supplier performance anomalies include activating backup suppliers, requiring bank performance guarantees, and adding third-party inspection procedures. The incremental optimization strategy package is integrated with the original strategy package to form a complete enhanced strategy solution that includes basic optimization and contingency response. The enhanced strategy solution is then fed into the engineering digital twin system for multiple rounds of Monte Carlo simulation to verify its effectiveness under the current project status and the latest market conditions, outputting new predicted cost curves and selecting the final executable solutions. This entire real-time feedback adjustment process and strategy re-optimization workflow form a dynamic closed loop for cost control, enabling the system to adapt to uncertainties and changes during project execution. The deployment plan for IoT data acquisition terminals needs to cover all key cost control points, and the real-time performance and reliability of data transmission directly affect the timeliness of deviation identification. The calculation model for the deviation coefficient needs to be calibrated regularly to adapt to changes in cost sensitivity at different project stages. The trigger frequency of the strategy re-optimization process needs to be set reasonably to avoid overly frequent adjustments that could lead to execution chaos. The design of incremental optimization strategy packages should emphasize modularity, facilitating integration and replacement with existing strategies. The cost-effectiveness of risk hedging measures needs to be carefully weighed to ensure that hedging costs do not exceed potential loss risks. The historical project case library needs to be continuously updated with the latest deviation handling cases to enrich the sources of contingency strategies. The suitability assessment factors and weights can be adjusted according to the specific characteristics of the project to reflect the personalized needs of different projects.

[0088] See Figure 4This chart demonstrates the practical application of a real-time feedback adjustment mechanism in engineering cost control. Through multi-source monitoring data acquired via IoT data acquisition terminals, the system can capture real-time data on actual material consumption and construction progress deviations. The time series data presented in the chart, including smart meter readings, water sensor data, material weighing information, and tower crane monitoring indicators, fully demonstrate the comprehensiveness and real-time nature of the data acquisition. The clearly marked strategy optimization trigger points in the chart represent the critical state where the execution deviation coefficient exceeds the fault tolerance threshold, at which point the system automatically initiates the strategy re-optimization process. Normalized monitoring values ​​facilitate comparative analysis of data from different dimensions, while the marked execution deviation coefficient and strategy success rate indicators intuitively reflect the overall effectiveness of the cost control system. This chart verifies the practical value of the engineering digital twin system in virtual simulation and strategy verification, showcasing a complete closed-loop control process from data acquisition to strategy execution, and providing an effective technical means to address uncertainties and changes during project execution.

[0089] Example 5: The cost knowledge graph update step packages the cost data flow, abnormal events, and execution strategies for the entire project lifecycle into a case data package, which is a structured data collection. Taking a super high-rise building project called "Binhai International Business Center" as an example, the cost data flow for the entire project lifecycle includes all financial records from project initiation, design, bidding, construction to final settlement. The records include the budget approval documents in the design phase, daily steel and concrete purchase and warehousing slips in the construction phase, monthly output value reports, subcontractor progress payment vouchers, and spare parts procurement plans in the operation and maintenance preparation phase. The abnormal event record details a major cost anomaly encountered in the middle of the project construction, described as "During the main structure construction phase, the domestic steel market price rose by more than 30% continuously within three months, resulting in raw material costs seriously exceeding the budget." The execution strategy document documents the response measures taken by the project team at the time. These measures included: "After verification by the engineering digital twin system, applying for and obtaining approval to use HRB400E steel bars to replace the originally designed HRB500 steel bars in some non-core load-bearing components"; "renegotiating with long-term steel bar supplier A to lock in subsequent supply prices at a floating upper limit"; and "urgently activating backup supplier B to share 40% of the procurement volume to mitigate price risks." All of this data was extracted, cleaned, correlated, and timestamped using a data packaging tool, ultimately generating a case data package file that conforms to the specified format.

[0090] Feature vectors and strategy effectiveness metrics are extracted from the case study data package and transformed into new nodes and edge relationships in a knowledge graph. The knowledge graph is a dataset that stores knowledge in a graph structure. Feature vectors are extracted from the cost data stream and anomaly events in the case study data package. These feature vectors include basic project attributes such as project type "super high-rise building," structural form "frame-core tube," and total building area "250,000 square meters"; cost characteristics such as steel cost percentage "18% of total construction cost" and cost deviation index "0.35"; and anomaly characteristics such as anomaly type "surge in raw material prices" and dominant factor "changes in market environment." Strategy effectiveness metrics quantify the results of the adopted strategies. These metrics include cost savings rate "through material substitution and supplier negotiations, steel cost overruns were controlled within 12%" and schedule impact "due to supplier switching and material certification, the local schedule was delayed by 15 days." In the knowledge graph, this information is transformed into new nodes. Node types include "Project Node" representing the "Binhai International Business Center" project itself, "Material Node" representing "HRB400E steel bars" and "HRB500 steel bars," "Supplier Node" representing "Supplier A" and "Supplier B," "Abnormal Event Node" representing "Steel Price Surge," and "Strategy Node" representing "Material Substitution Strategy" and "Multi-Source Procurement Strategy." Edge relationships describe the connections between nodes, with edge relationship types including "Project-Encounter-Abnormal Event," "Project-Application-Strategy," "Strategy-Target-Material," "Strategy-Involved-Supplier," and "Strategy-Generate-Effect."

[0091] The graph embedding algorithm updates node vector representations to optimize the semantic matching accuracy of subsequent case retrieval. This algorithm maps nodes and edges in a knowledge graph to a low-dimensional continuous vector space. The algorithm employs the TransR model, which learns different transformation matrices for different relation types. The model training process takes into account the structural information of the entire knowledge graph, including all nodes and edges. The TransR model projects each node from an entity space to a relation space through a transformation matrix corresponding to a specific relation. In the relation space, the objective function is optimized to ensure that the vector representations of two nodes with an "application" relation are as close as possible in space after transformation using the matrix corresponding to the "application" relation. After training, each node in the knowledge graph receives a fixed-length distributed vector representation that captures the node's semantic information and its structural context within the graph. When the "Binhai International Business Center" case is added to the graph and embedding training is completed, the vector representations of all relevant nodes in the graph are fine-tuned and updated. For example, the vector representation of the "material substitution strategy" node will be enhanced by this successful application case, making its relative positional relationship in the vector space with the "raw material price surge" anomaly event node and the "HRB400E steel bar" material node clearer. This vectorized representation allows for direct calculation of the cosine similarity between node vectors during subsequent case similarity retrieval. This vector-based similarity calculation is better able to understand deeper semantic relationships than traditional keyword matching methods, thereby improving retrieval accuracy.

[0092] The multi-project collaborative optimization steps establish a cost correlation analysis channel across engineering projects within the knowledge graph. This channel utilizes a path query and reasoning mechanism based on graph node and edge relationships. It identifies cost linkage effects caused by shared suppliers or identical construction techniques between different projects, leveraging the graph's global query capabilities. For example, the knowledge graph stores data for another ongoing project, "Jianghewan Residential Community," which also heavily utilizes steel reinforcement and shares the same steel reinforcement supplier, A, with the "Binhai International Business Center" project. A graph query can find all project nodes simultaneously connected to both the "Supplier A" node and the "Supplier Reinforcement" material node. Path analysis reveals that the "Binhai International Business Center" project node is connected to the "Multi-Source Procurement Strategy" node via an "Application" edge, and this strategy node is further connected to the "Supplier B" node via an "Involvement" edge. This correlation reveals that the supply stability or price fluctuations of "Supplier A" may have a cascading impact on multiple projects, and the "Multi-Source Procurement Strategy" may be an effective risk mitigation measure, its effectiveness having been validated in the "Binhai International Business Center" project. The cost linkage effect manifests itself in the fact that when a project needs to activate a backup supplier due to market changes, it may affect the backup supplier's ability to supply or quote for other projects.

[0093] When a new project triggers cost anomalies, the system simultaneously retrieves historical strategies from related projects and generates collaborative optimization suggestions. For example, a newly started "New City Metro Hub" project detected an abnormal increase in ready-mixed concrete prices during the foundation construction phase. After the cost anomaly detection model generates an alert, the system not only searches the knowledge graph for historical cases directly similar to "New City Metro Hub" in terms of project type and anomaly characteristics, but also automatically retrieves other projects potentially related to "New City Metro Hub" through cost correlation analysis. The search logic includes finding historical projects located in the same area as "New City Metro Hub," using the same main materials (such as ready-mixed concrete), or sharing suppliers (such as a single concrete mixing plant). The system retrieved the "Binhai International Business Center" project node because both are large-scale infrastructure projects, geographically located in the same city, and face similar regional market environments. The system further discovered that the "Binhai International Business Center" project node had successfully applied a "multi-source procurement strategy" to address material price risks. The collaborative optimization recommendations not only suggest that the "New City Metro Hub" project consider adopting a "multi-source procurement strategy" to introduce backup concrete suppliers, but also generate higher-level collaborative recommendations. For example, the recommendations might state: "Based on knowledge graph correlation analysis, the 'Binhai International Business Center' project shares regional concrete market information with your project. The 'Binhai International Business Center' project is currently in the decoration phase, and its demand for high-grade concrete is decreasing. It is recommended to coordinate with the management of the 'Binhai International Business Center' project to assess the possibility of transferring a portion of its locked-in long-term concrete supply contracts or its partner backup supplier resources to your project, in order to achieve the integration effect of regional procurement scale and jointly cope with market price fluctuations." These recommendations, generated based on deep knowledge graph correlation analysis, transcend the limitations of a single project and embody the core value of multi-project collaborative optimization.

[0094] The cost knowledge graph update process is a continuous learning process, with the experience of each completed project contributing to the graph's growth. The multi-project collaborative optimization step fully leverages the graph's networked knowledge structure to uncover systemic risks and global optimization opportunities hidden behind scattered projects. The construction of the knowledge graph relies on standardized data entry and entity linking technologies to ensure that information from newly added case data packages is accurately integrated into the existing knowledge network. Training the graph embedding algorithm requires sufficient computing resources to handle the ever-growing graph size. The accuracy of cross-project correlation analysis depends on the completeness and consistency of the data within the graph. Ultimately, the effectiveness of collaborative optimization recommendations needs to be evaluated and adopted by project decision-makers in real-world management scenarios.

[0095] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0096] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for engineering cost control based on big data, characterized in that, Includes the following steps: Collect the full lifecycle cost data stream of the project, including budget parameters in the design phase, material consumption parameters in the construction phase, and loss parameters in the operation and maintenance phase. After removing outliers through a data cleaning engine, a standardized cost dataset is generated. A dynamic cost feature library is constructed, and multi-dimensional features are extracted from the standardized cost dataset. The extracted features include time series fluctuation features, resource allocation discrete features, and supplier correlation features. Establish a cost anomaly detection model, input a dynamic cost feature library into a pre-trained deep learning network, output a cost deviation index, and trigger a cost correction instruction when the deviation index exceeds a preset threshold; Generate a set of cost optimization strategies, match similar scenarios in the historical engineering case library based on cost correction instructions, and call the strategy generator to output optimization strategies including material replacement schemes, schedule adjustment schemes, and supplier replacement schemes; The execution strategy verification process involves inputting the optimized strategy into the engineering digital twin system for virtual simulation, outputting the predicted cost curve after strategy execution, and selecting the strategy whose predicted cost curve is closest to the target value as the final execution plan. The specific steps for constructing the dynamic cost feature library include: The standardized cost dataset is sliced ​​along the time dimension to divide it into time-series data blocks arranged according to construction progress nodes; the feature decomposition algorithm is used to separate resource consumption rate features and capital flow cycle features from the time-series data blocks; a supplier association graph is established to extract the supply overlap and price fluctuation synergy features among various material suppliers; and the time-series features, resource features, and supplier features are fused into a dynamic cost feature vector set according to weight ratios. The specific steps for establishing the cost anomaly detection model include: Configure the structural parameters of the deep neural network, ensuring the number of nodes in the input layer matches the dimension of the dynamic cost feature vector; load historical engineering anomaly data as training samples and adjust the network weight parameters using the backpropagation algorithm; set up a cost deviation index calculation unit in the output layer, which is composed of the product of the Euclidean distance between the actual cost and the budgeted cost and the feature fluctuation amplitude; when the real-time input dynamic cost feature vector triggers the generation of the deviation index, output the anomaly type classification label simultaneously; The specific steps for generating the cost optimization strategy set include: A three-dimensional retrieval index is established in the historical engineering case library, with index dimensions including cost deviation range, project type, and combination of abnormal features; the nearest neighbor search algorithm is used to locate the top N historical cases with the highest similarity to the current abnormal scenario; the corrective measures and their effect data in the historical cases are analyzed to generate a list of strategy options with effect evaluation weights; and the single strategies of material, schedule, and supplier dimensions are cross-combined through a strategy combiner to generate a composite optimization strategy.

2. The engineering cost control method based on big data according to claim 1, characterized in that, The specific steps of the execution strategy verification process include: Import the current project's BIM model and resource scheduling plan into the engineering digital twin system; The composite optimization strategy is decomposed into executable material procurement instructions, construction schedule instructions, and supplier switching instructions; Run multiple rounds of Monte Carlo simulations to calculate the cost forecast distribution under different market volatility scenarios; The strategy with the median cost distribution closest to the budget target and the smallest variance was selected as the validation scheme.

3. The engineering cost control method based on big data according to claim 2, characterized in that, It also includes a real-time feedback adjustment step: During the strategy execution phase, deploy IoT data acquisition terminals to capture real-time data on actual material consumption and construction progress deviations. The real-time data is dynamically compared with the predicted cost curve to generate an execution deviation coefficient. When the deviation coefficient exceeds the fault tolerance threshold, the strategy re-optimization process is triggered and the digital twin verification is re-executed.

4. The engineering cost control method based on big data according to claim 3, characterized in that, The specific steps of the trigger strategy re-optimization process include: Analyze the main sources of execution deviations and identify the dominant factors in sudden changes in material prices, decreased construction efficiency, or abnormal supplier performance. Add search criteria to the historical case database based on the type of dominant factor to narrow down the scope of similar cases; Extract emergency strategies from new cases and assess their suitability for the current engineering scenario; Generate an incremental optimization strategy package that includes risk hedging measures.

5. The engineering cost control method based on big data according to claim 4, characterized in that, It also includes the cost knowledge graph update steps: The cost data stream, abnormal events, and execution strategies for the entire project lifecycle are packaged into a case data package. Extract feature vectors and policy performance metrics from case data packages and transform them into new nodes and edge relationships in the knowledge graph; The node vector representation is updated by graph embedding algorithm to optimize the semantic matching accuracy of subsequent case retrieval.

6. The engineering cost control method based on big data according to claim 5, characterized in that, It also includes multi-project collaborative optimization steps: Establish cost correlation analysis channels across engineering projects in the knowledge graph; Identify the cost linkage effects caused by sharing suppliers or the same construction techniques across different projects; When a new project triggers cost anomalies, the historical strategies of related projects are retrieved simultaneously, and collaborative optimization suggestions are generated.

7. A big data-based engineering cost control system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the big data-based engineering cost control method according to any one of claims 1 to 6.

Images